正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

🚨Announcing our #ICLR2025 Oral! 🔥Diffusion LMs are on the rise for parallel text generation! But unlike autoregressive LMs, they struggle with quality, fixed-length constraints & lack of KV caching. 🚀Introducing Block Diffusion—combining autoregressive and diffusion models for the best of both worlds! 👇1/7

Marianne Arriola @ICML

1,535 subscribers

150,079 次观看 • 1 年前 •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

0 条评论

暂无评论

原始帖子的评论将显示在这里

相关视频

Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models "Block diffusion sequentially generates blocks of tokens by performing diffusion within each block and conditioning on previous blocks. By combining strength from autoregressive and diffusion models, block diffusion overcomes the limitations of both approaches by supporting variable-length, higher-quality generation and improving inference efficiency with KV caching and parallel sampling."

Block Diffusion: Interpolating Between Autoregressive and Diffusion Language Models "Block diffusion sequentially generates blocks of tokens by performing diffusion within each block and conditioning on previous blocks. By combining strength from autoregressive and diffusion models, block diffusion overcomes the limitations of both approaches by supporting variable-length, higher-quality generation and improving inference efficiency with KV caching and parallel sampling."

Tanishq Mathew Abraham, Ph.D.

21,813 次观看 • 1 年前

MaskINT: Video Editing via Interpolative Non-autoregressive Masked Transformers paper page: Recent advances in generative AI have significantly enhanced image and video editing, particularly in the context of text prompt control. State-of-the-art approaches predominantly rely on diffusion models to accomplish these tasks. However, the computational demands of diffusion-based methods are substantial, often necessitating large-scale paired datasets for training, and therefore challenging the deployment in practical applications. This study addresses this challenge by breaking down the text-based video editing process into two separate stages. In the first stage, we leverage an existing text-to-image diffusion model to simultaneously edit a few keyframes without additional fine-tuning. In the second stage, we introduce an efficient model called MaskINT, which is built on non-autoregressive masked generative transformers and specializes in frame interpolation between the keyframes, benefiting from structural guidance provided by intermediate frames. Our comprehensive set of experiments illustrates the efficacy and efficiency of MaskINT when compared to other diffusion-based methodologies. This research offers a practical solution for text-based video editing and showcases the potential of non-autoregressive masked generative transformers in this domain.

MaskINT: Video Editing via Interpolative Non-autoregressive Masked Transformers paper page: Recent advances in generative AI have significantly enhanced image and video editing, particularly in the context of text prompt control. State-of-the-art approaches predominantly rely on diffusion models to accomplish these tasks. However, the computational demands of diffusion-based methods are substantial, often necessitating large-scale paired datasets for training, and therefore challenging the deployment in practical applications. This study addresses this challenge by breaking down the text-based video editing process into two separate stages. In the first stage, we leverage an existing text-to-image diffusion model to simultaneously edit a few keyframes without additional fine-tuning. In the second stage, we introduce an efficient model called MaskINT, which is built on non-autoregressive masked generative transformers and specializes in frame interpolation between the keyframes, benefiting from structural guidance provided by intermediate frames. Our comprehensive set of experiments illustrates the efficacy and efficiency of MaskINT when compared to other diffusion-based methodologies. This research offers a practical solution for text-based video editing and showcases the potential of non-autoregressive masked generative transformers in this domain.

AK

25,449 次观看 • 2 年前

Switti -- a new scale-wise transformer for text-to-image generation 🦾 🔥 Improved generation of fine-grained details. Outperforms existing T2I AR models and competes with state-of-the-art T2I diffusion models while being up to 7x faster.

Switti -- a new scale-wise transformer for text-to-image generation 🦾 🔥 Improved generation of fine-grained details. Outperforms existing T2I AR models and competes with state-of-the-art T2I diffusion models while being up to 7x faster.

Gradio

29,314 次观看 • 1 年前

Show-o One Single Transformer to Unify Multimodal Understanding and Generation discuss: We present a unified transformer, i.e., Show-o, that unifies multimodal understanding and generation. Unlike fully autoregressive models, Show-o unifies autoregressive and (discrete) diffusion modeling to adaptively handle inputs and outputs of various and mixed modalities. The unified model flexibly supports a wide range of vision-language tasks including visual question-answering, text-to-image generation, text-guided inpainting/extrapolation, and mixed-modality generation. Across various benchmarks, it demonstrates comparable or superior performance to existing individual models with an equivalent or larger number of parameters tailored for understanding or generation. This significantly highlights its potential as a next-generation foundation model.

Show-o One Single Transformer to Unify Multimodal Understanding and Generation discuss: We present a unified transformer, i.e., Show-o, that unifies multimodal understanding and generation. Unlike fully autoregressive models, Show-o unifies autoregressive and (discrete) diffusion modeling to adaptively handle inputs and outputs of various and mixed modalities. The unified model flexibly supports a wide range of vision-language tasks including visual question-answering, text-to-image generation, text-guided inpainting/extrapolation, and mixed-modality generation. Across various benchmarks, it demonstrates comparable or superior performance to existing individual models with an equivalent or larger number of parameters tailored for understanding or generation. This significantly highlights its potential as a next-generation foundation model.

AK

124,048 次观看 • 1 年前

1/ Happy to share UniDisc - Unified Multimodal Discrete Diffusion – We train a 1.5 billion parameter transformer model from scratch on 250 million image/caption pairs using a **discrete diffusion objective**. Our model has all the benefits of diffusion models but now in multimodal space! - flexible compute-quality tradeoff, zero-shot inpainting and editing, better control via classifier-free guidance and lower latency! We open source everything - our code, weights and the training dataset.

1/ Happy to share UniDisc - Unified Multimodal Discrete Diffusion – We train a 1.5 billion parameter transformer model from scratch on 250 million image/caption pairs using a discrete diffusion objective. Our model has all the benefits of diffusion models but now in multimodal space! - flexible compute-quality tradeoff, zero-shot inpainting and editing, better control via classifier-free guidance and lower latency! We open source everything - our code, weights and the training dataset.

Mihir Prabhudesai

104,934 次观看 • 1 年前

DimensionX: Create Any 3D and 4D Scenes from a Single Image with Controllable Video Diffusion TL;DR: Create 3/4DGS from Video Diffusion Note: Some first inference code released (not all yet). Contributions (cited): • We present DimensionX, a novel framework for generating photorealistic 3D and 4D scenes from only a single image using controllable video diffusion. • We propose ST-Director, which decouples the spatial and temporal priors in video diffusion models by learning (spatial and temporal) dimension-aware modules with our curated datasets. We further enhance the hybriddimension control with a training-free composition approach according to the essence of video diffusion denoising process. • To bridge the gap between video diffusion and real-world scenes, we design a trajectory-aware mechanism for 3D generation and an identity-preserving denoising approach for 4D generation, enabling more realistic and controllable scene synthesis. • Extensive experiments manifest that our DimensionX delivers superior performance in video, 3D, and 4D generation compared with baseline methods.

DimensionX: Create Any 3D and 4D Scenes from a Single Image with Controllable Video Diffusion TL;DR: Create 3/4DGS from Video Diffusion Note: Some first inference code released (not all yet). Contributions (cited): • We present DimensionX, a novel framework for generating photorealistic 3D and 4D scenes from only a single image using controllable video diffusion. • We propose ST-Director, which decouples the spatial and temporal priors in video diffusion models by learning (spatial and temporal) dimension-aware modules with our curated datasets. We further enhance the hybriddimension control with a training-free composition approach according to the essence of video diffusion denoising process. • To bridge the gap between video diffusion and real-world scenes, we design a trajectory-aware mechanism for 3D generation and an identity-preserving denoising approach for 4D generation, enabling more realistic and controllable scene synthesis. • Extensive experiments manifest that our DimensionX delivers superior performance in video, 3D, and 4D generation compared with baseline methods.

MrNeRF

17,052 次观看 • 1 年前

🚀 Self-speculation brings 6.75x real speedup for LLM generation with SGLang inference! Same model drafts future tokens in Diffusion mode → then verifies them in AR (causal) mode. One model and one KV cache. Just different attention masks. Thanks to perfect alignment, we get 2× longer acceptance lengths than MTP techniques (Eagle-3, MTP, dFlash). We run 2 forward passes… but the 2× higher acceptance means we break even - and with zero overhead from extra drafter, KV cache, or LM head that comes with MTP - those are not free. Last week we released Nemotron-Labs-Diffusion + Tri-mode LLMs! We did continued pre-training on Ministral-3 models by switching attention patterns (block causal bidirectional). Result: one model that runs AR mode, Diffusion mode, and Self-Speculation. Diffusion mode already shows high benchmark accuracy - excited to see what happens when someone beats left-to-right acceptance! 🔥 Github: Paper: SGLang inference: Try the models on HF:

🚀 Self-speculation brings 6.75x real speedup for LLM generation with SGLang inference! Same model drafts future tokens in Diffusion mode → then verifies them in AR (causal) mode. One model and one KV cache. Just different attention masks. Thanks to perfect alignment, we get 2× longer acceptance lengths than MTP techniques (Eagle-3, MTP, dFlash). We run 2 forward passes… but the 2× higher acceptance means we break even - and with zero overhead from extra drafter, KV cache, or LM head that comes with MTP - those are not free. Last week we released Nemotron-Labs-Diffusion + Tri-mode LLMs! We did continued pre-training on Ministral-3 models by switching attention patterns (block causal bidirectional). Result: one model that runs AR mode, Diffusion mode, and Self-Speculation. Diffusion mode already shows high benchmark accuracy - excited to see what happens when someone beats left-to-right acceptance! 🔥 Github: Paper: SGLang inference: Try the models on HF:

Pavlo Molchanov

66,554 次观看 • 2 个月前

Chop the gradients ✂️! We found that truncating decoder gradients in latent video diffusion to a fixed window allows us to finetune on videos with pixel-wise perceptual losses without running out of memory. Pixel losses have been essential for image generation and reconstruction, but until now, they haven't scaled to long-duration, high-resolution video diffusion due to recursive activation accumulation in causal decoders, leading to OOM during training 💥📉. Project: Video diffusion models can do a lot more 🚀 when you can backprop the decoder! Post-process neural rendered scenes, super-resolve videos, harmonize lighting in controlled synthetic driving scenes, and inpaint videos — all in a single step ⚡ with a quick finetune from a standard diffusion model.

Chop the gradients ✂️! We found that truncating decoder gradients in latent video diffusion to a fixed window allows us to finetune on videos with pixel-wise perceptual losses without running out of memory. Pixel losses have been essential for image generation and reconstruction, but until now, they haven't scaled to long-duration, high-resolution video diffusion due to recursive activation accumulation in causal decoders, leading to OOM during training 💥📉. Project: Video diffusion models can do a lot more 🚀 when you can backprop the decoder! Post-process neural rendered scenes, super-resolve videos, harmonize lighting in controlled synthetic driving scenes, and inpaint videos — all in a single step ⚡ with a quick finetune from a standard diffusion model.

Felix Heide

28,323 次观看 • 3 个月前

🎥 Video generation is hitting the memory wall. As videos get longer, the KV cache quietly explodes — and long-horizon consistency starts to break. We built Quant VideoGen: a training-free KV cache compression method for auto-regressive video diffusion. Instead of storing every KV in high precision, QVG exploits video’s spatiotemporal redundancy with semantic-aware smoothing + progressive residual quantization. 🚀 Up to 7× KV memory reduction ⚡ <4% overhead ✅ Strong long-video quality 🕹️ Deploy HYWorldPlay on your own RTX 5090 locally KV compression is becoming a core scaling primitive — not just for LLMs, but for video generation too. Paper: Code: (1/5)

🎥 Video generation is hitting the memory wall. As videos get longer, the KV cache quietly explodes — and long-horizon consistency starts to break. We built Quant VideoGen: a training-free KV cache compression method for auto-regressive video diffusion. Instead of storing every KV in high precision, QVG exploits video’s spatiotemporal redundancy with semantic-aware smoothing + progressive residual quantization. 🚀 Up to 7× KV memory reduction ⚡ <4% overhead ✅ Strong long-video quality 🕹️ Deploy HYWorldPlay on your own RTX 5090 locally KV compression is becoming a core scaling primitive — not just for LLMs, but for video generation too. Paper: Code: (1/5)

Haocheng Xi

65,008 次观看 • 3 个月前

Today, we are releasing Stable Video Diffusion, our first foundation model for generative AI video based on the image model, Stable Diffusion. As part of this research preview, the code, weights, and research paper are now available. Additionally, today you can sign up for our waitlist to access a new upcoming web experience featuring a Text-To-Video interface. To access the model & sign up for our waitlist, visit our website here:

Today, we are releasing Stable Video Diffusion, our first foundation model for generative AI video based on the image model, Stable Diffusion. As part of this research preview, the code, weights, and research paper are now available. Additionally, today you can sign up for our waitlist to access a new upcoming web experience featuring a Text-To-Video interface. To access the model & sign up for our waitlist, visit our website here:

Stability AI

1,024,532 次观看 • 2 年前

The Hidden Language of Diffusion Models paper page: tackle the challenge of understanding concept representations in text-to-image models by decomposing an input text prompt into a small set of interpretable elements. This is achieved by learning a pseudo-token that is a sparse weighted combination of tokens from the model's vocabulary, with the objective of reconstructing the images generated for the given concept. Applied over the state-of-the-art Stable Diffusion model, this decomposition reveals non-trivial and surprising structures in the representations of concepts. For example, we find that some concepts such as "a president" or "a composer" are dominated by specific instances (e.g., "Obama", "Biden") and their interpolations. Other concepts, such as "happiness" combine associated terms that can be concrete ("family", "laughter") or abstract ("friendship", "emotion"). In addition to peering into the inner workings of Stable Diffusion, our method also enables applications such as single-image decomposition to tokens, bias detection and mitigation, and semantic image manipulation

The Hidden Language of Diffusion Models paper page: tackle the challenge of understanding concept representations in text-to-image models by decomposing an input text prompt into a small set of interpretable elements. This is achieved by learning a pseudo-token that is a sparse weighted combination of tokens from the model's vocabulary, with the objective of reconstructing the images generated for the given concept. Applied over the state-of-the-art Stable Diffusion model, this decomposition reveals non-trivial and surprising structures in the representations of concepts. For example, we find that some concepts such as "a president" or "a composer" are dominated by specific instances (e.g., "Obama", "Biden") and their interpolations. Other concepts, such as "happiness" combine associated terms that can be concrete ("family", "laughter") or abstract ("friendship", "emotion"). In addition to peering into the inner workings of Stable Diffusion, our method also enables applications such as single-image decomposition to tokens, bias detection and mitigation, and semantic image manipulation

AK

41,746 次观看 • 3 年前

We took a 30B model and split it in two to write tokens in parallel instead of one at a time. Introducing Nemotron-Labs-TwoTower: a diffusion language model from NVIDIA Research adapted from Nemotron-3-Nano-30B-A3B. Here’s how it works: one half holds the context, the other writes the tokens, with both reusing the pretrained model instead of training a new one from scratch. We found it kept 98.7% of the original model’s quality at 2.42× faster generation.

We took a 30B model and split it in two to write tokens in parallel instead of one at a time. Introducing Nemotron-Labs-TwoTower: a diffusion language model from NVIDIA Research adapted from Nemotron-3-Nano-30B-A3B. Here’s how it works: one half holds the context, the other writes the tokens, with both reusing the pretrained model instead of training a new one from scratch. We found it kept 98.7% of the original model’s quality at 2.42× faster generation.

NVIDIA AI

759,036 次观看 • 27 天前

Decentralized Diffusion Models power stronger models trained on more accessible infrastructure. DDMs mitigate the networking bottleneck that locks training into expensive and power-hungry centralized clusters. They scale gracefully to billions of parameters and generate photorealistic images with just a week of training on eight independent GPU nodes. They’re easy to implement, adopt DiT hyperparameters directly and outperform standard models FLOP-for-FLOP.

Decentralized Diffusion Models power stronger models trained on more accessible infrastructure. DDMs mitigate the networking bottleneck that locks training into expensive and power-hungry centralized clusters. They scale gracefully to billions of parameters and generate photorealistic images with just a week of training on eight independent GPU nodes. They’re easy to implement, adopt DiT hyperparameters directly and outperform standard models FLOP-for-FLOP.

David McAllister

46,415 次观看 • 1 年前

Wonderland: Navigating 3D Scenes from a Single Image Contributions: • First, we introduce a representation for controllable 3D generation by leveraging the generative priors from camera-guided video diffusion models. Unlike image models, video diffusion models are trained on extensive video datasets. This enables them to capture comprehensive spatial relationships within scenes across multiple views and embed a form of "3D awareness" in their latent space, which allows us to maintain 3D consistency in novel view synthesis. • Second, to achieve controllable novel view generation, we empower video models with precise control over specified camera motions. We introduce a novel dual-branch conditioning mechanism that effectively incorporates desired diverse camera trajectories into the video diffusion model. This enables expansion of a single image into a multi-view consistent capture of a 3D scene with precise pose control. • Third, to achieve efficient 3D reconstruction, we directly transform video latents into 3DGS. We propose a novel latent-based large reconstruction model (LaLRM) that lifts video latents to 3D in a feed-forward manner. With this design, during inference, our model directly predicts 3DGS from a single input image, effectively aligning the generation and reconstruction tasks—and bridging image space and 3D space—through the video latent space. Compared with reconstructing scenes from images, the video latent space offers a 256× spatial-temporal reduction while retaining essential and consistent 3D structural details. Such a high degree of compression is crucial, as it allows the LaLRM to handle a wider range of 3D scenes within the reconstruction framework, with the same memory constraints.

Wonderland: Navigating 3D Scenes from a Single Image Contributions: • First, we introduce a representation for controllable 3D generation by leveraging the generative priors from camera-guided video diffusion models. Unlike image models, video diffusion models are trained on extensive video datasets. This enables them to capture comprehensive spatial relationships within scenes across multiple views and embed a form of "3D awareness" in their latent space, which allows us to maintain 3D consistency in novel view synthesis. • Second, to achieve controllable novel view generation, we empower video models with precise control over specified camera motions. We introduce a novel dual-branch conditioning mechanism that effectively incorporates desired diverse camera trajectories into the video diffusion model. This enables expansion of a single image into a multi-view consistent capture of a 3D scene with precise pose control. • Third, to achieve efficient 3D reconstruction, we directly transform video latents into 3DGS. We propose a novel latent-based large reconstruction model (LaLRM) that lifts video latents to 3D in a feed-forward manner. With this design, during inference, our model directly predicts 3DGS from a single input image, effectively aligning the generation and reconstruction tasks—and bridging image space and 3D space—through the video latent space. Compared with reconstructing scenes from images, the video latent space offers a 256× spatial-temporal reduction while retaining essential and consistent 3D structural details. Such a high degree of compression is crucial, as it allows the LaLRM to handle a wider range of 3D scenes within the reconstruction framework, with the same memory constraints.

MrNeRF

52,849 次观看 • 1 年前

World modeling and imitation learning have largely been considered two disparate worlds. In our recent work, Unified World Models, just accepted to #RSS2025, Chuning Zhu provides a dead-simple unifying solution: just train a joint diffusion model over actions and future states, but with *decoupled* diffusion time steps across these modalities. Manipulating these decoupled time steps then allows for marginalization or conditioning on actions or states; a single model can serve as a policy, forward dynamics model, video prediction model, or inverse dynamics model by simply setting diffusion timesteps carefully. The resulting model can leverage video datasets along with robot training data much more effectively, and shows improved robustness, generalization, and flexibility. This is exciting because it is frustratingly simple, scalable, and shows strong improvement on real-world robotics problems. Please refer to Chuning Zhu 's excellent thread for more details! More details/code can be found on our website and in the paper -

World modeling and imitation learning have largely been considered two disparate worlds. In our recent work, Unified World Models, just accepted to #RSS2025, Chuning Zhu provides a dead-simple unifying solution: just train a joint diffusion model over actions and future states, but with decoupled diffusion time steps across these modalities. Manipulating these decoupled time steps then allows for marginalization or conditioning on actions or states; a single model can serve as a policy, forward dynamics model, video prediction model, or inverse dynamics model by simply setting diffusion timesteps carefully. The resulting model can leverage video datasets along with robot training data much more effectively, and shows improved robustness, generalization, and flexibility. This is exciting because it is frustratingly simple, scalable, and shows strong improvement on real-world robotics problems. Please refer to Chuning Zhu 's excellent thread for more details! More details/code can be found on our website and in the paper -

Abhishek Gupta

11,430 次观看 • 1 年前

🚀New paper out - We present Video-MSG (Multimodal Sketch Guidance), a novel planning-based training-free guidance method for T2V models, improving control of spatial layout and object trajectories. 🔧 Key idea: • Generate a Video Sketch — a spatio-temporal plan with background, foreground, and motion in the pixel space. • Encode this structure directly into the latent space of the diffusion model during generation, which does not require fine-tuning or additional memory during inference. 🧵

🚀New paper out - We present Video-MSG (Multimodal Sketch Guidance), a novel planning-based training-free guidance method for T2V models, improving control of spatial layout and object trajectories. 🔧 Key idea: • Generate a Video Sketch — a spatio-temporal plan with background, foreground, and motion in the pixel space. • Encode this structure directly into the latent space of the diffusion model during generation, which does not require fine-tuning or additional memory during inference. 🧵

Jialu Li

35,060 次观看 • 1 年前

Here are more results from #RigidFormer: predicting physical dynamics with purely neural simulators — an attempt to learn physical dynamics in a scalable manner. 🤖 1) Controllable Articulated Body Simulation — More Results Additional Unitree G1 humanoid rollouts under controlled motion. Each sample uses a different initial state and control signal (direction and velocity). 🏺 2) Object Fragmentation Simulating the cracking and fragmentation process of objects. Thanks Žiga Kovačič for suggesting this experiment! 🎬 3) Combining Rigidformer with Diffusion-as-Shader for controllable video generation. Note: the meshes shown here are only for visualization — the network takes point clouds as input and predicts the updated state of each point.

Here are more results from #RigidFormer: predicting physical dynamics with purely neural simulators — an attempt to learn physical dynamics in a scalable manner. 🤖 1) Controllable Articulated Body Simulation — More Results Additional Unitree G1 humanoid rollouts under controlled motion. Each sample uses a different initial state and control signal (direction and velocity). 🏺 2) Object Fragmentation Simulating the cracking and fragmentation process of objects. Thanks Žiga Kovačič for suggesting this experiment! 🎬 3) Combining Rigidformer with Diffusion-as-Shader for controllable video generation. Note: the meshes shown here are only for visualization — the network takes point clouds as input and predicts the updated state of each point.

Zhiyang (Frank) Dou

20,955 次观看 • 2 个月前

[SIGGRAPH ASIA '25] Detail-Enhanced Gaussian Splatting for Large-Scale Volumetric Capture Contributions: - A two-stage approach to performance capture, combining a scene-scale capture rig and a single-actor facial capture rig. - A novel high-quality scene-scale volumetric performance capture rig, incorporating both static and dynamic cameras to track the performance of multiple actors. - A reconstruction pipeline for dynamic performance capture, featuring stable calibration of moving cameras and 4DGS with improved dynamic range and color fidelity. - A detail enhancement Diffusion Model, which supports 4K, RGB, and Alpha, with improved temporal stability.

[SIGGRAPH ASIA '25] Detail-Enhanced Gaussian Splatting for Large-Scale Volumetric Capture Contributions: - A two-stage approach to performance capture, combining a scene-scale capture rig and a single-actor facial capture rig. - A novel high-quality scene-scale volumetric performance capture rig, incorporating both static and dynamic cameras to track the performance of multiple actors. - A reconstruction pipeline for dynamic performance capture, featuring stable calibration of moving cameras and 4DGS with improved dynamic range and color fidelity. - A detail enhancement Diffusion Model, which supports 4K, RGB, and Alpha, with improved temporal stability.

MrNeRF

42,456 次观看 • 9 个月前

Most recent diffusion language model research (that I’ve seen) seems to be using masking as the noising process. It looks like, however, most closed-source models (Google Gemini Diffusion and possibly Inception Labs’ Mercury) use a different noising process, where instead of masking tokens, they replace them with different tokens (either with a random token or a semantically similar token). I wondered how they were getting such high throughput with the latter noising process, since I believed that optimizing inference with KVCache approximation would be more difficult (for various reasons). I visualized this noising process with tiny-diffusion and compared it to normal unmasking, and was very surprised to see how fast the generation “settles” into a reasonable output, and then only slightly refines afterwards, requiring much fewer steps in total. Unmasking (where tokens are never remasked, the typical implementation) is inherently limited in generation speed by the fact that an increase in tokens decoded per step leads to more errors due to the mismatch between individual and marginal token probability distributions we sample from. The token replacement noising process seems to have a much different set of characteristics. Because we sample each token per step, every token makes “progress” towards the final output each iteration (in addition to *potentially* giving other tokens more information in future steps). Generally, masking has outperformed other noising processes, which is probably why most research focused on it (using smaller models). But the paper referred to in the retweet shows that random replacement as a noising process may scale better as model size increases. Big labs might have noticed these results much earlier (due to having drastically more training resources and being able to test larger models), which may explain the discrepancy in the choice of noising process. I’m gonna test this with larger models, since tiny-diffusion only has 10M parameters.

Most recent diffusion language model research (that I’ve seen) seems to be using masking as the noising process. It looks like, however, most closed-source models (Google Gemini Diffusion and possibly Inception Labs’ Mercury) use a different noising process, where instead of masking tokens, they replace them with different tokens (either with a random token or a semantically similar token). I wondered how they were getting such high throughput with the latter noising process, since I believed that optimizing inference with KVCache approximation would be more difficult (for various reasons). I visualized this noising process with tiny-diffusion and compared it to normal unmasking, and was very surprised to see how fast the generation “settles” into a reasonable output, and then only slightly refines afterwards, requiring much fewer steps in total. Unmasking (where tokens are never remasked, the typical implementation) is inherently limited in generation speed by the fact that an increase in tokens decoded per step leads to more errors due to the mismatch between individual and marginal token probability distributions we sample from. The token replacement noising process seems to have a much different set of characteristics. Because we sample each token per step, every token makes “progress” towards the final output each iteration (in addition to potentially giving other tokens more information in future steps). Generally, masking has outperformed other noising processes, which is probably why most research focused on it (using smaller models). But the paper referred to in the retweet shows that random replacement as a noising process may scale better as model size increases. Big labs might have noticed these results much earlier (due to having drastically more training resources and being able to test larger models), which may explain the discrepancy in the choice of noising process. I’m gonna test this with larger models, since tiny-diffusion only has 10M parameters.

nathan (in sf)

40,440 次观看 • 6 个月前

You can't 3D reconstruct glass from images... ...WRONG! Thanks for video diffusion, now just about anything is possible! Introducing...Diffusion Knows Transparency (DKT) Transparent and reflective objects usually break robot vision and photogrammetry pipelines because they don't follow the "solid object" rules standard cameras expect. DKT is a new AI model that repurposes the "internal physics engine" found in video generation models to solve this problem. Researchers took a massive video diffusion model (WAN) and fine-tuned it using a custom-built synthetic dataset to turn it into a high-precision depth sensor. To train the AI, they built the first massive synthetic video library of transparent objects, 1.32 million frames of perfectly labeled glass and metal objects in motion. Without ever seeing a "real" labeled video of glass during training, the model (DKT) outperformed all previous specialized systems on real-world benchmarks (ClearPose, DREDS). They created a "lightweight" 1.3B parameter version that runs fast enough (0.17s per frame) to be used on actual robot hardware. Two reasons I find this project important: 1. It further proves that synthetic data will be essential for training the next generation vision models. 2. In real-world robotic tests, using DKT's depth maps nearly doubled the success rate of robot arms trying to pick up objects on tricky reflective or translucent surfaces. At home robots will need to interact with these types of objects on a daily basis. Check out the project page here: Code is LIVE! #Computervision #Robotics #AI

You can't 3D reconstruct glass from images... ...WRONG! Thanks for video diffusion, now just about anything is possible! Introducing...Diffusion Knows Transparency (DKT) Transparent and reflective objects usually break robot vision and photogrammetry pipelines because they don't follow the "solid object" rules standard cameras expect. DKT is a new AI model that repurposes the "internal physics engine" found in video generation models to solve this problem. Researchers took a massive video diffusion model (WAN) and fine-tuned it using a custom-built synthetic dataset to turn it into a high-precision depth sensor. To train the AI, they built the first massive synthetic video library of transparent objects, 1.32 million frames of perfectly labeled glass and metal objects in motion. Without ever seeing a "real" labeled video of glass during training, the model (DKT) outperformed all previous specialized systems on real-world benchmarks (ClearPose, DREDS). They created a "lightweight" 1.3B parameter version that runs fast enough (0.17s per frame) to be used on actual robot hardware. Two reasons I find this project important: 1. It further proves that synthetic data will be essential for training the next generation vision models. 2. In real-world robotic tests, using DKT's depth maps nearly doubled the success rate of robot arms trying to pick up objects on tricky reflective or translucent surfaces. At home robots will need to interact with these types of objects on a daily basis. Check out the project page here: Code is LIVE! #Computervision #Robotics #AI

Jonathan Stephens

17,712 次观看 • 6 个月前