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。

Masked diffusion is a discrete analog of continuous diffusion. Continuous diffusion gradually adds Gaussian noise, then learns to reverse that process to recover data. Masked diffusion instead progressively replaces tokens with a mask token, then learns to fill them back in.

Alec Helbling

11,723 subscribers

24,757 次观看 • 2 个月前 •via X (Twitter)

Anya Rossi• Live Now

Private livecam show

5 条评论

aidan ewart 的头像

aidan ewart2 个月前

why is noising by replacing with [MASK] better than noising by replacing with a randomly sampled token (which seems more analogous to the noising done in normal diffusions)?

Alec Helbling 的头像

Alec Helbling2 个月前

Link to code for visualizations.

Sohir Maskey 的头像

Sohir Maskey2 个月前

Actually uniform diffusion is the discrete analog to continuous diffusion with Gaussian prior see eg

dennis 的头像

dennis2 个月前

I wonder how efficient mixing/ diverse discrete would be compared to continuous?

andy 的头像

andy2 个月前

I never thought words could be corrected by images in this way!

相关视频

🏃Motion Tokenizer Bridging Semantic and Kinematic Conditions🏃‍♀️ #MoTok is a diffusion-based tokenizer that unifies *perception-planning-control*, combining the strengths of continuous diffusion & discrete tokens - Project: - Code:

🏃Motion Tokenizer Bridging Semantic and Kinematic Conditions🏃‍♀️ #MoTok is a diffusion-based tokenizer that unifies perception-planning-control, combining the strengths of continuous diffusion & discrete tokens - Project: - Code:

Ziwei Liu

19,259 次观看 • 7 个月前

A lot of people think that diffusion LLMs == BERT. But masked dLLMs are just one form of text diffusion, popularized by Subham Sahoo and his group at Cornell. There's much more to explore in text diffusion land! Check out our conversation (link below)!

A lot of people think that diffusion LLMs == BERT. But masked dLLMs are just one form of text diffusion, popularized by Subham Sahoo and his group at Cornell. There's much more to explore in text diffusion land! Check out our conversation (link below)!

Julia Turc

26,124 次观看 • 6 个月前

MinerU-Diffusion A 2.5B diffusion-based OCR model that replaces slow autoregressive decoding with parallel block-wise diffusion, achieving up to 3.2x faster inference while improving robustness on complex documents with tables, formulas, and layouts.

MinerU-Diffusion A 2.5B diffusion-based OCR model that replaces slow autoregressive decoding with parallel block-wise diffusion, achieving up to 3.2x faster inference while improving robustness on complex documents with tables, formulas, and layouts.

DailyPapers

15,304 次观看 • 7 个月前

If you have a policy that uses diffusion/flow (e.g. diffusion VLA), you can run RL where the actor chooses the noise, which is then denoised by the policy to produce an action. This method, which we call diffusion steering (DSRL), leads to a remarkably efficient RL method! 🧵👇

If you have a policy that uses diffusion/flow (e.g. diffusion VLA), you can run RL where the actor chooses the noise, which is then denoised by the policy to produce an action. This method, which we call diffusion steering (DSRL), leads to a remarkably efficient RL method! 🧵👇

Sergey Levine

153,232 次观看 • 1 年前

As someone who builds with LLMs and diffusion models, this 14-minute heat equation video is the closest thing to a "why noise reversal actually works" explainer I've ever seen released for free. Everyone thinks diffusion is a novel AI concept. It's the heat equation from 1822. The forward process spreads noise like heat, and the reverse process is what your model learns. Bookmark & watch this weekend. Same math, from 19th-century physics to 2025 image generation.

As someone who builds with LLMs and diffusion models, this 14-minute heat equation video is the closest thing to a "why noise reversal actually works" explainer I've ever seen released for free. Everyone thinks diffusion is a novel AI concept. It's the heat equation from 1822. The forward process spreads noise like heat, and the reverse process is what your model learns. Bookmark & watch this weekend. Same math, from 19th-century physics to 2025 image generation.

fintex

15,604 次观看 • 2 个月前

This is getting interesting. In this iteration, the magnifying glass is tracked in real-time, feeding only what is seen through it to the diffusion process, revealing the invisible within a piece of clay. Real-time diffusion with Stream Diffusion + SDXL Turbo. #StableDiffusion #SDXLTurbo #StreamDiffusion #StabilityAI #Realtime

This is getting interesting. In this iteration, the magnifying glass is tracked in real-time, feeding only what is seen through it to the diffusion process, revealing the invisible within a piece of clay. Real-time diffusion with Stream Diffusion + SDXL Turbo. #StableDiffusion #SDXLTurbo #StreamDiffusion #StabilityAI #Realtime

Hugues Bruyère

26,645 次观看 • 2 年前

"MeshFlow: Efficient Artistic Mesh Generation via MeshVAE and Flow-based Diffusion Transformer" TL;DR: learns a continuous mesh latent space and generates vertices and connectivity in parallel with flow matching, producing quality 3D meshes up to 18× faster than autoregressive.

"MeshFlow: Efficient Artistic Mesh Generation via MeshVAE and Flow-based Diffusion Transformer" TL;DR: learns a continuous mesh latent space and generates vertices and connectivity in parallel with flow matching, producing quality 3D meshes up to 18× faster than autoregressive.

Alexandre Morgand

43,375 次观看 • 4 个月前

Diffusion models spend the same compute on a blank wall as on a face. But you often know in advance where the detail will be. Introducing Level-of-Token (LoT) Diffusion: we turn that knowledge into a multiresolution token layout, with fine tokens where detail is needed and coarse tokens elsewhere. 1/9🧵

Diffusion models spend the same compute on a blank wall as on a face. But you often know in advance where the detail will be. Introducing Level-of-Token (LoT) Diffusion: we turn that knowledge into a multiresolution token layout, with fine tokens where detail is needed and coarse tokens elsewhere. 1/9🧵

Gordon Wetzstein

57,677 次观看 • 15 天前

Gemma 4 Diffusion landed in vLLM last week. Day 0. First diffusion LLM natively supported in vLLM. Instead of one token at a time, it predicts 256 tokens at once and iteratively denoises them in parallel. Result: 1,000+ tokens per second at batch size 1 on a single H100. Built on Model Runner V2. Google Gemma

Gemma 4 Diffusion landed in vLLM last week. Day 0. First diffusion LLM natively supported in vLLM. Instead of one token at a time, it predicts 256 tokens at once and iteratively denoises them in parallel. Result: 1,000+ tokens per second at batch size 1 on a single H100. Built on Model Runner V2. Google Gemma

Red Hat AI

17,670 次观看 • 4 个月前

Speed and quality can finally coexist in diffusion-based language generation. Introducing DiDi-Instruct, a Discrete Diffusion Divergence Instruct method that distills a pre-trained discrete diffusion language model (dLLM) into a few-step student for ultra-fast generation. Built on integral KL-divergence minimization, DiDi-Instruct achieves up to 64× faster decoding, surpasses both its teacher and GPT-2, and cuts training time by 20×. Ultra-Fast Language Generation via Discrete Diffusion Divergence Instruct Paper: Code: Project: Our report: 📬 #PapersAccepted by Jiqizhixin

Speed and quality can finally coexist in diffusion-based language generation. Introducing DiDi-Instruct, a Discrete Diffusion Divergence Instruct method that distills a pre-trained discrete diffusion language model (dLLM) into a few-step student for ultra-fast generation. Built on integral KL-divergence minimization, DiDi-Instruct achieves up to 64× faster decoding, surpasses both its teacher and GPT-2, and cuts training time by 20×. Ultra-Fast Language Generation via Discrete Diffusion Divergence Instruct Paper: Code: Project: Our report: 📬 #PapersAccepted by Jiqizhixin

机器之心 JIQIZHIXIN

18,196 次观看 • 11 个月前

LLM-grounded Diffusion: Enhancing Prompt Understanding of Text-to-Image Diffusion Models with Large Language Models paper page: github: Recent advancements in text-to-image generation with diffusion models have yielded remarkable results synthesizing highly realistic and diverse images. However, these models still encounter difficulties when generating images from prompts that demand spatial or common sense reasoning. We propose to equip diffusion models with enhanced reasoning capabilities by using off-the-shelf pretrained large language models (LLMs) in a novel two-stage generation process. First, we adapt an LLM to be a text-guided layout generator through in-context learning. When provided with an image prompt, an LLM outputs a scene layout in the form of bounding boxes along with corresponding individual descriptions. Second, we steer a diffusion model with a novel controller to generate images conditioned on the layout. Both stages utilize frozen pretrained models without any LLM or diffusion model parameter optimization. We validate the superiority of our design by demonstrating its ability to outperform the base diffusion model in accurately generating images according to prompts that necessitate both language and spatial reasoning. Additionally, our method naturally allows dialog-based scene specification and is able to handle prompts in a language that is not well-supported by the underlying diffusion model.

LLM-grounded Diffusion: Enhancing Prompt Understanding of Text-to-Image Diffusion Models with Large Language Models paper page: github: Recent advancements in text-to-image generation with diffusion models have yielded remarkable results synthesizing highly realistic and diverse images. However, these models still encounter difficulties when generating images from prompts that demand spatial or common sense reasoning. We propose to equip diffusion models with enhanced reasoning capabilities by using off-the-shelf pretrained large language models (LLMs) in a novel two-stage generation process. First, we adapt an LLM to be a text-guided layout generator through in-context learning. When provided with an image prompt, an LLM outputs a scene layout in the form of bounding boxes along with corresponding individual descriptions. Second, we steer a diffusion model with a novel controller to generate images conditioned on the layout. Both stages utilize frozen pretrained models without any LLM or diffusion model parameter optimization. We validate the superiority of our design by demonstrating its ability to outperform the base diffusion model in accurately generating images according to prompts that necessitate both language and spatial reasoning. Additionally, our method naturally allows dialog-based scene specification and is able to handle prompts in a language that is not well-supported by the underlying diffusion model.

AK

83,681 次观看 • 3 年前

today, sync. is excited to announce react-1 a ten billion parameter masked video diffusion model built for one purpose: giving you full creative control over performances in post

today, sync. is excited to announce react-1 a ten billion parameter masked video diffusion model built for one purpose: giving you full creative control over performances in post

sync. labs

27,258 次观看 • 10 个月前

Diffusion is the foundational ML framework behind state-of-the-art AI image and video generation, including Sora, Midjourney and Google Veo. In this episode of Decoded, Ankit Gupta sits down with Francois Chaubard to discuss how diffusion works, walk through a code sample, and explain why everyone training models should understand it. 00:00 Intro 00:33 What is diffusion? 02:50 What are applications of diffusion today? 04:06 Key innovations 07:01 Code examples 19:25 The "squint test" 22:27 Other areas diffusion is widely accessible 24:49 Outro

Diffusion is the foundational ML framework behind state-of-the-art AI image and video generation, including Sora, Midjourney and Google Veo. In this episode of Decoded, Ankit Gupta sits down with Francois Chaubard to discuss how diffusion works, walk through a code sample, and explain why everyone training models should understand it. 00:00 Intro 00:33 What is diffusion? 02:50 What are applications of diffusion today? 04:06 Key innovations 07:01 Code examples 19:25 The "squint test" 22:27 Other areas diffusion is widely accessible 24:49 Outro

Y Combinator

74,604 次观看 • 9 个月前

Rewrote tiny-diffusion to be 3x smaller! Went from 951 lines to just 364, all contained in one file. As simple as possible, but not simpler. I also added a tiny GPT implementation as a comparison (312 lines, inspired by Andrej Karpathy). The two implementations are ~80% identical. The model architecture, training loop, tokenization, etc, only differ in 19 lines of code. The main differences are contained within two functions (generate and get_batch). The reason to include the GPT implementation was to show how similar autoregressive LMs are to diffusion LMs on an architectural level. Only *1* line of code in the architecture needs to be modified to support masked language diffusion instead of next-token prediction (by disabling causal masking). Link to the repo is in the comments

Rewrote tiny-diffusion to be 3x smaller! Went from 951 lines to just 364, all contained in one file. As simple as possible, but not simpler. I also added a tiny GPT implementation as a comparison (312 lines, inspired by Andrej Karpathy). The two implementations are ~80% identical. The model architecture, training loop, tokenization, etc, only differ in 19 lines of code. The main differences are contained within two functions (generate and get_batch). The reason to include the GPT implementation was to show how similar autoregressive LMs are to diffusion LMs on an architectural level. Only 1 line of code in the architecture needs to be modified to support masked language diffusion instead of next-token prediction (by disabling causal masking). Link to the repo is in the comments

Nathan Barry

161,578 次观看 • 9 个月前

I'm thrilled to announce the launch of ⚡️Flash Diffusion from Jasper! Earlier this year, with our acquisition of Clipdrop, we launched the Jasper AI Research Lab in Paris. Today, we are excited to release our first piece of groundbreaking research: the open-source distillation method, "Flash Diffusion". Flash Diffusion accelerates inference by 500%, reduces computing costs, and produces higher-quality image outputs. Dive into the details and discover how Flash Diffusion is set to revolutionize the field of AI and image synthesis. Read all about it here: Try a demo on Hugging Face:

I'm thrilled to announce the launch of ⚡️Flash Diffusion from Jasper! Earlier this year, with our acquisition of Clipdrop, we launched the Jasper AI Research Lab in Paris. Today, we are excited to release our first piece of groundbreaking research: the open-source distillation method, "Flash Diffusion". Flash Diffusion accelerates inference by 500%, reduces computing costs, and produces higher-quality image outputs. Dive into the details and discover how Flash Diffusion is set to revolutionize the field of AI and image synthesis. Read all about it here: Try a demo on Hugging Face:

Timothy Young

10,180 次观看 • 2 年前

MVDream: Multi-view Diffusion for 3D Generation paper page: propose MVDream, a multi-view diffusion model that is able to generate geometrically consistent multi-view images from a given text prompt. By leveraging image diffusion models pre-trained on large-scale web datasets and a multi-view dataset rendered from 3D assets, the resulting multi-view diffusion model can achieve both the generalizability of 2D diffusion and the consistency of 3D data. Such a model can thus be applied as a multi-view prior for 3D generation via Score Distillation Sampling, where it greatly improves the stability of existing 2D-lifting methods by solving the 3D consistency problem. Finally, we show that the multi-view diffusion model can also be fine-tuned under a few shot setting for personalized 3D generation, i.e. DreamBooth3D application, where the consistency can be maintained after learning the subject identity.

MVDream: Multi-view Diffusion for 3D Generation paper page: propose MVDream, a multi-view diffusion model that is able to generate geometrically consistent multi-view images from a given text prompt. By leveraging image diffusion models pre-trained on large-scale web datasets and a multi-view dataset rendered from 3D assets, the resulting multi-view diffusion model can achieve both the generalizability of 2D diffusion and the consistency of 3D data. Such a model can thus be applied as a multi-view prior for 3D generation via Score Distillation Sampling, where it greatly improves the stability of existing 2D-lifting methods by solving the 3D consistency problem. Finally, we show that the multi-view diffusion model can also be fine-tuned under a few shot setting for personalized 3D generation, i.e. DreamBooth3D application, where the consistency can be maintained after learning the subject identity.

AK

294,519 次观看 • 3 年前

We just brought flow maps to language modeling for one-step sequence generation 💥 Discrete diffusion is not necessary -- continuous flows over one-hot encodings achieve SoTA performance and ≥8.3× faster generation 🔥 We believe this is a major step forward for discrete generative modeling and language modeling alike. 🚀 Full thread from first author Chanhyuk Lee:

We just brought flow maps to language modeling for one-step sequence generation 💥 Discrete diffusion is not necessary -- continuous flows over one-hot encodings achieve SoTA performance and ≥8.3× faster generation 🔥 We believe this is a major step forward for discrete generative modeling and language modeling alike. 🚀 Full thread from first author Chanhyuk Lee:

Nicholas Boffi

44,657 次观看 • 8 个月前