Загрузка видео...

Не удалось загрузить видео

Возникла проблема при загрузке этого видео. Это может быть связано с временными проблемами сети или видео может быть недоступно.

На главную

Check out 𝐌𝐨𝐭𝐢𝐨𝐧𝟐𝐕𝐞𝐜𝐒𝐞𝐭𝐬, a 4D diffusion model for dynamic surface reconstruction from imperfect observations of sparse, noisy, or partial point clouds. Main idea: we represent time-varying shapes via 4D neural representation with latent vector sets, and then explicitly learns the shape and motion distribution of non-rigid objects through an... show more

Matthias Niessner

48,725 subscribers

29,142 просмотров • 2 лет назад •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

Комментарии: 1

Фото профиля 田中義弘 | taziku CEO / AI × Creative

田中義弘 | taziku CEO / AI × Creative2 лет назад

Technology that may take 3D generative AI one step forward! Thanks for the info.

Похожие видео

SceNeRFlow: Time-Consistent Reconstruction of General Dynamic Scenes abs: paper page: Existing methods for the 4D reconstruction of general, non-rigidly deforming objects focus on novel-view synthesis and neglect correspondences. However, time consistency enables advanced downstream tasks like 3D editing, motion analysis, or virtual-asset creation. We propose SceNeRFlow to reconstruct a general, non-rigid scene in a time-consistent manner. Our dynamic-NeRF method takes multi-view RGB videos and background images from static cameras with known camera parameters as input. It then reconstructs the deformations of an estimated canonical model of the geometry and appearance in an online fashion. Since this canonical model is time-invariant, we obtain correspondences even for long-term, long-range motions. We employ neural scene representations to parametrize the components of our method. Like prior dynamic-NeRF methods, we use a backwards deformation model. We find non-trivial adaptations of this model necessary to handle larger motions: We decompose the deformations into a strongly regularized coarse component and a weakly regularized fine component, where the coarse component also extends the deformation field into the space surrounding the object, which enables tracking over time. We show experimentally that, unlike prior work that only handles small motion, our method enables the reconstruction of studio-scale motions.

SceNeRFlow: Time-Consistent Reconstruction of General Dynamic Scenes abs: paper page: Existing methods for the 4D reconstruction of general, non-rigidly deforming objects focus on novel-view synthesis and neglect correspondences. However, time consistency enables advanced downstream tasks like 3D editing, motion analysis, or virtual-asset creation. We propose SceNeRFlow to reconstruct a general, non-rigid scene in a time-consistent manner. Our dynamic-NeRF method takes multi-view RGB videos and background images from static cameras with known camera parameters as input. It then reconstructs the deformations of an estimated canonical model of the geometry and appearance in an online fashion. Since this canonical model is time-invariant, we obtain correspondences even for long-term, long-range motions. We employ neural scene representations to parametrize the components of our method. Like prior dynamic-NeRF methods, we use a backwards deformation model. We find non-trivial adaptations of this model necessary to handle larger motions: We decompose the deformations into a strongly regularized coarse component and a weakly regularized fine component, where the coarse component also extends the deformation field into the space surrounding the object, which enables tracking over time. We show experimentally that, unlike prior work that only handles small motion, our method enables the reconstruction of studio-scale motions.

AK

76,380 просмотров • 3 лет назад

[NeurIPS '24] DreamMesh4D: Video-to-4D Generation with Sparse-Controlled Gaussian-Mesh Hybrid Representation Abstract (excerpt) We introduce DreamMesh4D, a novel framework that combines mesh representation with sparse-controlled deformation technique to generate high-quality 4D object from a monocular video. To overcome the limitation of classical texture representation, we bind Gaussian splats to the surface of the triangular mesh for differentiable optimization of both the texture and mesh vertices. In particular, DreamMesh4D begins with a coarse mesh provided by a single image based 3D generation method. Sparse points are then uniformly sampled across the surface of the mesh, and are used to build a deformation graph to drive the motion of the 3D object for the sake of computational efficiency and providing additional constraint. For each step, transformations of sparse control points are predicted using a deformation network, and the mesh vertices as well as the bound surface Gaussians are deformed via a geometric skinning algorithm. The skinning algorithm is a hybrid approach combining LBS (linear blending skinning) and DQS (dual-quaternion skinning), mitigating drawbacks associated with both approaches. The static surface Gaussians and mesh vertices as well as the dynamic deformation network are learned via reference view photometric loss, score distillation loss as well as other regularization losses in a two-stage manner. Extensive experiments demonstrate that our method outperforms prior video-to-4D generation methods in terms of rendering quality and spatial-temporal consistency.

[NeurIPS '24] DreamMesh4D: Video-to-4D Generation with Sparse-Controlled Gaussian-Mesh Hybrid Representation Abstract (excerpt) We introduce DreamMesh4D, a novel framework that combines mesh representation with sparse-controlled deformation technique to generate high-quality 4D object from a monocular video. To overcome the limitation of classical texture representation, we bind Gaussian splats to the surface of the triangular mesh for differentiable optimization of both the texture and mesh vertices. In particular, DreamMesh4D begins with a coarse mesh provided by a single image based 3D generation method. Sparse points are then uniformly sampled across the surface of the mesh, and are used to build a deformation graph to drive the motion of the 3D object for the sake of computational efficiency and providing additional constraint. For each step, transformations of sparse control points are predicted using a deformation network, and the mesh vertices as well as the bound surface Gaussians are deformed via a geometric skinning algorithm. The skinning algorithm is a hybrid approach combining LBS (linear blending skinning) and DQS (dual-quaternion skinning), mitigating drawbacks associated with both approaches. The static surface Gaussians and mesh vertices as well as the dynamic deformation network are learned via reference view photometric loss, score distillation loss as well as other regularization losses in a two-stage manner. Extensive experiments demonstrate that our method outperforms prior video-to-4D generation methods in terms of rendering quality and spatial-temporal consistency.

MrNeRF

12,323 просмотров • 1 год назад

Fast View Synthesis of Casual Videos paper page: Novel view synthesis from an in-the-wild video is difficult due to challenges like scene dynamics and lack of parallax. While existing methods have shown promising results with implicit neural radiance fields, they are slow to train and render. This paper revisits explicit video representations to synthesize high-quality novel views from a monocular video efficiently. We treat static and dynamic video content separately. Specifically, we build a global static scene model using an extended plane-based scene representation to synthesize temporally coherent novel video. Our plane-based scene representation is augmented with spherical harmonics and displacement maps to capture view-dependent effects and model non-planar complex surface geometry. We opt to represent the dynamic content as per-frame point clouds for efficiency. While such representations are inconsistency-prone, minor temporal inconsistencies are perceptually masked due to motion. We develop a method to quickly estimate such a hybrid video representation and render novel views in real time. Our experiments show that our method can render high-quality novel views from an in-the-wild video with comparable quality to state-of-the-art methods while being 100x faster in training and enabling real-time rendering.

Fast View Synthesis of Casual Videos paper page: Novel view synthesis from an in-the-wild video is difficult due to challenges like scene dynamics and lack of parallax. While existing methods have shown promising results with implicit neural radiance fields, they are slow to train and render. This paper revisits explicit video representations to synthesize high-quality novel views from a monocular video efficiently. We treat static and dynamic video content separately. Specifically, we build a global static scene model using an extended plane-based scene representation to synthesize temporally coherent novel video. Our plane-based scene representation is augmented with spherical harmonics and displacement maps to capture view-dependent effects and model non-planar complex surface geometry. We opt to represent the dynamic content as per-frame point clouds for efficiency. While such representations are inconsistency-prone, minor temporal inconsistencies are perceptually masked due to motion. We develop a method to quickly estimate such a hybrid video representation and render novel views in real time. Our experiments show that our method can render high-quality novel views from an in-the-wild video with comparable quality to state-of-the-art methods while being 100x faster in training and enabling real-time rendering.

AK

20,668 просмотров • 2 лет назад

The Genie 3 release is a perfect moment to have a discussion about the future of 3D But first it would be nice to make the terminology more clear, specifically: What is a “spatial representation” Implicit vs Explicit Generalization vs Specialization Reconstruction vs Generation Production vs Execution Let’s start: For me, a spatial representation is just a way to describe a thing in the physical world The core property that makes it useful is consistency You can enforce consistency explicitly via rendering equations, geometric constraints, and physics Or implicitly, purely through training data Then, your representation parameters can be explicit, like points, gaussians, triangles, voxels, etc. Or implicit, weights or latent vectors Parameters alone are not the representation. It’s a combination of the parameters, the process that produces them, and the way you materialize them through a function (physics-based rendering, simulation, neural network, etc.) Generalization means you take data from multiple scene observations, and then produce a map from desired input to representation parameters Specialization means you take single-scene observations and directly fit a function parameters to describe thar scene Many representations can serve both of the approaches, as long as you keep them differentiable Both of the above can be used for reconstruction, where the main goal is to explain observations through a lens of physics (hard constraint) On the other hand, generation needs generalization, and its task is to produce statistically plausible results that could be conditioned on observations (soft constraint) Both tasks are not solved yet and they can complement each other in various ways Yet another important aspect is the difference between production and execution Production = process of going from inputs to parameters Execution = process of going from parameters to result It’s important to separate these, because most usecases require fast execution to be viable which is severely constrained by the hardware So, are *world models* like Genie an important step forward? Yes Do they make other representations obsolete? Maybe some of them - but there are tons of economically valuable tasks that won’t be solved by it, at least in any observable future

The Genie 3 release is a perfect moment to have a discussion about the future of 3D But first it would be nice to make the terminology more clear, specifically: What is a “spatial representation” Implicit vs Explicit Generalization vs Specialization Reconstruction vs Generation Production vs Execution Let’s start: For me, a spatial representation is just a way to describe a thing in the physical world The core property that makes it useful is consistency You can enforce consistency explicitly via rendering equations, geometric constraints, and physics Or implicitly, purely through training data Then, your representation parameters can be explicit, like points, gaussians, triangles, voxels, etc. Or implicit, weights or latent vectors Parameters alone are not the representation. It’s a combination of the parameters, the process that produces them, and the way you materialize them through a function (physics-based rendering, simulation, neural network, etc.) Generalization means you take data from multiple scene observations, and then produce a map from desired input to representation parameters Specialization means you take single-scene observations and directly fit a function parameters to describe thar scene Many representations can serve both of the approaches, as long as you keep them differentiable Both of the above can be used for reconstruction, where the main goal is to explain observations through a lens of physics (hard constraint) On the other hand, generation needs generalization, and its task is to produce statistically plausible results that could be conditioned on observations (soft constraint) Both tasks are not solved yet and they can complement each other in various ways Yet another important aspect is the difference between production and execution Production = process of going from inputs to parameters Execution = process of going from parameters to result It’s important to separate these, because most usecases require fast execution to be viable which is severely constrained by the hardware So, are world models like Genie an important step forward? Yes Do they make other representations obsolete? Maybe some of them - but there are tons of economically valuable tasks that won’t be solved by it, at least in any observable future

Lucky Iyinbor

13,959 просмотров • 6 месяцев назад

Excited to announce GR00T N1, the world’s first open foundation model for humanoid robots! We are on a mission to democratize Physical AI. The power of general robot brain, in the palm of your hand - with only 2B parameters, N1 learns from the most diverse physical action dataset ever compiled and punches above its weight: - Real humanoid teleoperation data. - Large-scale simulation data: we are open-sourcing 300K+ trajectories! - Neural trajectories: we apply SOTA video generation models to “hallucinate” new synthetic data that features accurate physics in pixels. Using Jensen’s words, “systematically infinite data”! - Latent actions: we develop novel algorithms to extract action tokens from in-the-wild human videos and neural generated videos. GR00T N1 is a single end-to-end neural net, from photons to actions: - Vision-Language Model (System 2) that interprets the physical world through vision and language instructions, enabling robots to reason about their environment and instructions, and plan the right actions. - Diffusion Transformer (System 1) that “renders” smooth and precise motor actions at 120 Hz, executing the latent plan made by System 2. We deploy N1 on GR1 robot, 1X Neo robot, and a large collection of simulation benchmarks. N1 achieves up to +30% boost in diverse manipulation tasks for household and industrial settings. While humanoid robots are the main focus of N1, our model also supports cross-embodiment. We finetune it to work on the $110 HuggingFace LeRobot SO100 robot arm! Open robot brain runs on open hardware. Sounds just right. Let’s solve robotics, together, one token at a time. Links to our Whitepaper, Github repo, HuggingFace model, and open dataset page in the thread: 🧵

Excited to announce GR00T N1, the world’s first open foundation model for humanoid robots! We are on a mission to democratize Physical AI. The power of general robot brain, in the palm of your hand - with only 2B parameters, N1 learns from the most diverse physical action dataset ever compiled and punches above its weight: - Real humanoid teleoperation data. - Large-scale simulation data: we are open-sourcing 300K+ trajectories! - Neural trajectories: we apply SOTA video generation models to “hallucinate” new synthetic data that features accurate physics in pixels. Using Jensen’s words, “systematically infinite data”! - Latent actions: we develop novel algorithms to extract action tokens from in-the-wild human videos and neural generated videos. GR00T N1 is a single end-to-end neural net, from photons to actions: - Vision-Language Model (System 2) that interprets the physical world through vision and language instructions, enabling robots to reason about their environment and instructions, and plan the right actions. - Diffusion Transformer (System 1) that “renders” smooth and precise motor actions at 120 Hz, executing the latent plan made by System 2. We deploy N1 on GR1 robot, 1X Neo robot, and a large collection of simulation benchmarks. N1 achieves up to +30% boost in diverse manipulation tasks for household and industrial settings. While humanoid robots are the main focus of N1, our model also supports cross-embodiment. We finetune it to work on the $110 HuggingFace LeRobot SO100 robot arm! Open robot brain runs on open hardware. Sounds just right. Let’s solve robotics, together, one token at a time. Links to our Whitepaper, Github repo, HuggingFace model, and open dataset page in the thread: 🧵

Jim Fan

466,610 просмотров • 1 год назад

This BlenderFusion paper basically says "screw trying to describe 3D edits through text" and just... use Blender :-) The idea is pretty straightforward -- instead of trying to cram 3D understanding into a diffusion model, use depth estimation & segmentation to project 2D images into 2.5D meshes, edit them in actual 3D software, then use a fine-tuned diffusion model to make the results photorealistic again. The clever bit is their "dual-stream architecture" -- the model sees both the original scene AND the edited Blender render in parallel, learning to preserve what matters while fixing the inevitable artifacts from transforming imperfect 2.5D/3D reconstructions. They train it with smart masking strategies so it learns when to ignore the original scene (for removals/replacements) and can manipulate objects independently of camera motion. What you get is pretty impressive control -- not just moving objects around, but changing materials, deforming shapes, swapping backgrounds, all while maintaining visual coherence. Neural Assets (one of my favorite papers last year) tried to crack this with learned object tokens, but it struggled with overlapping objects and loses fine details (due to low res DINO encodings). BlenderFusion just sidesteps the whole problem -- want to rotate something 173.5 degrees? Just rotate it in Blender. Want to duplicate an object 8 times? Copy paste away. The diffusion model's only job is making it look photorealistic, not figuring out the 3D underpinnings. The catch? Lacks temporal consistency for animation. Each viewpoint is generated independently, so while a single edit looks great, smoothly animating a car or camera down the street won't work -- you'd get flickering and inconsistencies between frames. That said, this approach is so much more intuitive for finer grain image editing than trying to describe your changes in text prompts. It's the kind of thing that makes you wonder why we're trying to do everything inside neural networks when perfectly good 3D tools already exist -- giving you the best of both worlds.

This BlenderFusion paper basically says "screw trying to describe 3D edits through text" and just... use Blender :-) The idea is pretty straightforward -- instead of trying to cram 3D understanding into a diffusion model, use depth estimation & segmentation to project 2D images into 2.5D meshes, edit them in actual 3D software, then use a fine-tuned diffusion model to make the results photorealistic again. The clever bit is their "dual-stream architecture" -- the model sees both the original scene AND the edited Blender render in parallel, learning to preserve what matters while fixing the inevitable artifacts from transforming imperfect 2.5D/3D reconstructions. They train it with smart masking strategies so it learns when to ignore the original scene (for removals/replacements) and can manipulate objects independently of camera motion. What you get is pretty impressive control -- not just moving objects around, but changing materials, deforming shapes, swapping backgrounds, all while maintaining visual coherence. Neural Assets (one of my favorite papers last year) tried to crack this with learned object tokens, but it struggled with overlapping objects and loses fine details (due to low res DINO encodings). BlenderFusion just sidesteps the whole problem -- want to rotate something 173.5 degrees? Just rotate it in Blender. Want to duplicate an object 8 times? Copy paste away. The diffusion model's only job is making it look photorealistic, not figuring out the 3D underpinnings. The catch? Lacks temporal consistency for animation. Each viewpoint is generated independently, so while a single edit looks great, smoothly animating a car or camera down the street won't work -- you'd get flickering and inconsistencies between frames. That said, this approach is so much more intuitive for finer grain image editing than trying to describe your changes in text prompts. It's the kind of thing that makes you wonder why we're trying to do everything inside neural networks when perfectly good 3D tools already exist -- giving you the best of both worlds.

Bilawal Sidhu

34,440 просмотров • 1 год назад

Want to create an avatar from a single image? FlexAvatar is a transformer model that creates full 360°, high-quality, and expressive 3D head avatar from just a single portrait image in minutes. Real-time Demo: FlexAvatar's lightweight architecture allows both animation and rendering in real-time, enabling interactive user experiences. To create a new 3D head avatar, only one image is required, e.g., from a webcam. The final avatar is ready after 2 minutes. Architecture: Under the hood, FlexAvatar adopts a transformer-based encoder-decoder design. The encoder maps the input image onto a latent avatar space, while the decoder produces 3D Gaussian attribute maps by incorporating the animation signal via cross-attention. The model learns all facial animations directly from the data without relying on pre-built 3D face models. This equips the avatars with realistic facial expressions. The internal avatar latent space can be conveniently used to integrate additional observations of a person via fitting. This enables use-cases where more than one image of a person is available, e.g., from a phone scan of the person. We train jointly on 2D monocular videos and multi-view data. However, in monocular videos, the animation signal leaks the target viewpoint, causing the model to produce incomplete 3D heads. We call this phenomenon entanglement of driving signal and target viewpoint. To prevent entanglement, we introduce bias sinks. These are learnable tokens that indicate whether a training sample stems from a monocular or a multi-view dataset. During training, the model learns to produce incomplete 3D heads only when the monocular token is present. During inference, FlexAvatar then always uses the multi-view token for which the model has learned to produce complete 3D heads. This simple design allows to combine the generalizability from monocular data with the quality of multi-view data. FlexAvatar summary: - Input: Single-image, phone scan, or monocular video - Output: Full 360° head avatar - Expressive animations - Real-time rendering and animation - Generalization to any portrait - Create a new avatar in 2 minutes - Use bias sinks to combine 2D and 3D data 🏠 🌍 🎥 Great work by Tobias Kirschstein and Simon Giebenhain!

Matthias Niessner

96,238 просмотров • 8 месяцев назад

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

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 просмотров • 1 год назад

Variational Autoencoder by hand ✍️ ~ 11 steps walkthrough below A VAE learns the structure of your data, the mean and variance of its hidden features, and then generates new data from that structure. A GAN only learns to fool a discriminator. It can make convincing fakes without ever knowing what the data is really made of. That is the difference, and it is the whole reason VAEs matter. In 2024 ICLR gave its first ever Test of Time Award to the VAE paper, "Auto-Encoding Variational Bayes" by Diederik Kingma and Max Welling, ten years on. How does it work? Goal: encode three inputs into a distribution, sample from it, decode it back, and read every loss gradient off the page. = 1. Given = Three training examples X1, X2, X3, copied to the bottom as their own targets. Reconstructing your own input is what puts the "auto", meaning self, in autoencoder. = 2. Encoder, layer 1 = Let us multiply the inputs by weights and biases, then apply ReLU, crossing out every negative. = 3. Mean and standard deviation = We multiply the features by two more weight sets. The first predicts the means μ of the latent distributions, the second their standard deviations σ. = 4. A random offset = Let us sample ε from a standard normal, mean 0 and variance 1, and multiply it by σ. This is a random step away from the mean, scaled by how uncertain each feature is. = 5. Mean plus offset = We add the offset back onto μ, and these become the decoder's inputs. Keeping the randomness out in ε is the reparameterization trick: it lets gradients flow straight through the sampling. = 6. Decoder, layer 1 = Let us multiply by weights and biases and apply ReLU again. Here -4 is crossed out. = 7. Decoder, layer 2 = We multiply once more. The output Y is the decoder's attempt to rebuild X from the sampled distribution. = 8. Gradient for the mean = Let us push μ toward 0. A lot of math, the SGVB estimator, collapses the KL gradient to simply μ itself. = 9. Gradient for the standard deviation = We want σ to approach 1. = 10. And its formula = That same math simplifies the gradient to σ minus 1/σ. = 11. Reconstruction gradient = We want the reconstruction Y to match the input X. Mean squared error simplifies its gradient to Y minus X. Takeaway: the two gradients you just calculated each sit at the heart of a modern method, so one VAE teaches you both. The KL divergence is the penalty RLHF like GRPO uses to keep a fine-tuned model from drifting off its base. The reconstruction loss, plain mean squared error, is exactly what trains a diffusion model to denoise. Draw one VAE by hand and you have quietly learned the core of both. 💾 Save this post!

Variational Autoencoder by hand ✍️ ~ 11 steps walkthrough below A VAE learns the structure of your data, the mean and variance of its hidden features, and then generates new data from that structure. A GAN only learns to fool a discriminator. It can make convincing fakes without ever knowing what the data is really made of. That is the difference, and it is the whole reason VAEs matter. In 2024 ICLR gave its first ever Test of Time Award to the VAE paper, "Auto-Encoding Variational Bayes" by Diederik Kingma and Max Welling, ten years on. How does it work? Goal: encode three inputs into a distribution, sample from it, decode it back, and read every loss gradient off the page. = 1. Given = Three training examples X1, X2, X3, copied to the bottom as their own targets. Reconstructing your own input is what puts the "auto", meaning self, in autoencoder. = 2. Encoder, layer 1 = Let us multiply the inputs by weights and biases, then apply ReLU, crossing out every negative. = 3. Mean and standard deviation = We multiply the features by two more weight sets. The first predicts the means μ of the latent distributions, the second their standard deviations σ. = 4. A random offset = Let us sample ε from a standard normal, mean 0 and variance 1, and multiply it by σ. This is a random step away from the mean, scaled by how uncertain each feature is. = 5. Mean plus offset = We add the offset back onto μ, and these become the decoder's inputs. Keeping the randomness out in ε is the reparameterization trick: it lets gradients flow straight through the sampling. = 6. Decoder, layer 1 = Let us multiply by weights and biases and apply ReLU again. Here -4 is crossed out. = 7. Decoder, layer 2 = We multiply once more. The output Y is the decoder's attempt to rebuild X from the sampled distribution. = 8. Gradient for the mean = Let us push μ toward 0. A lot of math, the SGVB estimator, collapses the KL gradient to simply μ itself. = 9. Gradient for the standard deviation = We want σ to approach 1. = 10. And its formula = That same math simplifies the gradient to σ minus 1/σ. = 11. Reconstruction gradient = We want the reconstruction Y to match the input X. Mean squared error simplifies its gradient to Y minus X. Takeaway: the two gradients you just calculated each sit at the heart of a modern method, so one VAE teaches you both. The KL divergence is the penalty RLHF like GRPO uses to keep a fine-tuned model from drifting off its base. The reconstruction loss, plain mean squared error, is exactly what trains a diffusion model to denoise. Draw one VAE by hand and you have quietly learned the core of both. 💾 Save this post!

Tom Yeh

17,011 просмотров • 1 месяц назад

FINALLY finishing up a MASSIVE PR from hell for the Sega Dreamcast port of Grand Theft Auto 3! This is an actual hardware capture now of the DC version under a high load, which would've previously been a slideshow, between the dynamic lighting from the sirens, the amount of rigid bodies in the physics simulation from the cars, and the high-speed chase placing high-demands on asset streaming... I went through all of the low-level common math infrastructure in both the engine and at the RenderWare driver layer and made numerous optimizations, before slowly working my way up to optimizing individual algorithms at the application layer using the new math routines. Firstly, the common low-level floating-point math routines for everything from trig and inverse square root operations to floor(), ceiling(), and clamp(), were replaced with what was meticulously found (in Compiler Explorer) to be the optimal patterns for GCC 14.2.0, targeting our SH architecture (sometimes favoring C builtins, sometimes inline SH4 ASM). Next, in the layer above, with inline SH4 assembly, the common matrix math and linear algebra routines were accelerated using the Dreamcast's vector instructions. Some cleverness went down here, such as cramming matrix metadata into unused insignificant bits of an element, combining loading two matrices with multiplying them, fast transposes, fast quaternion multiplication using 4 dot products, etc. Once the foundation was laid, some of the Renderware code such as the calculations for the lighting, updating bounding volumes, and deriving UV coordinates for specular environment maps on the cars was made faster automatically. The main gainz were actually made rewriting a decent amount of the collision intersection and contact resolution code, though, from using C++-style overloaded operators for multiplying single 4D vectors by a 4x4 matrix to doing batches of 4D vectors by the same matrix. This SUBSTANTIALLY reduced the number of times the backing 4x4 matrix bank had to be reloaded and allowed me to keep it resident within registers while it was being used by the intersection algorithms!

FINALLY finishing up a MASSIVE PR from hell for the Sega Dreamcast port of Grand Theft Auto 3! This is an actual hardware capture now of the DC version under a high load, which would've previously been a slideshow, between the dynamic lighting from the sirens, the amount of rigid bodies in the physics simulation from the cars, and the high-speed chase placing high-demands on asset streaming... I went through all of the low-level common math infrastructure in both the engine and at the RenderWare driver layer and made numerous optimizations, before slowly working my way up to optimizing individual algorithms at the application layer using the new math routines. Firstly, the common low-level floating-point math routines for everything from trig and inverse square root operations to floor(), ceiling(), and clamp(), were replaced with what was meticulously found (in Compiler Explorer) to be the optimal patterns for GCC 14.2.0, targeting our SH architecture (sometimes favoring C builtins, sometimes inline SH4 ASM). Next, in the layer above, with inline SH4 assembly, the common matrix math and linear algebra routines were accelerated using the Dreamcast's vector instructions. Some cleverness went down here, such as cramming matrix metadata into unused insignificant bits of an element, combining loading two matrices with multiplying them, fast transposes, fast quaternion multiplication using 4 dot products, etc. Once the foundation was laid, some of the Renderware code such as the calculations for the lighting, updating bounding volumes, and deriving UV coordinates for specular environment maps on the cars was made faster automatically. The main gainz were actually made rewriting a decent amount of the collision intersection and contact resolution code, though, from using C++-style overloaded operators for multiplying single 4D vectors by a 4x4 matrix to doing batches of 4D vectors by the same matrix. This SUBSTANTIALLY reduced the number of times the backing 4x4 matrix bank had to be reloaded and allowed me to keep it resident within registers while it was being used by the intersection algorithms!

Falco Girgis

114,621 просмотров • 1 год назад

Check out our #ICRA2024 paper "Actor-Critic Model Predictive Control." Model-free #reinforcementlearning (RL) is known for its strong task performance and flexibility in optimizing general reward formulations. On the other hand, #ModelPredictiveControl (MPC) benefits from robustness and online replanning capabilities. We combine both approaches by introducing a new framework called Actor-Critic Model Predictive Control. The key idea is to embed a differentiable MPC within an Actor-Critic RL framework. The proposed approach leverages the short-term predictive optimization capabilities of MPC with the exploratory and end-to-end training properties of RL. The resulting policy effectively manages both short-term decisions through the MPC-based actor and long-term prediction via the critic network, unifying the benefits of both model-based control and end-to-end learning. We validate our method in simulation and the real world with a quadcopter across various high-level tasks. We show that the proposed architecture can achieve real-time control performance, learn complex behaviors via trial and error, and retain the predictive properties of the MPC to better handle out-of-distribution behavior. Paper: Full Video with more details: Kudos to Ángel Romero, Yunlong Song IEEE ICRA University of Zurich UZH Science UZH Space Hub Aerial Core AUTOASSESS European Research Council (ERC)

Check out our #ICRA2024 paper "Actor-Critic Model Predictive Control." Model-free #reinforcementlearning (RL) is known for its strong task performance and flexibility in optimizing general reward formulations. On the other hand, #ModelPredictiveControl (MPC) benefits from robustness and online replanning capabilities. We combine both approaches by introducing a new framework called Actor-Critic Model Predictive Control. The key idea is to embed a differentiable MPC within an Actor-Critic RL framework. The proposed approach leverages the short-term predictive optimization capabilities of MPC with the exploratory and end-to-end training properties of RL. The resulting policy effectively manages both short-term decisions through the MPC-based actor and long-term prediction via the critic network, unifying the benefits of both model-based control and end-to-end learning. We validate our method in simulation and the real world with a quadcopter across various high-level tasks. We show that the proposed architecture can achieve real-time control performance, learn complex behaviors via trial and error, and retain the predictive properties of the MPC to better handle out-of-distribution behavior. Paper: Full Video with more details: Kudos to Ángel Romero, Yunlong Song IEEE ICRA University of Zurich UZH Science UZH Space Hub Aerial Core AUTOASSESS European Research Council (ERC)

Davide Scaramuzza

34,889 просмотров • 2 лет назад

[CLIP] by Hand ✍️ The CLIP (Contrastive Language–Image Pre-training) model, a groundbreaking work by OpenAI, redefines the intersection of computer vision and natural language processing. It is the basis of all the multi-modal foundation models we see today. How does CLIP work? Goal: 🟨 Learn a shared embedding space for text and image [1] Given ↳ A mini batch of 3 text-image pairs ↳ OpenAI used 400 million text-image pairs to train its original CLIP model. Process 1st pair: "big table" [2] 🟪 Text → 2 Vectors (3D) ↳ Look up word embedding vectors using word2vec. [3] 🟩 Image → 2 Vectors (4D) ↳ Divide the image into two patches. ↳ Flatten each patch [4] Process other pairs ↳ Repeat [2]-[3] [5] 🟪 Text Encoder & 🟩 Image Encoder ↳ Encode input vectors into feature vectors ↳ Here, both encoders are simple one layer perceptron (linear + ReLU) ↳ In practice, the encoders are usually transformer models. [6] 🟪 🟩 Mean Pooling: 2 → 1 vector ↳ Average 2 feature vectors into a single vector by averaging across the columns ↳ The goal is to have one vector to represent each image or text [7] 🟪 🟩 -> 🟨 Projection ↳ Note that the text and image feature vectors from the encoders have different dimensions (3D vs. 4D). ↳ Use a linear layer to project image and text vectors to a 2D shared embedding space. 🏋️ Contrastive Pre-training 🏋️ [8] Prepare for MatMul ↳ Copy text vectors (T1,T2,T3) ↳ Copy the transpose of image vectors (I1,I2,I3) ↳ They are all in the 2D shared embedding space. [9] 🟦 MatMul ↳ Multiply T and I matrices. ↳ This is equivalent to taking dot product between every pair of image and text vectors. ↳ The purpose is to use dot product to estimate the similarity between a pair of image-text. [10] 🟦 Softmax: e^x ↳ Raise e to the power of the number in each cell ↳ To simplify hand calculation, we approximate e^□ with 3^□. [11] 🟦 Softmax: ∑ ↳ Sum each row for 🟩 image→🟪 text ↳ Sum each column for 🟪 text→ 🟩 image [12] 🟦 Softmax: 1 / sum ↳ Divide each element by the column sum to obtain a similarity matrix for 🟪 text→🟩 image ↳ Divide each element by the row sum to obtain a similarity matrix for 🟩 image→🟪 text [13] 🟥 Loss Gradients ↳ The "Targets" for the similarity matrices are Identity Matrices. ↳ Why? If I and T come from the same pair (i=j), we want the highest value, which is 1, and 0 otherwise. ↳ Apply the simple equation of [Similarity - Target] to compute gradients of for both directions. ↳ Why so simple? Because when Softmax and Cross-Entropy Loss are used together, the math magically works out that way. ↳ These gradients kick off the backpropagation process to update weights and biases of the encoders and projection layers (red borders).

[CLIP] by Hand ✍️ The CLIP (Contrastive Language–Image Pre-training) model, a groundbreaking work by OpenAI, redefines the intersection of computer vision and natural language processing. It is the basis of all the multi-modal foundation models we see today. How does CLIP work? Goal: 🟨 Learn a shared embedding space for text and image [1] Given ↳ A mini batch of 3 text-image pairs ↳ OpenAI used 400 million text-image pairs to train its original CLIP model. Process 1st pair: "big table" [2] 🟪 Text → 2 Vectors (3D) ↳ Look up word embedding vectors using word2vec. [3] 🟩 Image → 2 Vectors (4D) ↳ Divide the image into two patches. ↳ Flatten each patch [4] Process other pairs ↳ Repeat [2]-[3] [5] 🟪 Text Encoder & 🟩 Image Encoder ↳ Encode input vectors into feature vectors ↳ Here, both encoders are simple one layer perceptron (linear + ReLU) ↳ In practice, the encoders are usually transformer models. [6] 🟪 🟩 Mean Pooling: 2 → 1 vector ↳ Average 2 feature vectors into a single vector by averaging across the columns ↳ The goal is to have one vector to represent each image or text [7] 🟪 🟩 -> 🟨 Projection ↳ Note that the text and image feature vectors from the encoders have different dimensions (3D vs. 4D). ↳ Use a linear layer to project image and text vectors to a 2D shared embedding space. 🏋️ Contrastive Pre-training 🏋️ [8] Prepare for MatMul ↳ Copy text vectors (T1,T2,T3) ↳ Copy the transpose of image vectors (I1,I2,I3) ↳ They are all in the 2D shared embedding space. [9] 🟦 MatMul ↳ Multiply T and I matrices. ↳ This is equivalent to taking dot product between every pair of image and text vectors. ↳ The purpose is to use dot product to estimate the similarity between a pair of image-text. [10] 🟦 Softmax: e^x ↳ Raise e to the power of the number in each cell ↳ To simplify hand calculation, we approximate e^□ with 3^□. [11] 🟦 Softmax: ∑ ↳ Sum each row for 🟩 image→🟪 text ↳ Sum each column for 🟪 text→ 🟩 image [12] 🟦 Softmax: 1 / sum ↳ Divide each element by the column sum to obtain a similarity matrix for 🟪 text→🟩 image ↳ Divide each element by the row sum to obtain a similarity matrix for 🟩 image→🟪 text [13] 🟥 Loss Gradients ↳ The "Targets" for the similarity matrices are Identity Matrices. ↳ Why? If I and T come from the same pair (i=j), we want the highest value, which is 1, and 0 otherwise. ↳ Apply the simple equation of [Similarity - Target] to compute gradients of for both directions. ↳ Why so simple? Because when Softmax and Cross-Entropy Loss are used together, the math magically works out that way. ↳ These gradients kick off the backpropagation process to update weights and biases of the encoders and projection layers (red borders).

Tom Yeh

67,874 просмотров • 2 лет назад

The Mathematics of Moving a Cursor with Neural Signals What might Neuralink Neuralink be doing Mathematically? Consider the task of moving a cursor without touching it. The machine is not looking for a full thought, a sentence, or an image. For this Control problem, the useful object is an intended movement state. sₜ = (pₜ, vₜ) Here, pₜ is the cursor position at time t, and vₜ is the velocity the user is trying to express. The implant records neural activity through many electrode channels, then the decoder tries to estimate vₜ from that activity. Neuralink’s PRIME material describes the N1 Implant as recording and transmitting brain activity with the goal of enabling computer control. For channel i, a simple population model is rᵢ(t) ≈ bᵢ + aᵢ max(0, dᵢ · vₜ) + ηᵢ(t) where rᵢ(t) is the measured activity, bᵢ is baseline activity, aᵢ is channel gain, dᵢ is the channel’s preferred movement direction, and ηᵢ(t) is noise. One channel is not the command. The useful signal is the pattern across many channels: rₜ = (r₁(t), r₂(t), …, rₙ(t)) The decoder subtracts the baseline vector b and applies a learned map W: v̂ₜ = W(rₜ − b) This gives an estimate of the intended velocity. The cursor then updates by pₜ₊₁ = pₜ + Δt v̂ₜ This is the loop shown in the render: neural activity -> decoded velocity -> cursor motion The cortical network and electrode threads show the measurement side. The N1 Implant is described as using 1,024 electrodes distributed across 64 flexible threads, each thinner than a human hair. The decoder panel shows the computational side with activity rₜ, decoded velocity v̂ₜ, and the cursor state pₜ changing over time. A noisy biological pattern becomes a state estimate. That estimate becomes motion on a screen. Therefore, the first lesson is not that Neuralink makes the brain a screen. For cursor control, the Mathematics is more precise: A small piece of intention is represented as a hidden state, measured through neural activity, decoded as a vector, and turned into action. #Neuralink #BrainComputerInterface #NeuralEngineering #Mathematics #StateEstimation #Neuroscience #MachineLearning #BiomedicalEngineering

The Mathematics of Moving a Cursor with Neural Signals What might Neuralink Neuralink be doing Mathematically? Consider the task of moving a cursor without touching it. The machine is not looking for a full thought, a sentence, or an image. For this Control problem, the useful object is an intended movement state. sₜ = (pₜ, vₜ) Here, pₜ is the cursor position at time t, and vₜ is the velocity the user is trying to express. The implant records neural activity through many electrode channels, then the decoder tries to estimate vₜ from that activity. Neuralink’s PRIME material describes the N1 Implant as recording and transmitting brain activity with the goal of enabling computer control. For channel i, a simple population model is rᵢ(t) ≈ bᵢ + aᵢ max(0, dᵢ · vₜ) + ηᵢ(t) where rᵢ(t) is the measured activity, bᵢ is baseline activity, aᵢ is channel gain, dᵢ is the channel’s preferred movement direction, and ηᵢ(t) is noise. One channel is not the command. The useful signal is the pattern across many channels: rₜ = (r₁(t), r₂(t), …, rₙ(t)) The decoder subtracts the baseline vector b and applies a learned map W: v̂ₜ = W(rₜ − b) This gives an estimate of the intended velocity. The cursor then updates by pₜ₊₁ = pₜ + Δt v̂ₜ This is the loop shown in the render: neural activity -> decoded velocity -> cursor motion The cortical network and electrode threads show the measurement side. The N1 Implant is described as using 1,024 electrodes distributed across 64 flexible threads, each thinner than a human hair. The decoder panel shows the computational side with activity rₜ, decoded velocity v̂ₜ, and the cursor state pₜ changing over time. A noisy biological pattern becomes a state estimate. That estimate becomes motion on a screen. Therefore, the first lesson is not that Neuralink makes the brain a screen. For cursor control, the Mathematics is more precise: A small piece of intention is represented as a hidden state, measured through neural activity, decoded as a vector, and turned into action. #Neuralink #BrainComputerInterface #NeuralEngineering #Mathematics #StateEstimation #Neuroscience #MachineLearning #BiomedicalEngineering

Mathelirium

14,520 просмотров • 3 месяцев назад

Glad that our work “Inference-Time Enhancement of Generative Robot Policies via Predictive World Modeling”, led by Han Qi, has been accepted to IEEE Robotics and Automation Letters! 🎉 We propose Generative Predictive Control (GPC): sample action proposals from a pretrained diffusion policy (“look back”), roll them out with a diffusion-based action-conditioned video world model (“look forward”), then rank or optimize the actions using either a learned reward model or VLM preferences. Conceptually, this is trajectory optimization / MPC with hybrid sampling + gradient optimization, interpreted through modern diffusion priors and video world models. Interestingly, we first posted the paper on arXiv in Feb 2025, when action-conditioned video world models for planning were still rare—now this direction is rapidly gaining traction. Still many open questions, e.g., • how to avoid local minima in planning • what representations work best for world models • how to balance physics priors vs. data-driven learning Paper:

Glad that our work “Inference-Time Enhancement of Generative Robot Policies via Predictive World Modeling”, led by Han Qi, has been accepted to IEEE Robotics and Automation Letters! 🎉 We propose Generative Predictive Control (GPC): sample action proposals from a pretrained diffusion policy (“look back”), roll them out with a diffusion-based action-conditioned video world model (“look forward”), then rank or optimize the actions using either a learned reward model or VLM preferences. Conceptually, this is trajectory optimization / MPC with hybrid sampling + gradient optimization, interpreted through modern diffusion priors and video world models. Interestingly, we first posted the paper on arXiv in Feb 2025, when action-conditioned video world models for planning were still rare—now this direction is rapidly gaining traction. Still many open questions, e.g., • how to avoid local minima in planning • what representations work best for world models • how to balance physics priors vs. data-driven learning Paper:

Heng Yang

18,994 просмотров • 5 месяцев назад

Introducing yes&, the social platform built for creation. Creativity has never happened in isolation. A style. A pose. A sound. An outfit. We draw inspiration from the work we admire, then turn it into something new. But inspiration and creation have always lived in separate places. We save Pinterest boards, bookmark TikToks, screenshot Instagram posts, and fill our camera rolls with ideas, yet rarely turn them into creations of our own. AI changes that. On yes&, anything that inspires you can become a starting point. Discover great work, explore the prompts, skills, and creative process behind it, then remix it into something that’s yours. Every remix preserves the lineage of the idea and recognizes the creators who helped shape it. When creating becomes as simple as taking a photo, creativity naturally becomes social. Ideas evolve through inspiration, remixing, and co-creation. Today, Markus Mak and I are excited to share that we’ve raised $3M in pre-seed funding, led by Accel , with support from Parable and an incredible group of partners, to build the infrastructure for collaborative creativity. Come make your next remix:

Introducing yes&, the social platform built for creation. Creativity has never happened in isolation. A style. A pose. A sound. An outfit. We draw inspiration from the work we admire, then turn it into something new. But inspiration and creation have always lived in separate places. We save Pinterest boards, bookmark TikToks, screenshot Instagram posts, and fill our camera rolls with ideas, yet rarely turn them into creations of our own. AI changes that. On yes&, anything that inspires you can become a starting point. Discover great work, explore the prompts, skills, and creative process behind it, then remix it into something that’s yours. Every remix preserves the lineage of the idea and recognizes the creators who helped shape it. When creating becomes as simple as taking a photo, creativity naturally becomes social. Ideas evolve through inspiration, remixing, and co-creation. Today, Markus Mak and I are excited to share that we’ve raised $3M in pre-seed funding, led by Accel , with support from Parable and an incredible group of partners, to build the infrastructure for collaborative creativity. Come make your next remix:

Lena

406,732 просмотров • 2 месяцев назад

New model: your robot can now pack your suitcase 🧳 Xiaomi has released a new robot foundation model. Called Xiaomi-Robotics-1, it is designed to have a robot pick things up and move them around. But first, DEFINITIONS: - Mixture-of-Transformers (MoT): An architecture where separate transformer "experts" (e.g., one for vision-language, one for actions) share a single attention stream, so each modality gets specialized parameters without losing joint reasoning. - Vision-language model (VLM): A model that jointly understands images and text. - Diffusion transformer: A transformer trained to turn noise into structured outputs by iterative denoising, here generating robot actions rather than images. - Action chunks: Short sequences of future actions (e.g., the next ~50 motor commands) predicted in one shot instead of one step at a time. - Flow matching: A faster version of diffusion. The model learns a straight-line velocity field from noise to the target action, so it needs only a few integration steps instead of many denoising ones. Its peculiarity comes from its two stage training: 1. 100,000 hours of video shot through a UMI rig: a handheld 3D-printed gripper with a camera, worn by humans doing ordinary tasks in homes, shops, factories and offices. 2. Adapt to actual robot bodies with ~10,000 hours of real-robot data. It replaces the standard approach of teleoperating a real robot for every hour of training data. Its architecture is a Mixture-of-Transformers pairing a pre-trained Qwen3-VL vision-language model with a diffusion transformer that emits action chunks via flow matching, released in 2.6B, 5.1B and 10.5B parameter variants. However, if you read the entire paper ("Scaling VLA Models with over 100K Hours"), you realize that all of the scaling experiments on 20k hours. Therefore the headline "out-of-the-box success climbing 26% → 75% as pre-training data grows" tops out at 100% of 20k hours! What the full corpus does to that curve is never shown -> and this where things would become interesting! Xiaomi's own conclusion is that model size has stopped mattering and data is the binding constraint. The performance gap among different model sizes are less pronounced than those observed across different data scales. This result suggests that model capacity at the billions-parameter scale may already be sufficient to capture the current dataset's distribution. Which further asks the same question: why not use the 100k video hours? Anyway, I would definitely love to have a couple robots at home that can cooperate to pack my suitcase with items relevant to my next destination:

New model: your robot can now pack your suitcase 🧳 Xiaomi has released a new robot foundation model. Called Xiaomi-Robotics-1, it is designed to have a robot pick things up and move them around. But first, DEFINITIONS: - Mixture-of-Transformers (MoT): An architecture where separate transformer "experts" (e.g., one for vision-language, one for actions) share a single attention stream, so each modality gets specialized parameters without losing joint reasoning. - Vision-language model (VLM): A model that jointly understands images and text. - Diffusion transformer: A transformer trained to turn noise into structured outputs by iterative denoising, here generating robot actions rather than images. - Action chunks: Short sequences of future actions (e.g., the next ~50 motor commands) predicted in one shot instead of one step at a time. - Flow matching: A faster version of diffusion. The model learns a straight-line velocity field from noise to the target action, so it needs only a few integration steps instead of many denoising ones. Its peculiarity comes from its two stage training: 1. 100,000 hours of video shot through a UMI rig: a handheld 3D-printed gripper with a camera, worn by humans doing ordinary tasks in homes, shops, factories and offices. 2. Adapt to actual robot bodies with ~10,000 hours of real-robot data. It replaces the standard approach of teleoperating a real robot for every hour of training data. Its architecture is a Mixture-of-Transformers pairing a pre-trained Qwen3-VL vision-language model with a diffusion transformer that emits action chunks via flow matching, released in 2.6B, 5.1B and 10.5B parameter variants. However, if you read the entire paper ("Scaling VLA Models with over 100K Hours"), you realize that all of the scaling experiments on 20k hours. Therefore the headline "out-of-the-box success climbing 26% → 75% as pre-training data grows" tops out at 100% of 20k hours! What the full corpus does to that curve is never shown -> and this where things would become interesting! Xiaomi's own conclusion is that model size has stopped mattering and data is the binding constraint. The performance gap among different model sizes are less pronounced than those observed across different data scales. This result suggests that model capacity at the billions-parameter scale may already be sufficient to capture the current dataset's distribution. Which further asks the same question: why not use the 100k video hours? Anyway, I would definitely love to have a couple robots at home that can cooperate to pack my suitcase with items relevant to my next destination:

Léo

15,662 просмотров • 23 дней назад

📢📢 𝐀𝐯𝐚𝐭𝟑𝐫 📢📢 Avat3r creates high-quality 3D head avatars from just a few input images in a single forward pass with a new dynamic 3DGS reconstruction model. Video: Project: Our core idea is to make Gaussian Reconstruction Models animatable. We find that a simple cross-attention to an expression code sequence is already sufficient to model complex facial expressions. We then incorporate position maps from DUSt3R and feature maps from Sapiens to facilitate the prediction task. While DUSt3R's position maps act as a pixel-aligned initialization for the Gaussians' positions, the Sapiens feature maps help the cross-view transformer to match corresponding image tokens in the 4 input images. One major challenge in creating a 3D head avatar from smartphone images comes from inconsistent facial expressions when the subject could not remain perfectly static during the capture. We eliminate this static requirement by simply showing our model input images with different facial expressions during training. This technique makes our model robust to inconsistent input images later on. Finally, we show that despite the model has been trained with 4 input images, one can even create a 3D head avatar when only a single image is available. To achieve this, we employ a pre-trained 3D GAN to lift the single image to 3D and then render the 4 input images for our model. This allows us to create 3D head avatars from single images and even highly out-of-distribution examples like AI generated faces, paintings or statues. Great work by Tobias Kirschstein from his internship at Meta with Javier Romero, Artem Sevastopolsky, and Shunsuke Saito

Matthias Niessner

74,763 просмотров • 1 год назад

Evan Spiegel on the lesson that killed his first startup and led to Snapchat: "We focused on building the perfect product for way too long before we got feedback. We worked for like eighteen months to build this perfect full-featured product, which was in direct contravention to how I was always taught to build things. Build a prototype, build an MVP, get it in front of people, learn as quickly as possible. But we had spent all this time building this perfect piece of software and we hadn't thought enough about distribution. While we built this great piece of software, our competitor at the time, Naviance, had secured distribution through all the different college counselors. What piece of software are you going to choose to help your kid get into college? The one recommended by the college counselors or the one from two kids at Stanford? I think it's a pretty easy choice. So we saw very early that we had no distribution advantage. Even if we loved our software, people weren't going to use it because we didn't have a scalable way to get it in people's hands. Around that time when we saw the emergence of the App Store on iPhone, it was very clear that was a distribution channel we could really use and benefit from. But we also needed to build things we could build quickly, things we really were going to use together with our friends so we could be the first early customers. Ultimately Peekaboo and Snapchat represented that." This is exactly the asymmetry Peter Thiel describes in Zero to One: superior distribution by itself can create a monopoly, even with no product differentiation. The converse is not true. Thiel puts an even finer point on it: most businesses get zero distribution channels to work. Poor sales rather than bad product is the most common cause of failure. If you can get just one distribution channel to work, you have a great business. Estée Lauder said: it's not enough to have the most wonderful product in the world. You must be able to sell it. Many founders over-index on product perfection and under-invest in distribution. The world is full of great products nobody knows about, and mediocre products with massive distribution that dominate markets. Evan Spiegel lived it. He built the better product, watched it lose, and then built something he could actually distribute.

Evan Spiegel on the lesson that killed his first startup and led to Snapchat: "We focused on building the perfect product for way too long before we got feedback. We worked for like eighteen months to build this perfect full-featured product, which was in direct contravention to how I was always taught to build things. Build a prototype, build an MVP, get it in front of people, learn as quickly as possible. But we had spent all this time building this perfect piece of software and we hadn't thought enough about distribution. While we built this great piece of software, our competitor at the time, Naviance, had secured distribution through all the different college counselors. What piece of software are you going to choose to help your kid get into college? The one recommended by the college counselors or the one from two kids at Stanford? I think it's a pretty easy choice. So we saw very early that we had no distribution advantage. Even if we loved our software, people weren't going to use it because we didn't have a scalable way to get it in people's hands. Around that time when we saw the emergence of the App Store on iPhone, it was very clear that was a distribution channel we could really use and benefit from. But we also needed to build things we could build quickly, things we really were going to use together with our friends so we could be the first early customers. Ultimately Peekaboo and Snapchat represented that." This is exactly the asymmetry Peter Thiel describes in Zero to One: superior distribution by itself can create a monopoly, even with no product differentiation. The converse is not true. Thiel puts an even finer point on it: most businesses get zero distribution channels to work. Poor sales rather than bad product is the most common cause of failure. If you can get just one distribution channel to work, you have a great business. Estée Lauder said: it's not enough to have the most wonderful product in the world. You must be able to sell it. Many founders over-index on product perfection and under-invest in distribution. The world is full of great products nobody knows about, and mediocre products with massive distribution that dominate markets. Evan Spiegel lived it. He built the better product, watched it lose, and then built something he could actually distribute.

David Senra

27,151 просмотров • 4 месяцев назад

CLIP by hand ✍️ ~ 13 steps walkthrough below CLIP, Contrastive Language-Image Pre-training, is OpenAI's answer to a question that sounds impossible: how do you put a sentence and a picture in the same space? CLIP shipped when OpenAI was still open, and those embeddings were shared far and wide. Almost every multimodal model you use today descends from them. How does it work? Goal: learn one shared embedding space for text and images. = 1. Given = A mini batch of three text-image pairs. OpenAI trained the original on 400 million. = 2. Text to vectors = Let us look up each word with word2vec. = 3. Image to vectors = We cut each image into two patches and flatten them. Now text and pixels are both just numbers. = 4. The other pairs = Repeat steps 2 and 3 for the rest of the batch. = 5. Encode = Let us push both sides through their encoders, a linear layer and a ReLU. In practice these are transformers, but the shape of the operation is the same. = 6. Mean pooling = We average across the columns, so each image and each sentence collapses to a single vector. = 7. Projection = The text vectors are 3D and the image vectors are 4D, so they cannot be compared at all. A linear layer projects both to 2D. That 2D space is the shared embedding space, and getting here is the whole point of the model. = 8. Prepare for matmul = Let us copy the text vectors down and the transposed image vectors across. = 9. MatMul = We multiply, which takes the dot product of every text vector with every image vector. Each cell is one estimate of how well a sentence matches a picture. = 10. Softmax, e to the power = Raise e to each cell. To keep it hand sized we approximate e with 3. = 11. Softmax, sum = Sum each row for image to text, each column for text to image. = 12. Softmax, normalize = Divide, and out come two similarity matrices, one per direction. = 13. Loss gradients = The targets are identity matrices: a pair that belongs together should score 1, every other cell 0. Subtract the target from the similarity and you have the gradients, in both directions. The takeaway: pairing a picture with a sentence comes down to a single dot product. Everything before step 9 is the work of getting them into one shared space, so that the dot product finally means something. 💾 Save this post!

CLIP by hand ✍️ ~ 13 steps walkthrough below CLIP, Contrastive Language-Image Pre-training, is OpenAI's answer to a question that sounds impossible: how do you put a sentence and a picture in the same space? CLIP shipped when OpenAI was still open, and those embeddings were shared far and wide. Almost every multimodal model you use today descends from them. How does it work? Goal: learn one shared embedding space for text and images. = 1. Given = A mini batch of three text-image pairs. OpenAI trained the original on 400 million. = 2. Text to vectors = Let us look up each word with word2vec. = 3. Image to vectors = We cut each image into two patches and flatten them. Now text and pixels are both just numbers. = 4. The other pairs = Repeat steps 2 and 3 for the rest of the batch. = 5. Encode = Let us push both sides through their encoders, a linear layer and a ReLU. In practice these are transformers, but the shape of the operation is the same. = 6. Mean pooling = We average across the columns, so each image and each sentence collapses to a single vector. = 7. Projection = The text vectors are 3D and the image vectors are 4D, so they cannot be compared at all. A linear layer projects both to 2D. That 2D space is the shared embedding space, and getting here is the whole point of the model. = 8. Prepare for matmul = Let us copy the text vectors down and the transposed image vectors across. = 9. MatMul = We multiply, which takes the dot product of every text vector with every image vector. Each cell is one estimate of how well a sentence matches a picture. = 10. Softmax, e to the power = Raise e to each cell. To keep it hand sized we approximate e with 3. = 11. Softmax, sum = Sum each row for image to text, each column for text to image. = 12. Softmax, normalize = Divide, and out come two similarity matrices, one per direction. = 13. Loss gradients = The targets are identity matrices: a pair that belongs together should score 1, every other cell 0. Subtract the target from the similarity and you have the gradients, in both directions. The takeaway: pairing a picture with a sentence comes down to a single dot product. Everything before step 9 is the work of getting them into one shared space, so that the dot product finally means something. 💾 Save this post!

Tom Yeh

20,750 просмотров • 27 дней назад

Marc Andreessen on the idea maze and getting a prototype working before raising money “You see people sometimes who say, ‘I want to start a company’ and then they work through the process of coming up with an idea. Generally those don’t work as well as the case where somebody has the idea first and then they realize there’s an opportunity to build a company.” This was the case for Marc himself - they had Mosaic working before they started Netscape. Larry Page and Sergey Brin had the Google search engine working at Stanford. Pierre Omidyar had eBay working before he left his job. “It’s also far easier to raise money. The ideal pitch we receive is ‘Here’s the thing that works. Would you like to invest in our company or not?’ That’s so much easier than 30 slides with a dream.” At a16z Balaji Srinivasan came up with the concept of the Idea Maze, which Marc explains: “There’s a mythology that these ideas arrive like magic or people stumble into them… The reality usually with the big successes is that the founder has been chewing on the problem for 5-10 years before they start the company… So they’re a true domain expert.” He continues: “We call it the idea maze because for any idea there’s all these different permutations. Who should the customer be? What shape should the product have? How should we take it to market?… The really smart founders have thought through all these scenarios by the time they go out to raise money. And they have detailed answers on every one of those fronts because they put so much thought into it. The more haphazard founders haven’t thought about any of that. And it’s the detailed ones who tend to do much better.” Source: Lex Fridman (Jun 2023)

Marc Andreessen on the idea maze and getting a prototype working before raising money “You see people sometimes who say, ‘I want to start a company’ and then they work through the process of coming up with an idea. Generally those don’t work as well as the case where somebody has the idea first and then they realize there’s an opportunity to build a company.” This was the case for Marc himself - they had Mosaic working before they started Netscape. Larry Page and Sergey Brin had the Google search engine working at Stanford. Pierre Omidyar had eBay working before he left his job. “It’s also far easier to raise money. The ideal pitch we receive is ‘Here’s the thing that works. Would you like to invest in our company or not?’ That’s so much easier than 30 slides with a dream.” At a16z Balaji Srinivasan came up with the concept of the Idea Maze, which Marc explains: “There’s a mythology that these ideas arrive like magic or people stumble into them… The reality usually with the big successes is that the founder has been chewing on the problem for 5-10 years before they start the company… So they’re a true domain expert.” He continues: “We call it the idea maze because for any idea there’s all these different permutations. Who should the customer be? What shape should the product have? How should we take it to market?… The really smart founders have thought through all these scenarios by the time they go out to raise money. And they have detailed answers on every one of those fronts because they put so much thought into it. The more haphazard founders haven’t thought about any of that. And it’s the detailed ones who tend to do much better.” Source: Lex Fridman (Jun 2023)

Startup Archive

21,345 просмотров • 1 месяц назад