Video yükleniyor...

Video Yüklenemedi

Ana Sayfaya Dön

Original Design: Medieval Dog Model: XenoBuster: Synthetic Xenomorph Hunter Modified & painted : Zio Yeh Printer: Phrozen Sonic Mighty Revo MAX Download the model on PIXUP 3D :

ZioYeh's profile picture

Zio Yeh

4,161 subscribers

21,744 görüntüleme • 1 ay önce •via X (Twitter)

LIVE

Anya Rossi• Live Now

Private livecam show

0 Yorum

Yorum bulunmuyor

Orijinal gönderinin yorumları burada görünecek

Benzer Videolar

INTRODUCING FLOPPI IN 3D!!! Thanks to: Andy-O is WRAPPING UP HNAUTS?! for working with me on the original design Valz for designing the banger 3D model Flapcats🔞 for not only rigging the model, but also making this animation and also making a tall version of Floppi #floppiart
0:21

Sensitive content

INTRODUCING FLOPPI IN 3D!!! Thanks to: Andy-O is WRAPPING UP HNAUTS?! for working with me on the original design Valz for designing the banger 3D model Flapcats🔞 for not only rigging the model, but also making this animation and also making a tall version of Floppi #floppiart

MENTLE

822,990 görüntüleme • 3 yıl önce

Check out how Perception designed the main title for the end credits of #TheMarvels in this design breakdown. Using #Cinema4D, we created a 3D model of our type. A render of the type was brought into #Photoshop, painted over and mapped back on to the 3D geometry to create the illustrated look. Each outline, edge highlight and grunge texture was meticulously placed to get the perfect style.
0:50

Check out how Perception designed the main title for the end credits of #TheMarvels in this design breakdown. Using #Cinema4D, we created a 3D model of our type. A render of the type was brought into #Photoshop, painted over and mapped back on to the 3D geometry to create the illustrated look. Each outline, edge highlight and grunge texture was meticulously placed to get the perfect style.

Perception

57,146 görüntüleme • 2 yıl önce

AN OPENCLAW AI AGENT JUST DESIGNED A 3D CHARACTER IN BLENDER AND SENT IT TO A 3D PRINTER. SHE DIDN'T TOUCH A SINGLE BUTTON. This is what running an OpenClaw AI agent on a Mac Mini actually looks like. She described what she wanted. The agent did everything else. Blender opened automatically. Character designed. Put in space. Animated with blinking eyes and stomping feet. Landing page built and live. Model sliced. File sent to 3D printer. Printer started moving. All from conversation. Zero code. Zero design skills. Zero manual work. Here's why this matters beyond the demo: Custom 3D printed figures sell for $30-$200 each on Etsy and at conventions. The people making serious money aren't artists. They're the ones who can produce custom designs fast. An AI agent that goes from description to printed figure autonomously is a business model not a toy. "I was just talking to the guy and he made it happen." Bookmark this. Full demo in the video below.
1:29

AN OPENCLAW AI AGENT JUST DESIGNED A 3D CHARACTER IN BLENDER AND SENT IT TO A 3D PRINTER. SHE DIDN'T TOUCH A SINGLE BUTTON. This is what running an OpenClaw AI agent on a Mac Mini actually looks like. She described what she wanted. The agent did everything else. Blender opened automatically. Character designed. Put in space. Animated with blinking eyes and stomping feet. Landing page built and live. Model sliced. File sent to 3D printer. Printer started moving. All from conversation. Zero code. Zero design skills. Zero manual work. Here's why this matters beyond the demo: Custom 3D printed figures sell for $30-$200 each on Etsy and at conventions. The people making serious money aren't artists. They're the ones who can produce custom designs fast. An AI agent that goes from description to printed figure autonomously is a business model not a toy. "I was just talking to the guy and he made it happen." Bookmark this. Full demo in the video below.

SCOTTY BEAM

17,209 görüntüleme • 2 ay önce

Alibaba just released a coding model that hits 82 percent on SWE-Bench Verified. That is the highest score ever published for an open-source model. The weights are free. The license is Apache 2.0. You can run it today. The model is Qwen 4 Coder 32B. Here is what 82 percent on SWE-Bench Verified actually means. SWE-Bench Verified tests whether an AI can autonomously resolve real bugs pulled from real production GitHub repositories. Not synthetic exercises. Real open-source projects that real teams depend on. A model gets a bug report, reads the code, writes a fix, and either passes the test suite or it does not. At 82 percent, Qwen 4 Coder 32B resolves 82 out of every 100 real production bugs it is given. Without a human guiding it. On code it has never seen before. For comparison: Qwen 4 Coder 32B: 82 percent SWE-Bench Verified. Open source. Apache 2.0. Claude Fable 5: 80.3 percent SWE-Bench Pro. $10 input / $50 output per million tokens. Currently suspended. GPT-5.6 Sol: Competitive on Terminal-Bench. $5 input / $30 output per million tokens. An open-weight model that you can download and run for free just beat both of them on the benchmark designed to measure real software engineering capability. Here is the architecture. Qwen 4 Coder 32B is a 32 billion parameter dense model. Not a Mixture-of-Experts. Every parameter is active on every request. This matters for inference: a dense 32B model runs on 22 gigabytes of VRAM, which fits on a single high-end consumer GPU or a MacBook Pro with 64GB of unified memory. The smaller variant, Qwen 4 Coder 4B, runs at approximately 135 tokens per second on an M5 Max and fits inside 8 gigabytes of RAM. For a model with usable coding capability, that is a new bar for what fits in a single laptop. The training methodology continued Alibaba's approach of reinforcement learning on verifiable coding tasks. The model gets rewarded when its code passes tests. It gets penalized when it fails. Over millions of training steps, the model learns to write code that actually runs rather than code that looks plausible. License: Apache 2.0. Full commercial use. No attribution requirement. No revenue threshold. No monthly active user ceiling. Weights: Hugging Face, available today. Runs on: vLLM, Ollama, SGLang, and any standard GGUF-compatible inference engine. Qwen 4 32B also runs at approximately 135 tokens per second on an M5 Max chip, setting a new bar for what a sub-8GB model can do on Apple Silicon. The open-source coding model just beat the best closed-source model in the world on the benchmark designed to test whether AI can actually do software engineering. The weights are free. The subscription is optional. Source: Autom8Labs AI Insight July 2026, State of Open Source LLMs June 2026, Kunal Ganglani blog June 2026.
0:25

Alibaba just released a coding model that hits 82 percent on SWE-Bench Verified. That is the highest score ever published for an open-source model. The weights are free. The license is Apache 2.0. You can run it today. The model is Qwen 4 Coder 32B. Here is what 82 percent on SWE-Bench Verified actually means. SWE-Bench Verified tests whether an AI can autonomously resolve real bugs pulled from real production GitHub repositories. Not synthetic exercises. Real open-source projects that real teams depend on. A model gets a bug report, reads the code, writes a fix, and either passes the test suite or it does not. At 82 percent, Qwen 4 Coder 32B resolves 82 out of every 100 real production bugs it is given. Without a human guiding it. On code it has never seen before. For comparison: Qwen 4 Coder 32B: 82 percent SWE-Bench Verified. Open source. Apache 2.0. Claude Fable 5: 80.3 percent SWE-Bench Pro. $10 input / $50 output per million tokens. Currently suspended. GPT-5.6 Sol: Competitive on Terminal-Bench. $5 input / $30 output per million tokens. An open-weight model that you can download and run for free just beat both of them on the benchmark designed to measure real software engineering capability. Here is the architecture. Qwen 4 Coder 32B is a 32 billion parameter dense model. Not a Mixture-of-Experts. Every parameter is active on every request. This matters for inference: a dense 32B model runs on 22 gigabytes of VRAM, which fits on a single high-end consumer GPU or a MacBook Pro with 64GB of unified memory. The smaller variant, Qwen 4 Coder 4B, runs at approximately 135 tokens per second on an M5 Max and fits inside 8 gigabytes of RAM. For a model with usable coding capability, that is a new bar for what fits in a single laptop. The training methodology continued Alibaba's approach of reinforcement learning on verifiable coding tasks. The model gets rewarded when its code passes tests. It gets penalized when it fails. Over millions of training steps, the model learns to write code that actually runs rather than code that looks plausible. License: Apache 2.0. Full commercial use. No attribution requirement. No revenue threshold. No monthly active user ceiling. Weights: Hugging Face, available today. Runs on: vLLM, Ollama, SGLang, and any standard GGUF-compatible inference engine. Qwen 4 32B also runs at approximately 135 tokens per second on an M5 Max chip, setting a new bar for what a sub-8GB model can do on Apple Silicon. The open-source coding model just beat the best closed-source model in the world on the benchmark designed to test whether AI can actually do software engineering. The weights are free. The subscription is optional. Source: Autom8Labs AI Insight July 2026, State of Open Source LLMs June 2026, Kunal Ganglani blog June 2026.

Harman

41,278 görüntüleme • 1 ay önce

Open-source magnetic tactile sensor for $5! 🧲 Researchers introduced a magnetic tactile sensor that's low-cost, and easy to fabricate, democratizing tactile sensing for robotics. Operating in unstructured environments like homes and offices requires robots to sense forces during physical interaction. Yet the lack of a versatile, accessible tactile sensor has led to fragmented solutions and often force-unaware, sensorless approaches. Building an eFlesh sensor requires four components: a hobbyist 3D printer, off-the-shelf magnets (less than $5), a CAD model, and a magnetometer circuit board. The sensor is 3D printed with magnets embedded in the middle layer. Based on chosen mechanical properties, magnets displace in response to contact forces, measured by a magnetometer underneath. An open-source design tool converts simple OBJ/STL files into 3D-printable STLs. This enables application-specific sensors for robot hands, grippers, quadruped feet, and more. Slip detection generalizes to unseen objects with 95% accuracy. Visual-tactile control policies improve manipulation by 40% over vision-only baselines, achieving 90% success on precise tasks like plug insertion and credit card swiping. All design files, code, trained models, and conversion tools are openly available. Project page: ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →
0:57

Open-source magnetic tactile sensor for $5! 🧲 Researchers introduced a magnetic tactile sensor that's low-cost, and easy to fabricate, democratizing tactile sensing for robotics. Operating in unstructured environments like homes and offices requires robots to sense forces during physical interaction. Yet the lack of a versatile, accessible tactile sensor has led to fragmented solutions and often force-unaware, sensorless approaches. Building an eFlesh sensor requires four components: a hobbyist 3D printer, off-the-shelf magnets (less than $5), a CAD model, and a magnetometer circuit board. The sensor is 3D printed with magnets embedded in the middle layer. Based on chosen mechanical properties, magnets displace in response to contact forces, measured by a magnetometer underneath. An open-source design tool converts simple OBJ/STL files into 3D-printable STLs. This enables application-specific sensors for robot hands, grippers, quadruped feet, and more. Slip detection generalizes to unseen objects with 95% accuracy. Visual-tactile control policies improve manipulation by 40% over vision-only baselines, achieving 90% success on precise tasks like plug insertion and credit card swiping. All design files, code, trained models, and conversion tools are openly available. Project page: ~~ ♻️ Join the weekly robotics newsletter, and never miss any news →

Lukas Ziegler

289,626 görüntüleme • 4 ay önce

CapCut Design Studio changed the way I think about creating with AI. For a long time, creating with AI felt less like creating and more like managing files—generate, export, switch tools, edit, re-upload, repeat. The ideas were flowing, but the workflow kept getting in the way. What's changing now isn't just image quality, it's the creative process itself. This video shows how I used CapCut Design Studio to build an autumn fashion poster from scratch. I started on the infinite canvas to explore styles, then used the pencil to easily remove unwanted elements and even swap out outfits to find the perfect look for my model. Also, I used the text editing feature to instantly translate my poster into multiple languages without losing the original aesthetic. If I ever need more options, the 'One More' feature automatically creates multiple creative variations with one click. Instead of switching between multiple tools, I was able to keep the entire creative process connected in one place. AI design is moving from "generate and download" to "create, refine, and scale"—and that's the shift I'm most excited about. What's the biggest bottleneck in your current design workflow? #capcutdesignstudio #capcut
0:58

CapCut Design Studio changed the way I think about creating with AI. For a long time, creating with AI felt less like creating and more like managing files—generate, export, switch tools, edit, re-upload, repeat. The ideas were flowing, but the workflow kept getting in the way. What's changing now isn't just image quality, it's the creative process itself. This video shows how I used CapCut Design Studio to build an autumn fashion poster from scratch. I started on the infinite canvas to explore styles, then used the pencil to easily remove unwanted elements and even swap out outfits to find the perfect look for my model. Also, I used the text editing feature to instantly translate my poster into multiple languages without losing the original aesthetic. If I ever need more options, the 'One More' feature automatically creates multiple creative variations with one click. Instead of switching between multiple tools, I was able to keep the entire creative process connected in one place. AI design is moving from "generate and download" to "create, refine, and scale"—and that's the shift I'm most excited about. What's the biggest bottleneck in your current design workflow? #capcutdesignstudio #capcut

Farlane

41,971 görüntüleme • 2 ay önce

We've done it! 🐈 This is our tribute to and the entire Ordinals community. The OP_CAT_AR support is officially ON-CHAIN and accessible to everyone. The support covers all 3333 Ordinals, allowing you to obtain the 3D model of any cat from the collection [GLTF File], exportable to all 3D modeling software for creating animations to participate in QUANTUM CATS contests or just to pet your cats in “real life”. You can download the AR File and open it on your iPhone for real-time augmented reality with your cat 👁️Quick Guide: Visit: Choose the cat you want to export and enter the Quantum Cat ID. Click Export and choose the file type. Have fun animating your cat or creating augmented reality videos. Unleash your creativity and stand out for a chance to get on the QUANTUM CATS whitelist. Behind this experiment, there was a lot of sacrifice. We had to recreate the entire 3D collection and have an artist in our team design the features. Subsequently, we coded the AR support and integrated it ON-CHAIN, all within less than 5 days. This is our gift to the community🐱 Additionally, we're hosting a giveaway🎁 Share your AR experience with OP CATs [Quote this tweet with AR-Video] and tag , and use the hashtag #OP_CAT_AR. 🎉We'll be giving away various prizes For team #OP_CAT_IS_ALIVE: RAF edgerunner G.G.
0:32

We've done it! 🐈 This is our tribute to and the entire Ordinals community. The OP_CAT_AR support is officially ON-CHAIN and accessible to everyone. The support covers all 3333 Ordinals, allowing you to obtain the 3D model of any cat from the collection [GLTF File], exportable to all 3D modeling software for creating animations to participate in QUANTUM CATS contests or just to pet your cats in “real life”. You can download the AR File and open it on your iPhone for real-time augmented reality with your cat 👁️Quick Guide: Visit: Choose the cat you want to export and enter the Quantum Cat ID. Click Export and choose the file type. Have fun animating your cat or creating augmented reality videos. Unleash your creativity and stand out for a chance to get on the QUANTUM CATS whitelist. Behind this experiment, there was a lot of sacrifice. We had to recreate the entire 3D collection and have an artist in our team design the features. Subsequently, we coded the AR support and integrated it ON-CHAIN, all within less than 5 days. This is our gift to the community🐱 Additionally, we're hosting a giveaway🎁 Share your AR experience with OP CATs [Quote this tweet with AR-Video] and tag , and use the hashtag #OP_CAT_AR. 🎉We'll be giving away various prizes For team #OP_CAT_IS_ALIVE: RAF edgerunner G.G.

RAF

15,720 görüntüleme • 2 yıl önce

NEW RESEARCH: You can now create a new robot optimized for any given task! I love this new project by Huy Ha, Shuran Song, and others. Called "Transformer Transformer: A Unified Model for Motion-Conditioned Robot Co-design", it generates a robot's physical design and its controller together from a task spec. DEFINITIONS: - Reward function: A scoring rule that assigns a number to how well a behavior achieves the task. Here, it is the objective the generated design is pushed to maximize (e.g., track the target motion with low error). - Tokenizing: dividing continuous or structured data (a robot's links, joints, motor specs, states, actions) into a discrete vocabulary of symbols a transformer can process, the same step that turned pixels and audio into "language" for these models. - Diffusion transformer (DiT): A transformer trained to turn random noise into structured output through iterative denoising. Here, it generates robot bodies and trajectories instead of images. - MuJoCo: The standard fast physics simulator for robotics research (DeepMind-maintained). The Menagerie is its curated zoo of ready-to-use robot models. - CMA-ES: Covariance Matrix Adaptation Evolution Strategy, the workhorse black-box optimizer: it evolves a population of candidate designs, keeps the best, and needs thousands of simulator rollouts. - Bimanual multi-trajectory optimization: Finding one design/controller that performs well across several target motions for a two-armed robot at once, harder than optimizing for a single arm and a single motion. - BERT/MAE masked-modeling trick: Train one model to fill in whatever parts of the input you hide (words for BERT, image patches for MAE); at inference, choosing what to mask chooses the task, so masking the body makes it a designer and masking the actions makes it a controller. In practice, you give it a target end-effector motion and a reward function, and it outputs a complete embodiment (link, joint, motor, and inertial property), as well as a controller to drive it. It works by tokenizing both the body (links/joints/motors) and the dynamics (states/actions) into a compact scheme called RoboTokens, training a diffusion transformer (DiT) over them. The same model predicts dynamics using those predictions ("Dynamics Self-Guidance") to push generated designs toward higher reward at inference time. Masking different token types (using the BERT/MAE masked-modeling trick) lets the one model do three jobs: generate an embodiment, control an arbitrary embodiment, or design one conditioned on a motion. It is trained on 11 robots from the MuJoCo Menagerie (0.65 kg hand to 67.5 kg quadruped, 6–35 joints), and validated in sim and on a physical ALOHA doing cloth flinging. I like the fact that this approach inverts the entire recent robotics ideas: designing a policy for a fixed robot -> designing the robot for a fixed task. Every other approach assumes the body is given and learns a controller. Transformer Transformer takes the task (target motion + reward), then generates the body and controller jointly. In practice, it is a ~180× speedup over the standard optimizer at equal-or-better quality. It reaches "CMA-ES-level quality in seconds" and finishes bimanual multi-trajectory optimization in that is worth underlining nowadays! Also worth mentioning: this is the lab behind UMI and Handroid, that I mentioned here previously! The team seems extremely creative, i love these out-of-the-box approaches. Enjoy watching the demo of robot optimization in 3D, data acquisition, then real-life testing:
0:28

NEW RESEARCH: You can now create a new robot optimized for any given task! I love this new project by Huy Ha, Shuran Song, and others. Called "Transformer Transformer: A Unified Model for Motion-Conditioned Robot Co-design", it generates a robot's physical design and its controller together from a task spec. DEFINITIONS: - Reward function: A scoring rule that assigns a number to how well a behavior achieves the task. Here, it is the objective the generated design is pushed to maximize (e.g., track the target motion with low error). - Tokenizing: dividing continuous or structured data (a robot's links, joints, motor specs, states, actions) into a discrete vocabulary of symbols a transformer can process, the same step that turned pixels and audio into "language" for these models. - Diffusion transformer (DiT): A transformer trained to turn random noise into structured output through iterative denoising. Here, it generates robot bodies and trajectories instead of images. - MuJoCo: The standard fast physics simulator for robotics research (DeepMind-maintained). The Menagerie is its curated zoo of ready-to-use robot models. - CMA-ES: Covariance Matrix Adaptation Evolution Strategy, the workhorse black-box optimizer: it evolves a population of candidate designs, keeps the best, and needs thousands of simulator rollouts. - Bimanual multi-trajectory optimization: Finding one design/controller that performs well across several target motions for a two-armed robot at once, harder than optimizing for a single arm and a single motion. - BERT/MAE masked-modeling trick: Train one model to fill in whatever parts of the input you hide (words for BERT, image patches for MAE); at inference, choosing what to mask chooses the task, so masking the body makes it a designer and masking the actions makes it a controller. In practice, you give it a target end-effector motion and a reward function, and it outputs a complete embodiment (link, joint, motor, and inertial property), as well as a controller to drive it. It works by tokenizing both the body (links/joints/motors) and the dynamics (states/actions) into a compact scheme called RoboTokens, training a diffusion transformer (DiT) over them. The same model predicts dynamics using those predictions ("Dynamics Self-Guidance") to push generated designs toward higher reward at inference time. Masking different token types (using the BERT/MAE masked-modeling trick) lets the one model do three jobs: generate an embodiment, control an arbitrary embodiment, or design one conditioned on a motion. It is trained on 11 robots from the MuJoCo Menagerie (0.65 kg hand to 67.5 kg quadruped, 6–35 joints), and validated in sim and on a physical ALOHA doing cloth flinging. I like the fact that this approach inverts the entire recent robotics ideas: designing a policy for a fixed robot -> designing the robot for a fixed task. Every other approach assumes the body is given and learns a controller. Transformer Transformer takes the task (target motion + reward), then generates the body and controller jointly. In practice, it is a ~180× speedup over the standard optimizer at equal-or-better quality. It reaches "CMA-ES-level quality in seconds" and finishes bimanual multi-trajectory optimization in that is worth underlining nowadays! Also worth mentioning: this is the lab behind UMI and Handroid, that I mentioned here previously! The team seems extremely creative, i love these out-of-the-box approaches. Enjoy watching the demo of robot optimization in 3D, data acquisition, then real-life testing:

Léo

26,353 görüntüleme • 1 ay önce

🚨 Paper Alert 🚨 ➡️Paper Title: Articulate3D: Zero-Shot Text-Driven 3D Object Posing 🌟Few pointers from the paper 🎯Authors of this paper proposed a training-free method, “Articulate3D”, to pose a 3D asset through language control. 🎯Despite advances in vision and language models, this task remains surprisingly challenging. 🎯To achieve this goal, they decomposed the problem into two steps. 🎯They modified a powerful image-generator to create target images conditioned on the input image and a text instruction. 🎯They then align the mesh to the target images through a multi-view pose optimisation step. 🎯 In detail, they introduced a self-attention rewiring mechanism (RSActrl) that decouples the source structure from pose within an image generative model, allowing it to maintain a consistent structure across varying poses. 🎯They observed that differentiable rendering is an unreliable signal for articulation optimisation; instead, they used keypoints to establish correspondences between input and target images. 🎯The effectiveness of Articulate3D is demonstrated across a diverse range of 3D objects and free-form text prompts, successfully manipulating poses while maintaining the original identity of the mesh. 🎯Quantitative evaluations and a comparative user study, in which their method was preferred over 85% of the time, confirm its superiority over existing approaches. 🏢Organization: University of Oxford , Google DeepMind 🧙Paper Authors: Oishi Deb, Anjun Hu, Ashkan Khakzar, Philip Torr, Christian Rupprecht 📝 Read the Full Paper here: 🗂️ Project Page: 🎥 Be sure to watch the attached Demo Video - Sound on 🔊🔊 Find this Valuable 💎 ? ♻️QT and teach your network something new Follow me 👣, naveen manwani , for the latest updates on Tech and AI-related news, insightful research papers, and exciting announcements.
0:46

🚨 Paper Alert 🚨 ➡️Paper Title: Articulate3D: Zero-Shot Text-Driven 3D Object Posing 🌟Few pointers from the paper 🎯Authors of this paper proposed a training-free method, “Articulate3D”, to pose a 3D asset through language control. 🎯Despite advances in vision and language models, this task remains surprisingly challenging. 🎯To achieve this goal, they decomposed the problem into two steps. 🎯They modified a powerful image-generator to create target images conditioned on the input image and a text instruction. 🎯They then align the mesh to the target images through a multi-view pose optimisation step. 🎯 In detail, they introduced a self-attention rewiring mechanism (RSActrl) that decouples the source structure from pose within an image generative model, allowing it to maintain a consistent structure across varying poses. 🎯They observed that differentiable rendering is an unreliable signal for articulation optimisation; instead, they used keypoints to establish correspondences between input and target images. 🎯The effectiveness of Articulate3D is demonstrated across a diverse range of 3D objects and free-form text prompts, successfully manipulating poses while maintaining the original identity of the mesh. 🎯Quantitative evaluations and a comparative user study, in which their method was preferred over 85% of the time, confirm its superiority over existing approaches. 🏢Organization: University of Oxford , Google DeepMind 🧙Paper Authors: Oishi Deb, Anjun Hu, Ashkan Khakzar, Philip Torr, Christian Rupprecht 📝 Read the Full Paper here: 🗂️ Project Page: 🎥 Be sure to watch the attached Demo Video - Sound on 🔊🔊 Find this Valuable 💎 ? ♻️QT and teach your network something new Follow me 👣, naveen manwani , for the latest updates on Tech and AI-related news, insightful research papers, and exciting announcements.

naveen manwani

14,334 görüntüleme • 1 yıl önce

I’m thrilled to announce that we just released GraspGen, a multi-year project we have been cooking at NVIDIA Robotics 🚀 GraspGen: A Diffusion-Based Framework for 6-DOF Grasping Grasping is a foundational challenge in robotics 🤖 — whether for industrial picking or general-purpose humanoids. VLA + real data collection is all the rage now but is expensive and scales poorly for this task. For every new gripper and/or scene, you’ll have to recollect the dataset in this paradigm for the best perf. 💡Key Idea: Since grasping is such a well-defined task in simulation - why can’t we just scale synthetic data generation and train a generative model for grasping? By embracing modularity and standardized grasp formats, we can make this a turnkey technology that works zero-shot for multiple settings. GraspGen is a modular framework for diffusion-based 6-DOF grasp generation that scales across embodiment types, observability conditions, clutter, task complexity. Key Features: ✅ Multi-embodiment support: suction, parallel-jaw, and multi-fingered grippers ✅ Generalization to partial + complete 3D point clouds ✅ Generalization to single-objects + cluttered scenes ✅ Modular design uses other robotics modules and foundation models (SAM2, cuRobo, FoundationStereo, FoundationPose). This allows GraspGen to focus on only one thing - grasp generation ✅ Training recipe: grasp discriminator is trained with On-Generator data from the diffusion model - so that it learns to correct the mistakes (if any) of the diffusion generator ✅ Real-time performance (~20 Hz) before any GPU acceleration; low memory footprint 📊 Results: • SOTA on the FetchBench [Han et al. CoRL 2024] benchmark • Zero-shot sim-to-real transfer on unknown objects and cluttered scenes • Dataset of 53M simulated grasps across 8K objects from Objaverse 📄 arXiv: 🌐 Website: 💻 Code: A huge thank you to everyone involved in this journey — excited to see what the community builds on top of it! Joint work with Clemens Eppner , Balakumar Sundaralingam , Yu-Wei, Jun Yamada Wentao Yuan and other collaborators #robotics #diffusionmodels #physicalAI #simtoreal
0:39

I’m thrilled to announce that we just released GraspGen, a multi-year project we have been cooking at NVIDIA Robotics 🚀 GraspGen: A Diffusion-Based Framework for 6-DOF Grasping Grasping is a foundational challenge in robotics 🤖 — whether for industrial picking or general-purpose humanoids. VLA + real data collection is all the rage now but is expensive and scales poorly for this task. For every new gripper and/or scene, you’ll have to recollect the dataset in this paradigm for the best perf. 💡Key Idea: Since grasping is such a well-defined task in simulation - why can’t we just scale synthetic data generation and train a generative model for grasping? By embracing modularity and standardized grasp formats, we can make this a turnkey technology that works zero-shot for multiple settings. GraspGen is a modular framework for diffusion-based 6-DOF grasp generation that scales across embodiment types, observability conditions, clutter, task complexity. Key Features: ✅ Multi-embodiment support: suction, parallel-jaw, and multi-fingered grippers ✅ Generalization to partial + complete 3D point clouds ✅ Generalization to single-objects + cluttered scenes ✅ Modular design uses other robotics modules and foundation models (SAM2, cuRobo, FoundationStereo, FoundationPose). This allows GraspGen to focus on only one thing - grasp generation ✅ Training recipe: grasp discriminator is trained with On-Generator data from the diffusion model - so that it learns to correct the mistakes (if any) of the diffusion generator ✅ Real-time performance (~20 Hz) before any GPU acceleration; low memory footprint 📊 Results: • SOTA on the FetchBench [Han et al. CoRL 2024] benchmark • Zero-shot sim-to-real transfer on unknown objects and cluttered scenes • Dataset of 53M simulated grasps across 8K objects from Objaverse 📄 arXiv: 🌐 Website: 💻 Code: A huge thank you to everyone involved in this journey — excited to see what the community builds on top of it! Joint work with Clemens Eppner , Balakumar Sundaralingam , Yu-Wei, Jun Yamada Wentao Yuan and other collaborators #robotics #diffusionmodels #physicalAI #simtoreal

Adithya Murali

24,106 görüntüleme • 1 yıl önce

It's finally time for the massive update I know EVERYONE has been eagerly waiting for... Presenting to you over 6 straight minutes of spliced footage directly captured from my actual physical Sega Dreamcast... RUNNING THE LATEST BUILD OF OUR SONIC MANIA PORT!!! I can't even begin to describe to you how polished and fantastic it's starting to feel... as though it belonged on the Dreamcast all-along... Thanks to the hard work of jnmartin, sonicfreak94, and a few others, not only is the Dreamcast port nearing completion, but... wait for it... THE 3D STAGES THAT STRUGGLED ON EVERY UNOFFICIAL PORT ARE NOW ALL RUNNING FULLSPEED!!!! Trust me when I say that this took an absolutely ENORMOUS amount of effort from jnmartin and the crew to pull off, considering the low-end for this game was the original Nintendo Switch, and ports running on consoles with twice our processing power struggled to run these levels fullspeed. The first and most obvious thing is that the 3D software renderer was ditched and all rendering was done natively with our PowerVR GPU... which actually wasn't as simple as it sounds. The tilemaps for these levels have had to be tessellated into a bajillion PVR quads and transformed and rendered as individual polygons to look correct and run faster than a slideshow. Mr anonymous Ocarina of Time chad developer came up with a pretty slick LoD scheme for drawing tiles closest to the player at 1x1 pixel sizes with sizes increasing with distance up to 2x2 and 4x4 pixels, allowing them to reduce the overall number of tile vertices that have to be transformed by the SH4 CPU and submitted to the PVR GPU, by strategically keeping the majority of the polygon detail closest to the camera, with detail decreasing as the tiles get further away. Next, the 3D geometry was preprocessed and converted from being triangle-based to being triangle-strip based, drastically reducing the number of vertices per model that our 200Mhz SH4 CPU had to transform (with plain integer arithmetic and no FPU vectorization, since this is all slow-ass fixed-point integer math)! Finally, it was discovered that the lighting was a disproportionately significant contributor to the TnL load on the SH4 for processing vertices. Jnmartin came to the realization that 99% of the time only the Y axis direction is considered for lighting equation intensity with just a white color. So this simplified lighting model got baked into the renderer, which gave another round of gainz. So in the end, after all of this work came together, a draw distance of 75% the distance of the retail Sonic Mania versions was achieved for the decorations, plus the character models have their full geometry count and have not been simplified on a freaking Sega Dreamcast with a 200Mhz SH4 CPU and only 16MB of RAM! 🔥
6:32

It's finally time for the massive update I know EVERYONE has been eagerly waiting for... Presenting to you over 6 straight minutes of spliced footage directly captured from my actual physical Sega Dreamcast... RUNNING THE LATEST BUILD OF OUR SONIC MANIA PORT!!! I can't even begin to describe to you how polished and fantastic it's starting to feel... as though it belonged on the Dreamcast all-along... Thanks to the hard work of jnmartin, sonicfreak94, and a few others, not only is the Dreamcast port nearing completion, but... wait for it... THE 3D STAGES THAT STRUGGLED ON EVERY UNOFFICIAL PORT ARE NOW ALL RUNNING FULLSPEED!!!! Trust me when I say that this took an absolutely ENORMOUS amount of effort from jnmartin and the crew to pull off, considering the low-end for this game was the original Nintendo Switch, and ports running on consoles with twice our processing power struggled to run these levels fullspeed. The first and most obvious thing is that the 3D software renderer was ditched and all rendering was done natively with our PowerVR GPU... which actually wasn't as simple as it sounds. The tilemaps for these levels have had to be tessellated into a bajillion PVR quads and transformed and rendered as individual polygons to look correct and run faster than a slideshow. Mr anonymous Ocarina of Time chad developer came up with a pretty slick LoD scheme for drawing tiles closest to the player at 1x1 pixel sizes with sizes increasing with distance up to 2x2 and 4x4 pixels, allowing them to reduce the overall number of tile vertices that have to be transformed by the SH4 CPU and submitted to the PVR GPU, by strategically keeping the majority of the polygon detail closest to the camera, with detail decreasing as the tiles get further away. Next, the 3D geometry was preprocessed and converted from being triangle-based to being triangle-strip based, drastically reducing the number of vertices per model that our 200Mhz SH4 CPU had to transform (with plain integer arithmetic and no FPU vectorization, since this is all slow-ass fixed-point integer math)! Finally, it was discovered that the lighting was a disproportionately significant contributor to the TnL load on the SH4 for processing vertices. Jnmartin came to the realization that 99% of the time only the Y axis direction is considered for lighting equation intensity with just a white color. So this simplified lighting model got baked into the renderer, which gave another round of gainz. So in the end, after all of this work came together, a draw distance of 75% the distance of the retail Sonic Mania versions was achieved for the decorations, plus the character models have their full geometry count and have not been simplified on a freaking Sega Dreamcast with a 200Mhz SH4 CPU and only 16MB of RAM! 🔥

Falco Girgis

88,952 görüntüleme • 5 ay önce

The Concorde could fly London to New York in ~3.5 hours (half the time of a normal airliner). In operation from 1976 to 2003, the plane was a technical marvel and faster than the speed of sound. But it had poor business economics. Let’s start with the insane specs: ▫️Got as high as 10k miles ▫️Flew around earth in 30 hours ▫️Mach 2.04 speed (more than 2x speed of sound) ▫️Arrived in New York at earlier time than it left London (banker folk loved this) ▫️Delta wings and nose tip that could point down (so pilots could see runway, because it had to land at specific angle) Jointly run by British Airways and Air France, there were only 14 Concordes put to work (flew 50,000 total flights). What was wrong with the business model? ▫️SMALL MARKET: The thin aerodynamic design could only seat 109, so tickets were very expensive: $11k (a Boeing 747 can do >800 passengers) ▫️HUGE MAINTENANCE: Planes got so hot at the max height (110 degrees), that they expanded 30cm and sealant for fuel hardened. It took 28 hours to turnaround the plane (a normal one can do <2 hours). ▫️OUTRAGEOUS FUEL NEEDS: Each flight required 28,000 litres, for max 109 passengers. Commmercial planes needed 4x less fuel on a per passenger basis. ▫️NOISE RESTRICTIONS: Many cities wouldn’t allow the Concorde to fly because of how glass-shatteringly long it is on take-off, which limited destinations. *** Development of the Concorde — it had an “e” to placate the French — cost $2.8B and was paid for by the UK and French governments. The airlines were actually able to fly profitably for a number of years. But the economics were warped because the planes were given to them for free and the airlines didn’t have to capitalize costs. The business ultimately shuttered in 2003 following a tragic crash in 2000 and industry-wide slowdown post-9/11. Will we ever have sub-4 hour flights from London to NY again? A batch of sonic plane startups could make it happen by the end of this decade.
0:39

The Concorde could fly London to New York in ~3.5 hours (half the time of a normal airliner). In operation from 1976 to 2003, the plane was a technical marvel and faster than the speed of sound. But it had poor business economics. Let’s start with the insane specs: ▫️Got as high as 10k miles ▫️Flew around earth in 30 hours ▫️Mach 2.04 speed (more than 2x speed of sound) ▫️Arrived in New York at earlier time than it left London (banker folk loved this) ▫️Delta wings and nose tip that could point down (so pilots could see runway, because it had to land at specific angle) Jointly run by British Airways and Air France, there were only 14 Concordes put to work (flew 50,000 total flights). What was wrong with the business model? ▫️SMALL MARKET: The thin aerodynamic design could only seat 109, so tickets were very expensive: $11k (a Boeing 747 can do >800 passengers) ▫️HUGE MAINTENANCE: Planes got so hot at the max height (110 degrees), that they expanded 30cm and sealant for fuel hardened. It took 28 hours to turnaround the plane (a normal one can do <2 hours). ▫️OUTRAGEOUS FUEL NEEDS: Each flight required 28,000 litres, for max 109 passengers. Commmercial planes needed 4x less fuel on a per passenger basis. ▫️NOISE RESTRICTIONS: Many cities wouldn’t allow the Concorde to fly because of how glass-shatteringly long it is on take-off, which limited destinations. *** Development of the Concorde — it had an “e” to placate the French — cost $2.8B and was paid for by the UK and French governments. The airlines were actually able to fly profitably for a number of years. But the economics were warped because the planes were given to them for free and the airlines didn’t have to capitalize costs. The business ultimately shuttered in 2003 following a tragic crash in 2000 and industry-wide slowdown post-9/11. Will we ever have sub-4 hour flights from London to NY again? A batch of sonic plane startups could make it happen by the end of this decade.

Trung Phan

3,100,805 görüntüleme • 2 yıl önce

VALENTINES DAY IS A PAGAN HOLIDAY CALLED LUPERCALIA The first man named Valentine, was Lupercus (The hunter). Nimrod King of Babylon, worshipper of Baal, was known as the mighty Hunter and was the Romans Lupercus. Lupercalia is a blood ritual falling on 13, 14, and 15. Valentine's Day or Lupercalia is just another occult holyday celebrating Osiris, Ra, Moloch, Baal, Lucifer, Saturn, etc all which represent the occult esoteric mystery schools and Babylonian Mystery Religion sun worship. This is who they are, astrotheologists and today is one of their ritual Holydays. Keep your children close as sacrifice will be taking place it started yesterday will continue today at the peak and also tomorrow. Not to mention last night was also a full moon "Snow moon" Futhermore.... keep in mind King Nimrod built Babylon and commissioned the building of the tower of Babel to reach God and become God. "Valentines day is currently celebrated on February 14th. In ancient Babylonian times the solstice occurred on January 6th, Nimrod’s true birthday. It was the custom of antiquity for the mother of a male child to present herself for purification on the 40th day after the day of birth. The 40th day after January 6th is February 15th. The Lupercalia celebration began on the evening of February 14th hence our current holiday, St. Valentine’s Day. Nimrod was the original St. Valentine but he wasn’t called that during Lupercalia annual celebration. The name Valentine originated in Rome. Valentine was also known as Saturn (Satan), and Nimrod, the Roman Babylonian god who hid from his pursuers in a secret place. The Latin word Saturn is derived from the Semitic-speaking Babylonians. It means “Be hid,” “Hide self,” “Secret,” “Conceal.” The original Semitic (Hebrew) word, from which the Latin Saturn is derived, is used 83 times in the Old Testament. According to ancient tradition, Saturn (Nimrod) fled from his pursuers to Italy. The Apennine Mountains of Italy were anciently named the mountains of Nembrod, or Nimrod. Nimrod briefly hid out at the site where Rome was later built. The ancient name of Rome, before it was rebuilt in 753 B.C., was Saturnia, the site of Saturn’s (Nimrod’s) hiding. There he was found and slain for his crimes. Later, professing Christians in Constantine’s day made Nimrod the St. Valentine of the heathen, “A Saint of the Church,” and continued to honor him under the name of a Christian martyr. Valentine was a common Roman name. The parents of Rome often gave this name to their children in honor of the famous man who was first called Valentine in antiquity. That famous man was Lupercus, “The Hunter” (Nimrod). The Greeks called Lupercus, “Pan.” The Semites called Pan, “Baal,” according to “Classical Dictionaries.” Baal is often mentioned in the Bible and this was merely another name for Nimrod, “The mighty Hunter” (Genesis 10:9). The hunter, Nimrod was the Lupercus, or wolf hunter of the Romans. And St. Valentine’s Day was originally a day set aside by the pagans in his honor. Valentine comes from the Latin word “Valentinus,” a proper name derived from the word “Valens,” meaning, “To be strong,” as illustrated in the Webster’s Unabridged Dictionary. It means strong, powerful, mighty. Another name for Nimrod was “Sanctus,” and “Santa.” The Romans acquired the symbol of the heart from the Babylonians. In the Babylonian language for “Heart” was “Bal.” The heart, or “Bal,” was merely a symbol of Nimrod the “Baal” (the name of Satan), or Lord of the Babylonians. Another Nimrod name as a child, was “Cupid,” meaning desire (Encyclopedia Britannica, art., “Cupid”). It is said that when Nimrod’s mother saw him, she lusted after him, and desired him. Nimrod became her “Cupid,” her desired one. As Nimrod grew up, he became the child-hero of many women who desired him. He was their Cupid. In the Book of Daniel he is called the “Desire of women” (Daniel 11:37). Moffatt translates the word as Tammuz, a Babylonian name of Nimrod. He provoked so many women to jealousy that an idol of him was often called the “Idol of jealousy” (Ezekiel 8:5). The pagans commemorated their hero-hunter Nimrod, or Baal, by sending heart shaped love tokens to one another on the evening of February 14th as a symbol of him. The heart shape that we use today is given out on Valentines Day in forms of chocolate and candy, because it’s considered a day of love and lovemaking. The shape of this heart is not in the shape of a human heart. It derives from an extinct plant called, “Silphium.” Silphium was the most effective birth control of its time. It was so popular that its shape was used on Greek coins." Even more interesting.... Lupercalia = Wolf Festival (literally) Lupercalia comes from Lupercus, an old Italic god. Lupercus = “he who wards off wolves” or “wolf god.” The festival was held at the Lupercal Cave, where: Romulus and Remus were suckled by a she-wolf (lupa) Rome’s origin story is wolf-raised So from the jump: Rome = founded by wolves Lupercalia = ritual honoring wolf power The Luperci: Men Who Became Wolves (Symbolically) The priests of the festival were called Luperci. During the rite they: Sacrificed goats and a dog (both linked to wildness and liminality), smeared blood on their foreheads Laughed it off, ritual rebirth, stripped nearly naked and ran through the city whipping people with goat hides This matters because in Indo-European cultures: Nudity + blood + animal skins = transformation rites These are the same elements later associated with berserkers and werewolves. The Luperci weren’t pretending to be wolves. They were ritually assuming wolf nature. Lupus, Lupa, Lycanthropy. Same Root Family Latin lupus = wolf Greek lykos = wolf Lycanthropy = lykos (wolf) + anthropos (man) Roman writers knew this connection. Pliny the Elder records Arcadian wolf-men legends where: A man eats human flesh during a ritual. Becomes a wolf for 9 years. Can return to human form if he abstains. Lupercalia happened mid-February, a dangerous threshold: Winter to Spring Death to Fertility Order to Chaos Werewolves always appear in liminal spaces: - Full moons - Forest edges - Night - Seasonal transitions Same psychology. Same symbolism. When Christianity absorbed Rome: Lupercalia was banned. Wolf symbolism was flipped from sacred protector to demonic shapeshifter. Saint Augustine and later medieval clergy reframed wolf-men as servants of Satan and turned ritual transformation into curses. Now research: Dogman Sources: Video credit: A call for Vengence Youtube
12:02

VALENTINES DAY IS A PAGAN HOLIDAY CALLED LUPERCALIA The first man named Valentine, was Lupercus (The hunter). Nimrod King of Babylon, worshipper of Baal, was known as the mighty Hunter and was the Romans Lupercus. Lupercalia is a blood ritual falling on 13, 14, and 15. Valentine's Day or Lupercalia is just another occult holyday celebrating Osiris, Ra, Moloch, Baal, Lucifer, Saturn, etc all which represent the occult esoteric mystery schools and Babylonian Mystery Religion sun worship. This is who they are, astrotheologists and today is one of their ritual Holydays. Keep your children close as sacrifice will be taking place it started yesterday will continue today at the peak and also tomorrow. Not to mention last night was also a full moon "Snow moon" Futhermore.... keep in mind King Nimrod built Babylon and commissioned the building of the tower of Babel to reach God and become God. "Valentines day is currently celebrated on February 14th. In ancient Babylonian times the solstice occurred on January 6th, Nimrod’s true birthday. It was the custom of antiquity for the mother of a male child to present herself for purification on the 40th day after the day of birth. The 40th day after January 6th is February 15th. The Lupercalia celebration began on the evening of February 14th hence our current holiday, St. Valentine’s Day. Nimrod was the original St. Valentine but he wasn’t called that during Lupercalia annual celebration. The name Valentine originated in Rome. Valentine was also known as Saturn (Satan), and Nimrod, the Roman Babylonian god who hid from his pursuers in a secret place. The Latin word Saturn is derived from the Semitic-speaking Babylonians. It means “Be hid,” “Hide self,” “Secret,” “Conceal.” The original Semitic (Hebrew) word, from which the Latin Saturn is derived, is used 83 times in the Old Testament. According to ancient tradition, Saturn (Nimrod) fled from his pursuers to Italy. The Apennine Mountains of Italy were anciently named the mountains of Nembrod, or Nimrod. Nimrod briefly hid out at the site where Rome was later built. The ancient name of Rome, before it was rebuilt in 753 B.C., was Saturnia, the site of Saturn’s (Nimrod’s) hiding. There he was found and slain for his crimes. Later, professing Christians in Constantine’s day made Nimrod the St. Valentine of the heathen, “A Saint of the Church,” and continued to honor him under the name of a Christian martyr. Valentine was a common Roman name. The parents of Rome often gave this name to their children in honor of the famous man who was first called Valentine in antiquity. That famous man was Lupercus, “The Hunter” (Nimrod). The Greeks called Lupercus, “Pan.” The Semites called Pan, “Baal,” according to “Classical Dictionaries.” Baal is often mentioned in the Bible and this was merely another name for Nimrod, “The mighty Hunter” (Genesis 10:9). The hunter, Nimrod was the Lupercus, or wolf hunter of the Romans. And St. Valentine’s Day was originally a day set aside by the pagans in his honor. Valentine comes from the Latin word “Valentinus,” a proper name derived from the word “Valens,” meaning, “To be strong,” as illustrated in the Webster’s Unabridged Dictionary. It means strong, powerful, mighty. Another name for Nimrod was “Sanctus,” and “Santa.” The Romans acquired the symbol of the heart from the Babylonians. In the Babylonian language for “Heart” was “Bal.” The heart, or “Bal,” was merely a symbol of Nimrod the “Baal” (the name of Satan), or Lord of the Babylonians. Another Nimrod name as a child, was “Cupid,” meaning desire (Encyclopedia Britannica, art., “Cupid”). It is said that when Nimrod’s mother saw him, she lusted after him, and desired him. Nimrod became her “Cupid,” her desired one. As Nimrod grew up, he became the child-hero of many women who desired him. He was their Cupid. In the Book of Daniel he is called the “Desire of women” (Daniel 11:37). Moffatt translates the word as Tammuz, a Babylonian name of Nimrod. He provoked so many women to jealousy that an idol of him was often called the “Idol of jealousy” (Ezekiel 8:5). The pagans commemorated their hero-hunter Nimrod, or Baal, by sending heart shaped love tokens to one another on the evening of February 14th as a symbol of him. The heart shape that we use today is given out on Valentines Day in forms of chocolate and candy, because it’s considered a day of love and lovemaking. The shape of this heart is not in the shape of a human heart. It derives from an extinct plant called, “Silphium.” Silphium was the most effective birth control of its time. It was so popular that its shape was used on Greek coins." Even more interesting.... Lupercalia = Wolf Festival (literally) Lupercalia comes from Lupercus, an old Italic god. Lupercus = “he who wards off wolves” or “wolf god.” The festival was held at the Lupercal Cave, where: Romulus and Remus were suckled by a she-wolf (lupa) Rome’s origin story is wolf-raised So from the jump: Rome = founded by wolves Lupercalia = ritual honoring wolf power The Luperci: Men Who Became Wolves (Symbolically) The priests of the festival were called Luperci. During the rite they: Sacrificed goats and a dog (both linked to wildness and liminality), smeared blood on their foreheads Laughed it off, ritual rebirth, stripped nearly naked and ran through the city whipping people with goat hides This matters because in Indo-European cultures: Nudity + blood + animal skins = transformation rites These are the same elements later associated with berserkers and werewolves. The Luperci weren’t pretending to be wolves. They were ritually assuming wolf nature. Lupus, Lupa, Lycanthropy. Same Root Family Latin lupus = wolf Greek lykos = wolf Lycanthropy = lykos (wolf) + anthropos (man) Roman writers knew this connection. Pliny the Elder records Arcadian wolf-men legends where: A man eats human flesh during a ritual. Becomes a wolf for 9 years. Can return to human form if he abstains. Lupercalia happened mid-February, a dangerous threshold: Winter to Spring Death to Fertility Order to Chaos Werewolves always appear in liminal spaces: - Full moons - Forest edges - Night - Seasonal transitions Same psychology. Same symbolism. When Christianity absorbed Rome: Lupercalia was banned. Wolf symbolism was flipped from sacred protector to demonic shapeshifter. Saint Augustine and later medieval clergy reframed wolf-men as servants of Satan and turned ritual transformation into curses. Now research: Dogman Sources: Video credit: A call for Vengence Youtube

Redpill Drifter

1,802,645 görüntüleme • 6 ay önce

** Sega Genesis 3D Engine Update 8 ** Significant improvements all round as you can see and hear from the last update !! Foremost - A huge thanks to Toni Gálvez - Megastyle - BG. who has joined the project to create a bit of 16bit low poly magic. Toni's an Amiga fan but also crazy about game dev in general, he's worked on GBC, GBA, PC, MD, PSP, C64, CPC, MSX... and others. Gaming titles include War Times, Metal Gear, Rocketman, Tintin & Asterix to name a few. He's provided the great new ship model you see on screen - new striped buildings, all the backgrounds / palettes etc. There's a lot of models he's given me which need to be added, also he will be planning a lot of the level design. Very happy to have him help me turn this into something more than a tech demo as I have my hands tied pushing the MD as far as it can go haha - there is no cpu cycle to be spared. Also many thanks to my good friend CYBERDEOUS - Crouzet Laurent for the Music for this showing , I wanted to have the music load occurring so we have a realistic benchmark for performance and he was only too obliging. If you're into MD chiptunes check him out !! Since last update : New player model , substantially more detailed than the Arwing. Last update had a 23 triangle Arwing , this update has a 39 triangle custom model from Toni. We had several to choose from , others will be used for enemies . 3D Buffer size increased 25% to 256x160. This was quite tricky as I'm close to the DMA limit even with an extended vblank . Spent a few days thinking of how to do this as like anything retro every solution has a drawback, finally got a workable solution. It makes a big difference to have a bit more vertical height . Z Rotation added ( the screen tilting left to right ) , small hit to vertex transform on cpu thanks to look up tables doing the heavy lifting, saving 4 multiplies per vertex. Multiple speed ups in rendering code. Onscreen paths with no range checking used until Z is close enough to cause clipping , partial onscreen drawing pathes that need to check boundaries, quad rendering completely rewritten - was very very painfull to get right . I found out the hard way that things are great when they are not rotating in the Z axis haha . Partial buffer draw optimisations - which have helped with the massive dma load , sending up to a 20kb buffer in a single frame needs a lot of optimisation. Min / Max tile lines are analysed and only sent if dirtied , reducing most buffer swaps substantially. Still some issues to sort out , at times you can see the flicker near top of screen when frames are near full height . I need to optimise that a bit. Due to the onscreen buffer system a full Sprite background had to be implemented almost Neo Geo style. This flips the usual MD rendering system on its head as it uses both foreground and background layers for a foreground 3d plane and sprites for the background. This presents a few issues, one is to get a tilt effect on the background by using narrow sprites (16x32) we run out of sprites when trying to cover the screen. Thankfully the MD is not limited to 80 sprites, to fix this a 114 sprite multiplexor is used to draw the background, its completely made up of 16x32 sprites ! Why do things this way ? speed . Its the interleaved foreground/background layers that allow a double buffered ram system writing to write to vram using dma in a completely linear fashion - virtually no tile translation needed. The negative is you have no planes for the background, that's where the sprites come in . Thanks to H40 mode we still have a few sprites we can use for effects in the forground also . Thankfully we can implement a fairly good tilt still for the background using sprites, in future updates this will be able to move horizontally also and a bit of vertical movement. XGM1 music driver in use to simulate music cpu load, XGM2 unfortunately with the massive DMA needed to shift the 3d buffers would slow down at times rendering it unusable, XGM1 plays at full speed - albiet with a bit more of a cpu hit. Together with the sprite multiplexor and the music driver active theres a 10 % hit to cpu so I've had to play around with draw distances / object heights and other optimisations to offset that. Not to mention the larger buffer takes more cpu to fill also. Everything is placeholder so will be changed with proper stage design. We are averaging 20 FPS in the current video, I'll push for more as always !! Progress continues on my other projects , updates soon on those - retirement can't come quick enough . #SGDK #SegaGenesis #SegaMegadrive
1:17

** Sega Genesis 3D Engine Update 8 ** Significant improvements all round as you can see and hear from the last update !! Foremost - A huge thanks to Toni Gálvez - Megastyle - BG. who has joined the project to create a bit of 16bit low poly magic. Toni's an Amiga fan but also crazy about game dev in general, he's worked on GBC, GBA, PC, MD, PSP, C64, CPC, MSX... and others. Gaming titles include War Times, Metal Gear, Rocketman, Tintin & Asterix to name a few. He's provided the great new ship model you see on screen - new striped buildings, all the backgrounds / palettes etc. There's a lot of models he's given me which need to be added, also he will be planning a lot of the level design. Very happy to have him help me turn this into something more than a tech demo as I have my hands tied pushing the MD as far as it can go haha - there is no cpu cycle to be spared. Also many thanks to my good friend CYBERDEOUS - Crouzet Laurent for the Music for this showing , I wanted to have the music load occurring so we have a realistic benchmark for performance and he was only too obliging. If you're into MD chiptunes check him out !! Since last update : New player model , substantially more detailed than the Arwing. Last update had a 23 triangle Arwing , this update has a 39 triangle custom model from Toni. We had several to choose from , others will be used for enemies . 3D Buffer size increased 25% to 256x160. This was quite tricky as I'm close to the DMA limit even with an extended vblank . Spent a few days thinking of how to do this as like anything retro every solution has a drawback, finally got a workable solution. It makes a big difference to have a bit more vertical height . Z Rotation added ( the screen tilting left to right ) , small hit to vertex transform on cpu thanks to look up tables doing the heavy lifting, saving 4 multiplies per vertex. Multiple speed ups in rendering code. Onscreen paths with no range checking used until Z is close enough to cause clipping , partial onscreen drawing pathes that need to check boundaries, quad rendering completely rewritten - was very very painfull to get right . I found out the hard way that things are great when they are not rotating in the Z axis haha . Partial buffer draw optimisations - which have helped with the massive dma load , sending up to a 20kb buffer in a single frame needs a lot of optimisation. Min / Max tile lines are analysed and only sent if dirtied , reducing most buffer swaps substantially. Still some issues to sort out , at times you can see the flicker near top of screen when frames are near full height . I need to optimise that a bit. Due to the onscreen buffer system a full Sprite background had to be implemented almost Neo Geo style. This flips the usual MD rendering system on its head as it uses both foreground and background layers for a foreground 3d plane and sprites for the background. This presents a few issues, one is to get a tilt effect on the background by using narrow sprites (16x32) we run out of sprites when trying to cover the screen. Thankfully the MD is not limited to 80 sprites, to fix this a 114 sprite multiplexor is used to draw the background, its completely made up of 16x32 sprites ! Why do things this way ? speed . Its the interleaved foreground/background layers that allow a double buffered ram system writing to write to vram using dma in a completely linear fashion - virtually no tile translation needed. The negative is you have no planes for the background, that's where the sprites come in . Thanks to H40 mode we still have a few sprites we can use for effects in the forground also . Thankfully we can implement a fairly good tilt still for the background using sprites, in future updates this will be able to move horizontally also and a bit of vertical movement. XGM1 music driver in use to simulate music cpu load, XGM2 unfortunately with the massive DMA needed to shift the 3d buffers would slow down at times rendering it unusable, XGM1 plays at full speed - albiet with a bit more of a cpu hit. Together with the sprite multiplexor and the music driver active theres a 10 % hit to cpu so I've had to play around with draw distances / object heights and other optimisations to offset that. Not to mention the larger buffer takes more cpu to fill also. Everything is placeholder so will be changed with proper stage design. We are averaging 20 FPS in the current video, I'll push for more as always !! Progress continues on my other projects , updates soon on those - retirement can't come quick enough . #SGDK #SegaGenesis #SegaMegadrive

Shannon Birt

34,524 görüntüleme • 1 ay önce

kimi k3 vs gpt 5.6 sol vs fable 5 vs grok 4.5 Kimi.ai just dropped kimi k3 – a 2.8t param native multimodal model, the first open 3t-class release. key facts: • 1m token context. stable latentmoe activating 16 of 896 experts, built on kimi delta attention (kda) and attention residuals • quantization-aware training from the sft stage onward – mxfp4 weights, mxfp8 activations. moonshot claims ~2.5x scaling efficiency over k2 • max thinking effort by default. low- and high-effort modes are "coming in updates" – there is no way to turn the thinking down today, and you feel it in every run • pricing: $0.30/mtok cache-hit input, $3.00/mtok cache-miss, $15.00/mtok output. claims >90% cache hit rate on coding workloads • benchmarks: swe marathon 42.0 (1st – fable 5: 35.0, sol: 39.0, opus 4.8: 40.0), terminal bench 2.1 88.3, browsecomp 91.2 (1st), program bench 77.8 (1st), gpqa-diamond 93.5. loses frontierswe 81.2 vs fable's 86.6, and deepswe 67.5 vs sol's 73.0 our test – 3 prompts, single-file html, Three.js, fully procedural, no assets: 1. photorealistic european roulette wheel – 37 pockets in the real sequence, mahogany clearcoat bowl, chrome turret, diamond deflectors, flick-to-spin, ball that spirals inward and settles on a mathematically real number 2. las vegas slot machine – 3 reels behind transmissive glass, drag the chrome lever to play, mechanical odometer counters modelled in 3d, coin physics on win 3. full pinball table – 6.5° tilted playfield, flipper impulse physics, spline ramps, drop targets, 6 bumpers, mechanical score reels in the backbox we ran the test on AI/ML API platform results: - cost #1 grok 4.5 – $0.30 #2 kimi k3 – $0.71 #3 gpt 5.6 sol – $2.05 #4 fable 5 – $7.69 - tokens #1 grok 4.5 – 34,241 #2 gpt 5.6 sol – 51,748 #3 fable 5 – 144,126 #4 kimi k3 – 157,999 - lines of code #1 gpt 5.6 sol – 3,054 #2 grok 4.5 – 3,047 #3 kimi k3 – 2,255 #4 fable 5 – 1,950 - generation time #1 grok 4.5 – 5.1 min #2 gpt 5.6 sol – 22.0 min #3 fable 5 – 31.5 min #4 kimi k3 – 75.6 min observations: • kimi k3 is cheap and it is slow. 75.6 minutes across three prompts against grok's 5.1. it is 2.4x grok's price and 15x grok's wall clock. the roulette took 15 min, the slot 18, the pinball 42 • it failed 2 of 3. only the roulette works. the slot machine has reel cutouts on both faces of the cabinet and the symbols face backwards – you can only read your spin by walking around to the rear of the machine. the pinball table stands vertically on its edge with the legs floating detached beside it. • 81% of kimi's output tokens are reasoning, not code. grok: 22%. you are not paying for a bigger answer, you are paying for a longer argument with itself • price per 100 shipped lines – grok $0.010, kimi $0.031, sol $0.067, fable $0.394. a 39x spread for the same three files kimi k3's code quality: upsides: • the roulette is genuinely good – procedural wood grain with real specular breakup, correct european sequence (0-32-15-19-4...), chrome turret, diamond deflectors, clean console • the pinball artwork is the best in the test – a synthwave "nova strike / deep space" field with six individually coloured neon bumper rings, a retro sun on a grid horizon, a nova burst, and a scoring legend printed on the apron. no other model printed the rules on the machine. it is a beautiful texture on a broken object • physics reasoning is real – it derived a 480hz substep for the collider, worked out ball settle conditions and termination guarantees, and checked every ramp exit vector by hand before writing any of it • it is the only model that saw the importmap trap coming. sol shipped a blank white page twice because three.js addons import the bare specifier 'three' and die without an import map downsides: • it dodged that trap on the slot by loading three.js r128 through classic script tags – a 2021 build with no working transmission. its slot glass rendered fully opaque and buried all three reels behind a white pane. the code asks for transmission: 0.93, ior: 1.5 – correct, and silently ignored by a renderer that predates the feature • after 42 minutes and 212k characters of reasoning, the pinball cabinet is not assembled. the table stands vertically on its edge like a wardrobe – the prompt asked for 6.5° from horizontal, it delivered 90°. the legs float detached in the void beside it. head-on it photographs beautifully; orbit ten degrees and it is a painted slab with four chrome rods hovering nearby • the playfield z-fights with the glass – hard black banding across the whole field as soon as you pull the camera back a note on the pinball, in fairness to kimi: nobody passed it. every model shipped broken ball physics and controls you cannot trust. it is the hardest prompt we have run and the whole field failed it, each in its own way kimi k3 reasons better than anything else here and it shows exactly where reasoning pays – physics constants, sequences, edge cases, traps the others walked into follow thehype. for 24/7 ai news, analysis and breakdowns
0:36

kimi k3 vs gpt 5.6 sol vs fable 5 vs grok 4.5 Kimi.ai just dropped kimi k3 – a 2.8t param native multimodal model, the first open 3t-class release. key facts: • 1m token context. stable latentmoe activating 16 of 896 experts, built on kimi delta attention (kda) and attention residuals • quantization-aware training from the sft stage onward – mxfp4 weights, mxfp8 activations. moonshot claims ~2.5x scaling efficiency over k2 • max thinking effort by default. low- and high-effort modes are "coming in updates" – there is no way to turn the thinking down today, and you feel it in every run • pricing: $0.30/mtok cache-hit input, $3.00/mtok cache-miss, $15.00/mtok output. claims >90% cache hit rate on coding workloads • benchmarks: swe marathon 42.0 (1st – fable 5: 35.0, sol: 39.0, opus 4.8: 40.0), terminal bench 2.1 88.3, browsecomp 91.2 (1st), program bench 77.8 (1st), gpqa-diamond 93.5. loses frontierswe 81.2 vs fable's 86.6, and deepswe 67.5 vs sol's 73.0 our test – 3 prompts, single-file html, Three.js, fully procedural, no assets: 1. photorealistic european roulette wheel – 37 pockets in the real sequence, mahogany clearcoat bowl, chrome turret, diamond deflectors, flick-to-spin, ball that spirals inward and settles on a mathematically real number 2. las vegas slot machine – 3 reels behind transmissive glass, drag the chrome lever to play, mechanical odometer counters modelled in 3d, coin physics on win 3. full pinball table – 6.5° tilted playfield, flipper impulse physics, spline ramps, drop targets, 6 bumpers, mechanical score reels in the backbox we ran the test on AI/ML API platform results: - cost #1 grok 4.5 – $0.30 #2 kimi k3 – $0.71 #3 gpt 5.6 sol – $2.05 #4 fable 5 – $7.69 - tokens #1 grok 4.5 – 34,241 #2 gpt 5.6 sol – 51,748 #3 fable 5 – 144,126 #4 kimi k3 – 157,999 - lines of code #1 gpt 5.6 sol – 3,054 #2 grok 4.5 – 3,047 #3 kimi k3 – 2,255 #4 fable 5 – 1,950 - generation time #1 grok 4.5 – 5.1 min #2 gpt 5.6 sol – 22.0 min #3 fable 5 – 31.5 min #4 kimi k3 – 75.6 min observations: • kimi k3 is cheap and it is slow. 75.6 minutes across three prompts against grok's 5.1. it is 2.4x grok's price and 15x grok's wall clock. the roulette took 15 min, the slot 18, the pinball 42 • it failed 2 of 3. only the roulette works. the slot machine has reel cutouts on both faces of the cabinet and the symbols face backwards – you can only read your spin by walking around to the rear of the machine. the pinball table stands vertically on its edge with the legs floating detached beside it. • 81% of kimi's output tokens are reasoning, not code. grok: 22%. you are not paying for a bigger answer, you are paying for a longer argument with itself • price per 100 shipped lines – grok $0.010, kimi $0.031, sol $0.067, fable $0.394. a 39x spread for the same three files kimi k3's code quality: upsides: • the roulette is genuinely good – procedural wood grain with real specular breakup, correct european sequence (0-32-15-19-4...), chrome turret, diamond deflectors, clean console • the pinball artwork is the best in the test – a synthwave "nova strike / deep space" field with six individually coloured neon bumper rings, a retro sun on a grid horizon, a nova burst, and a scoring legend printed on the apron. no other model printed the rules on the machine. it is a beautiful texture on a broken object • physics reasoning is real – it derived a 480hz substep for the collider, worked out ball settle conditions and termination guarantees, and checked every ramp exit vector by hand before writing any of it • it is the only model that saw the importmap trap coming. sol shipped a blank white page twice because three.js addons import the bare specifier 'three' and die without an import map downsides: • it dodged that trap on the slot by loading three.js r128 through classic script tags – a 2021 build with no working transmission. its slot glass rendered fully opaque and buried all three reels behind a white pane. the code asks for transmission: 0.93, ior: 1.5 – correct, and silently ignored by a renderer that predates the feature • after 42 minutes and 212k characters of reasoning, the pinball cabinet is not assembled. the table stands vertically on its edge like a wardrobe – the prompt asked for 6.5° from horizontal, it delivered 90°. the legs float detached in the void beside it. head-on it photographs beautifully; orbit ten degrees and it is a painted slab with four chrome rods hovering nearby • the playfield z-fights with the glass – hard black banding across the whole field as soon as you pull the camera back a note on the pinball, in fairness to kimi: nobody passed it. every model shipped broken ball physics and controls you cannot trust. it is the hardest prompt we have run and the whole field failed it, each in its own way kimi k3 reasons better than anything else here and it shows exactly where reasoning pays – physics constants, sequences, edge cases, traps the others walked into follow thehype. for 24/7 ai news, analysis and breakdowns

thehype.

2,183,762 görüntüleme • 1 ay önce

Can we compile matter - for instance, a pine cone - and derive new active materials, end-to-end from observation to manufacturing? If physical systems can be formalized as composable mathematics, we can point AI that has been shown to resolve long-open mathematical problems at matter itself. Our new work turns bioinspired engineering from analogy into formal compilation: biology and mechanics become explicit, checkable, and executable, so AI reasoning can produce physical designs. This is the first end-to-end demonstration in which a formally compositional multiscale model is carried from a biological hierarchy, through engineered design and fabrication specification, to executable manufacturing code - and then to a physically tested artifact. Background: Humans have long been inspired by biology to advance technology, but this has usually been an ad hoc process rather than a mathematically rigorous one. Natural materials such as pinecones achieve adaptive behavior through mechanisms organized across many scales. Engineering typically translates those mechanisms by analogy: identify a biological principle, build something inspired by it, and validate each new design as a separate case. This can produce remarkable results, but the knowledge does not readily compound. Instead, we represent each scale as a dynamical module with explicit states, stimuli, governing laws, and interfaces. Every scale-to-scale map must preserve the stimulus - response dynamics: evolve the fine-scale system and then map upward, or map upward first and then evolve. The two paths must agree. Because this condition is preserved under composition, locally valid interfaces remain consistent when assembled into the full hierarchy. We then carry that structure into an engineered system, translate the target behavior into a verified fabrication specification, and compile it into G-code: the toolpaths, deposition sequence, temperatures, speeds, and other commands executed by a 3D printer. The intermediate translations are explicit, checkable, and executable rather than completed through an ad hoc handoff. The formal guarantee is that given valid local models and interfaces, their composition remains valid. Whether those models and manufacturing assumptions accurately capture physical reality remains an empirical question. That is why we fabricated and tested the results. We generated four actuator classes by crossing two stimuli - humidity and heat - with two responses: bending and twisting. The fourth, thermal twisting, required no new pipeline and no separate derivation within the framework. It emerged by composing a thermal stimulus module already validated in one case with a twisting module validated in another. The generated G-code produced the intended motion without manual redesign, and all four predictions fell within one experimental standard deviation of the measured response. Why this matters: 1⃣For AI in science, this provides a physics-aware type system against which generative proposals can be checked - and rejected at the interface - before expensive simulation, fabrication, or experiment. It is roughly analogous to proof checking, but for the composition of physical mechanisms. 2⃣For engineering, the accessible design space can scale with a library of validated components rather than with the number of individually derived cases. 3⃣The mathematics, category theory, carries all the way into a physical object on a print bed. This points toward scientific knowledge as executable infrastructure: models that are not only described in papers, but typed, composable, verifiable, and able to compile into experiments. Excellent work led by my student Lee Marom with Skylar Tibbits & Gioele Zardini. Paper published in J. Mech. Phys. Solids along with code, Grasshopper scripts, and manufacturing G-code below.
1:21

Can we compile matter - for instance, a pine cone - and derive new active materials, end-to-end from observation to manufacturing? If physical systems can be formalized as composable mathematics, we can point AI that has been shown to resolve long-open mathematical problems at matter itself. Our new work turns bioinspired engineering from analogy into formal compilation: biology and mechanics become explicit, checkable, and executable, so AI reasoning can produce physical designs. This is the first end-to-end demonstration in which a formally compositional multiscale model is carried from a biological hierarchy, through engineered design and fabrication specification, to executable manufacturing code - and then to a physically tested artifact. Background: Humans have long been inspired by biology to advance technology, but this has usually been an ad hoc process rather than a mathematically rigorous one. Natural materials such as pinecones achieve adaptive behavior through mechanisms organized across many scales. Engineering typically translates those mechanisms by analogy: identify a biological principle, build something inspired by it, and validate each new design as a separate case. This can produce remarkable results, but the knowledge does not readily compound. Instead, we represent each scale as a dynamical module with explicit states, stimuli, governing laws, and interfaces. Every scale-to-scale map must preserve the stimulus - response dynamics: evolve the fine-scale system and then map upward, or map upward first and then evolve. The two paths must agree. Because this condition is preserved under composition, locally valid interfaces remain consistent when assembled into the full hierarchy. We then carry that structure into an engineered system, translate the target behavior into a verified fabrication specification, and compile it into G-code: the toolpaths, deposition sequence, temperatures, speeds, and other commands executed by a 3D printer. The intermediate translations are explicit, checkable, and executable rather than completed through an ad hoc handoff. The formal guarantee is that given valid local models and interfaces, their composition remains valid. Whether those models and manufacturing assumptions accurately capture physical reality remains an empirical question. That is why we fabricated and tested the results. We generated four actuator classes by crossing two stimuli - humidity and heat - with two responses: bending and twisting. The fourth, thermal twisting, required no new pipeline and no separate derivation within the framework. It emerged by composing a thermal stimulus module already validated in one case with a twisting module validated in another. The generated G-code produced the intended motion without manual redesign, and all four predictions fell within one experimental standard deviation of the measured response. Why this matters: 1⃣For AI in science, this provides a physics-aware type system against which generative proposals can be checked - and rejected at the interface - before expensive simulation, fabrication, or experiment. It is roughly analogous to proof checking, but for the composition of physical mechanisms. 2⃣For engineering, the accessible design space can scale with a library of validated components rather than with the number of individually derived cases. 3⃣The mathematics, category theory, carries all the way into a physical object on a print bed. This points toward scientific knowledge as executable infrastructure: models that are not only described in papers, but typed, composable, verifiable, and able to compile into experiments. Excellent work led by my student Lee Marom with Skylar Tibbits & Gioele Zardini. Paper published in J. Mech. Phys. Solids along with code, Grasshopper scripts, and manufacturing G-code below.

Markus J. Buehler

128,390 görüntüleme • 1 ay önce

My first test with the new Gemini Deep Think 3 🔥🔥🔥 Build a complete Three.js scene in a single HTML file that renders a fully 3D interior room indistinguishable from a classical oil painting hanging in a museum. The Painterly Rendering System Write custom GLSL shaders that replace all standard rendering with oil paint simulation. Every pixel must feel painted by hand. The system needs these layers working together. Brushstroke normals. Generate a procedural brushstroke normal map using layered directional noise at varying scales. Large bold strokes for walls and floors following the plane direction. Small delicate strokes for fine details like metal and glass. Circular strokes for rounded objects. The brushstrokes must catch sidelight and cast tiny shadows into their grooves exactly like real impasto paint on canvas. Paint thickness. Use parallax occlusion mapping to give highlights genuine physical thickness. Where the original painter would load their brush with white or yellow to hit a bright highlight, the paint should visibly sit above the surface. In dark shadow areas, the paint should appear thinner, letting canvas weave show through slightly. Color palette. Restrict the entire scene to a historical oil palette. Titanium white, naples yellow, yellow ochre, raw sienna, burnt sienna, burnt umber, raw umber, ivory black, vermillion used sparingly, and a muted blue-grey. No modern saturated colors. All color mixing should feel subtractive and warm. Edge treatment. No hard edges anywhere in the scene. Object silhouettes must soften and blur slightly as if the painter's brush feathered where one form meets another. Implement this as a screen-space edge-detection pass that blurs based on depth discontinuity and overlays brushstroke texture at boundaries. Canvas texture. The entire final image must have a linen canvas weave overlay rendered with its own normal map that interacts with the scene lighting. As you orbit, the canvas tooth should glint differently. This sells the illusion more than anything else. Varnish layer. Apply a post-processing pass that simulates aged oil varnish. A warm amber tint that is slightly uneven, thicker in corners and thinner in center. A subtle gloss reflection that shifts as you move the camera. Very fine craquelure, hairline crack patterns, visible only when you zoom in close. The Scene. A Dutch Golden Age Study. Model everything procedurally with no external assets. A small intimate room with rough plastered walls in thick paint, warm grey-ochre. A single tall window on the left wall. Leaded glass with thick mullions letting in one dominant shaft of warm light. The window glass should have slight imperfections like bubbles and waviness visible in the paint treatment. A heavy dark wood table positioned center-left. On the table place a brass candlestick with a half-melted candle where the wax drips are modeled and painted with naples yellow impasto highlights. A pewter plate with a half-peeled lemon, the peel curling off the edge of the plate, the exposed fruit flesh a jewel of thick yellow paint catching light. A partially unfolded letter with a broken red wax seal. A small glass of dark wine catching a single highlight. A dark velvet cloth draped from the table edge falling in heavy folds to the floor. The velvet should have that characteristic oil painting treatment where shadows go almost black and the fabric catches light in soft broken highlights. The floor is wide dark wooden planks painted with long horizontal brushstrokes. Against the back wall, barely visible in shadow, a tall wooden cabinet with a few old leather-bound books leaning against each other. A single beam of light from the window cuts diagonally across the scene illuminating floating dust motes painted as soft tiny dots of naples yellow, not CG particles. The rest of the room falls into rich warm shadow. Lighting One dominant directional light from the left simulating window light. Warm, strong, with soft VSM shadows. A very subtle fill from the right in cool blue-grey at perhaps 5% intensity. No ambient light. The shadows should be genuinely dark and warm. This is Vermeer lighting. The contrast between the luminous light-struck areas and the deep velvety shadows is what makes the painting breathe. Post-Processing Chain Render pass. Painterly edge softening pass. Canvas texture overlay pass. Varnish and aging pass. Heavy vignette like a dark gallery frame encroaching. Very subtle bloom only on the brightest impasto highlights. Film grain that mimics canvas tooth texture rather than photographic noise. Interaction OrbitControls with very slow damping at 0.02 and limited orbit range so you can look around the painting but not flip it upside down. Slow zoom. The feeling should be like leaning closer to a painting in a museum and discovering more detail. The Standard When someone opens this file and takes a screenshot, people should genuinely argue whether it is a photograph of a real oil painting or a digital render.
0:36

My first test with the new Gemini Deep Think 3 🔥🔥🔥 Build a complete Three.js scene in a single HTML file that renders a fully 3D interior room indistinguishable from a classical oil painting hanging in a museum. The Painterly Rendering System Write custom GLSL shaders that replace all standard rendering with oil paint simulation. Every pixel must feel painted by hand. The system needs these layers working together. Brushstroke normals. Generate a procedural brushstroke normal map using layered directional noise at varying scales. Large bold strokes for walls and floors following the plane direction. Small delicate strokes for fine details like metal and glass. Circular strokes for rounded objects. The brushstrokes must catch sidelight and cast tiny shadows into their grooves exactly like real impasto paint on canvas. Paint thickness. Use parallax occlusion mapping to give highlights genuine physical thickness. Where the original painter would load their brush with white or yellow to hit a bright highlight, the paint should visibly sit above the surface. In dark shadow areas, the paint should appear thinner, letting canvas weave show through slightly. Color palette. Restrict the entire scene to a historical oil palette. Titanium white, naples yellow, yellow ochre, raw sienna, burnt sienna, burnt umber, raw umber, ivory black, vermillion used sparingly, and a muted blue-grey. No modern saturated colors. All color mixing should feel subtractive and warm. Edge treatment. No hard edges anywhere in the scene. Object silhouettes must soften and blur slightly as if the painter's brush feathered where one form meets another. Implement this as a screen-space edge-detection pass that blurs based on depth discontinuity and overlays brushstroke texture at boundaries. Canvas texture. The entire final image must have a linen canvas weave overlay rendered with its own normal map that interacts with the scene lighting. As you orbit, the canvas tooth should glint differently. This sells the illusion more than anything else. Varnish layer. Apply a post-processing pass that simulates aged oil varnish. A warm amber tint that is slightly uneven, thicker in corners and thinner in center. A subtle gloss reflection that shifts as you move the camera. Very fine craquelure, hairline crack patterns, visible only when you zoom in close. The Scene. A Dutch Golden Age Study. Model everything procedurally with no external assets. A small intimate room with rough plastered walls in thick paint, warm grey-ochre. A single tall window on the left wall. Leaded glass with thick mullions letting in one dominant shaft of warm light. The window glass should have slight imperfections like bubbles and waviness visible in the paint treatment. A heavy dark wood table positioned center-left. On the table place a brass candlestick with a half-melted candle where the wax drips are modeled and painted with naples yellow impasto highlights. A pewter plate with a half-peeled lemon, the peel curling off the edge of the plate, the exposed fruit flesh a jewel of thick yellow paint catching light. A partially unfolded letter with a broken red wax seal. A small glass of dark wine catching a single highlight. A dark velvet cloth draped from the table edge falling in heavy folds to the floor. The velvet should have that characteristic oil painting treatment where shadows go almost black and the fabric catches light in soft broken highlights. The floor is wide dark wooden planks painted with long horizontal brushstrokes. Against the back wall, barely visible in shadow, a tall wooden cabinet with a few old leather-bound books leaning against each other. A single beam of light from the window cuts diagonally across the scene illuminating floating dust motes painted as soft tiny dots of naples yellow, not CG particles. The rest of the room falls into rich warm shadow. Lighting One dominant directional light from the left simulating window light. Warm, strong, with soft VSM shadows. A very subtle fill from the right in cool blue-grey at perhaps 5% intensity. No ambient light. The shadows should be genuinely dark and warm. This is Vermeer lighting. The contrast between the luminous light-struck areas and the deep velvety shadows is what makes the painting breathe. Post-Processing Chain Render pass. Painterly edge softening pass. Canvas texture overlay pass. Varnish and aging pass. Heavy vignette like a dark gallery frame encroaching. Very subtle bloom only on the brightest impasto highlights. Film grain that mimics canvas tooth texture rather than photographic noise. Interaction OrbitControls with very slow damping at 0.02 and limited orbit range so you can look around the painting but not flip it upside down. Slow zoom. The feeling should be like leaning closer to a painting in a museum and discovering more detail. The Standard When someone opens this file and takes a screenshot, people should genuinely argue whether it is a photograph of a real oil painting or a digital render.

Emily

20,940 görüntüleme • 6 ay önce

First drive & impressions of FSD 14.2 in South Central Texas. I just received the update on my 2025 Model 3 & went out to test it on some areas where FSD has had difficulties in previous versions including the previous FSD 14.1.7 and also to get overall impressions. Overall, it did well over about 45 miles of driving on rural highways, congested areas with lots of traffic, construction, grocery store busy parking lots & rural post office destinations. Although in general it did very well in a variety of situations, a few stood out that I want to highlight where I did have disconnects & unusual behaviors that seem to have been introduced with FSD 14.2. Summary of Strengths: Smooth, confident most of the time, and polished. Felt comfortable at all times while FSD was driving the car. This includes all of the driver settings (Sloth, Chill, Standard, Hurry & Mad Max) Opportunities & areas with challenges: - Navigation database mismatches really becoming a glaring deficiency with the approach Tesla AI is using, not just because of recent or ongoing construction, but also with established routes, roads & destinations - FSD 14.2 stopped unexpectedly at a gas station that was not part of the route or any destination in the route. After parking, I tapped the Start Self Driving button and the car resumed as if it was starting at an intermediate stopping point & it continued on to the original destination. - A second strange Navigation challenge happened when arriving at my home. For no apparent reason, it tried to go to my neighbors house … something no previous FSD version has done. Very strange. - Construction prevented FSD from entering into the selected grocery store entrance, but instead of taking several obvious nearby detour routes or doing a U-turn, it re-mapped the route & drove a few miles past the grocery store & into a residential area, did a turn around & then rerouted all the way back to the grocery store. This was several miles of added driving & passed numerous paths that were much closer & available. - Parking in angled parking spaces continues to be a big challenge for FSD 14.2 … it took up two parking spaces & at an odd angle when parking at a rural Post Office, disregarding the painted signs & angled marks on the parking spaces & the sign posts for the parking spaces. It also attempted to drive the wrong way in the one-way parking area. - Speed profiles seem a bit better, but speeding, especially in my neighborhood area is still excessive. For some reason, a posted 25 mph section showed a 65 mph speed sign on the MCU display & FSD tried to drive to that speed. This might be related to navigation databases mismatches issues mentioned earlier. In summary, I think Tesla FSD 14.2 shows some promising signs of progress, but still has several areas that need additional work, the biggest of which is the navigation databases mismatch issues. I’d like to see some way for owners to provide inputs to the FSD team to provide navigation map updates that can help fix these issues. Also, having a way to “train” FSD in specific areas, say at your home so it knows how to properly enter the driveway, back into the garage, select which side of the garage, etc… would be very helpful.
8:44

First drive & impressions of FSD 14.2 in South Central Texas. I just received the update on my 2025 Model 3 & went out to test it on some areas where FSD has had difficulties in previous versions including the previous FSD 14.1.7 and also to get overall impressions. Overall, it did well over about 45 miles of driving on rural highways, congested areas with lots of traffic, construction, grocery store busy parking lots & rural post office destinations. Although in general it did very well in a variety of situations, a few stood out that I want to highlight where I did have disconnects & unusual behaviors that seem to have been introduced with FSD 14.2. Summary of Strengths: Smooth, confident most of the time, and polished. Felt comfortable at all times while FSD was driving the car. This includes all of the driver settings (Sloth, Chill, Standard, Hurry & Mad Max) Opportunities & areas with challenges: - Navigation database mismatches really becoming a glaring deficiency with the approach Tesla AI is using, not just because of recent or ongoing construction, but also with established routes, roads & destinations - FSD 14.2 stopped unexpectedly at a gas station that was not part of the route or any destination in the route. After parking, I tapped the Start Self Driving button and the car resumed as if it was starting at an intermediate stopping point & it continued on to the original destination. - A second strange Navigation challenge happened when arriving at my home. For no apparent reason, it tried to go to my neighbors house … something no previous FSD version has done. Very strange. - Construction prevented FSD from entering into the selected grocery store entrance, but instead of taking several obvious nearby detour routes or doing a U-turn, it re-mapped the route & drove a few miles past the grocery store & into a residential area, did a turn around & then rerouted all the way back to the grocery store. This was several miles of added driving & passed numerous paths that were much closer & available. - Parking in angled parking spaces continues to be a big challenge for FSD 14.2 … it took up two parking spaces & at an odd angle when parking at a rural Post Office, disregarding the painted signs & angled marks on the parking spaces & the sign posts for the parking spaces. It also attempted to drive the wrong way in the one-way parking area. - Speed profiles seem a bit better, but speeding, especially in my neighborhood area is still excessive. For some reason, a posted 25 mph section showed a 65 mph speed sign on the MCU display & FSD tried to drive to that speed. This might be related to navigation databases mismatches issues mentioned earlier. In summary, I think Tesla FSD 14.2 shows some promising signs of progress, but still has several areas that need additional work, the biggest of which is the navigation databases mismatch issues. I’d like to see some way for owners to provide inputs to the FSD team to provide navigation map updates that can help fix these issues. Also, having a way to “train” FSD in specific areas, say at your home so it knows how to properly enter the driveway, back into the garage, select which side of the garage, etc… would be very helpful.

Joe Tegtmeyer 🚀 🤠🛸😎

66,927 görüntüleme • 9 ay önce