Gordon Wetzstein

Generative priors enable 4D reconstruction from monocular video with unprecedented photorealism!

We are hiring at Rhoda AI! Join us if you are excited about world models and robotics and want to work on the challenge of our generation, i.e., generalist robotics, with an amazing team and real customers.

Aside from the official code release, I am thrilled to share that Spectral Progressive Diffusion a.k.a. SPEED (arxiv.org/abs/2605.18736) is now integrated into @lmsysorg's SGLang (@sgl_project)! 🚀 Instead of always running diffusion at full resolution, SPEED progressively grows resolution across denoising steps, drastically cutting token count and achieving >2× speedup with no quality loss. SPEED is now supported in SGLang for FLUX.1 & 2, Z-Image, Qwen-Image, and Wan. Support for Ideogram 4 incoming. Try it out now: docs.sglang.io/docs/sglang-di…. [1/4]

Speed up your diffusion or flow matching model using our plug & play SPEED module! No distillation necessary.

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/foveate… 🌐 Website: bchao1.github.io/foveated-diffu… 🤗 Weights: huggingface.co/bchao1/foveate… 🚀 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!

I integrated our recent work, Spectral Progressive Diffusion a.k.a. SPEED (arxiv.org/abs/2605.18736), with the open-source @ideogram_ai model released yesterday. Turns out our method works out of the box and speeds up inference by up to 1.6× while preserving the high image quality! [1/4]

Work led by @SoniaMinseoKim, together with @axlevy0. 📄 Paper: arxiv.org/abs/2505.17353 💻 Code: github.com/SoniaMinseoKim… 🌐 Project page: soniaminseokim.github.io/ddiff/

In short, DDiff is an ADMM-inspired diffusion solver that delivers: ✅ Higher image quality across metrics ✅ Robustness to high measurement noise ✅ Better measurement fidelity / lower residual ✅ Faster sampling ✅ Latent-diffusion compatibility ✅ Provable convergence under mild assumptions, a first for diffusion-based posterior optimization

🔍 Ill-posed inverse problems (deblurring, inpainting, phase retrieval) need strong priors, and diffusion models are great ones. But how you use them matters. Posterior samplers fold the prior & data term into the reverse diffusion, needing costly score backprop or inner MCMC. MAP variable-splitting methods decouple them, but lack a dual variable to fix constraint violations. DDiff tackles both. At #CVPR2026 🧵👇

Work led by @howard_xhc, together with @jan_on_x and @boyang_deng. 📄 Paper: arxiv.org/abs/2606.02565 🌐 Project page: howardxiao.ca/foveated/ Please check out our paper Fast Forward on July 19 and our paper presentation on July 23 at 2:50 PM at #SIGGRAPH2026!

We further demonstrate that our framework works in real-world settings on a 200 MP Samsung ISOCELL HP2 prototype. For both object tracking and scene text recognition tasks, our sensor attention policy reproduces the smooth pursuit and saccadic scanpaths seen in simulations, achieving 200 MP bandwidth-efficient sensing under real-world conditions with sensor noise. [5/6]

The era of ultra-high-resolution imaging has arrived. Modern image sensors exceeding 200 MP resolution are common in smartphones, with over 400 MP sensors under development. However, the large number of pixels poses significant challenges for acquisition and processing, especially on edge devices. Which pixels should be acquired, and when, for bandwidth-efficient imaging and perception? We introduce Policy-based Foveated Imaging and Perception, an on-device, real-time, predictive, and task-aware framework that dynamically allocates sensor resolution to prioritize important regions under specific perception objectives. This paper will be presented at #SIGGRAPH2026! [1/6]

@CSProfKGD Excited to speak at this workshop!

Attending #CVPR2026, Denver? We have a stellar speaker lineup for “GeoFreeNVS: Geometry-Free Novel View Synthesis and Controllable Video Models.” Come early. We anticipate seats will fill up FAST! 1/2

Pixel diffusions are getting hot🔥, but are they really good at texture details? Here's a one-shot comparison (no cherry picking) - HiDream: plain x0 prediction - AsymFLUX.2: plain AsymFlow prediction - L2P: velocity prediction through detailer head