Neta Shaul retweeted
Adaptive tokenization is becoming an essential tool for making visual generation more efficient. In our new work, Level-of-Token (LoT) Diffusion, we allocate tokens based on where details are.
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🧵
Neta Shaul retweeted
Great to see our Parallel Decoding Distillation (PDD) already being applied to MiniMax-H3🚀
Alibaba scaled it up just a month after our paper release (before our code even dropped) - guess that proves how simple it is to implement!
Learn more about PDD: xitter.cf/julberner/status/20825…
Another cool stuff from Alibaba!
MiniMax-H3 Acc LoRAs.
Parallel Decoding Distillation for efficient video gen in a few steps.
FL2VA + Ref2VA
huggingface.co/alibaba-pai/M…
Neta Shaul retweeted
My talk on accelerated diffusion and flow models at @BerkeleyLab's DL4Sci School is online!
Check it out for an overview on how to:
- Reduce the number of denoising steps
- Accelerate each step
📺 youtube.com/watch?v=ZnRzw9iK… (jump to 22:18 to learn more about the video below)
1/ New paper: Parallel Decoding Distillation (PDD)!🚀
To accelerate diffusion/flow models, trajectory distillation usually merges consecutive ODE steps into 1 larger step. PDD changes this by predicting MULTIPLE ODE steps in 1 network eval.🧵
w/ @julberner @ArashVahdat Chao Liu
9/ In addition to video, we validate PDD on text-to-image generation using Qwen-Image at 1024x1024 resolution.
At 4 NFE, PDD consistently improves compared to DMD2 (Qwen-Lightning) across the three benchmarks: One-IG, DPG-Bench, and GenEval. 👇
Code release is coming soon!
Our project page and paper are available now at:
Project page: research.nvidia.com/labs/gen…
Paper: arxiv.org/abs/2607.26004
Neta Shaul retweeted
Why choose between autoregressive (fast, streaming) and bidirectional (high-quality) video diffusion when a single model can do both even better?
Excited to share Flex-Forcing, accepted as spotlight at #ICML2026!
📍Say hi to Chao Liu at our poster: July 9 | 2:30 - 4:15 PM KST | Hall A #3909
Joint work with @horseeeMa @wn8_nie @ArashVahdat @XinchaoWang3
Flow Matching obtains its training supervising velocity by conditioning on data samples. But what if you don’t have a dataset, only an unnormalized density? Flow Sampling gets the supervising drift by conditioning on the source instead of the target.
Catch @aaronjhavens at #ICML
Excited to be at ICML to present our Spotlight paper: Flow Sampling!
We propose a simple fixed-point objective for learning diffusion samplers, built on the flow matching marginal construction. More details to come soon!
Paper: arxiv.org/abs/2605.03984
Neta Shaul retweeted
Check out our project’ code and the cool demo that @BrianCChao did!
Excited to release the code and model weights for our recent paper "Foveated Diffusion: Efficient Spatially Adaptive Image and Video Generation"! We hope this inspires continued research on mixed-resolution diffusion and efficient token allocation for image and video generation.
📰 Paper: arxiv.org/abs/2603.23491
💻 Code: github.com/bchao1/foveated_d…
🌐 Website: bchao1.github.io/foveated-di…
🤗 Weights: huggingface.co/bchao1/foveat…
🚀 Demo: huggingface.co/spaces/bchao1…
We put together a fun demo on Hugging Face Spaces — try it out here: huggingface.co/spaces/bchao1…. Sketch your own tokenization layouts and pick model variants to run mixed-resolution diffusion live!
Congrats @YarivLior! Very interesting direction for accelerating flow/diffusion models🚀
After multiple requests for the code of the visuals from my talk about Transition Matching, I made a notebook that reproduces the DTM vs. FM GIF!
This demo is a good way to build intuition on how TM and FM differ.
github.com/neta93/visual-ill…
@urielsinger
Very cool! Transition matching for fast inference.
🚀🎬We introduce TMD (Transition Matching Distillation): 480p videos generated from text prompts in < 3 NFEs!
1️⃣Main backbone for feature extraction and lightweight head for iterative refinement
2️⃣Distilled from Wan2.1 14B T2V combining MeanFlow & DMD2
🔗research.nvidia.com/labs/gen…
Incredible atmosphere at the poster session today! Thanks to everyone who visited 🙌
I’ll be at #NeurIPS2025 until Dec 8. If you missed the poster or just want to chat, my DMs are open.
Shoutout to @EliahuHorwitz for the pictures!
#NeurIPS2025 "Transition Matching" explores generative Markov processes with expressive transition kernels, going beyond the Gaussian kernel used in diffusion and flow models.
Interested? Let's chat! 📍 Poster #3609 🕒 Wed at 11am - 2pm
📄 arxiv.org/abs/2506.23589
