Do coding agents understand the dynamics of the world well enough to reconstruct it from video?
We introduce 4DCodeBench to evaluate this ability through 4D inverse graphics.
https://t.co/QlkzTPZpbi
<🧵>
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🧵
I’m happy to share that I will be presenting my first paper “Dual Ascent Diffusion for Inverse Problems” at #CVPR this week! Please check out our poster on 🕞Sat, Jun 6, 2026 3:45 PM – 5:45 PM PDT at 📍ExHall A & F 204!
🔍 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 🧵👇
@GordonWetzstein Very excited to present this paper “Dual Ascent Diffusion for Inverse Problems” at #CVPR this week! Please check out our poster on 🕞Sat, Jun 6, 2026 3:45 PM – 5:45 PM PDT at 📍ExHall A & F 204!
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
Just released our paper CrystalBoltz💎: an end-to-end framework that steers diffusion using experimental X-ray data, recovering structures in minutes when sequence-only prediction fails, even under large conformational changes. Paper and project page below, would love to discuss!
AlphaFold-like models such as Boltz-2 and AlphaFold3 learn powerful priors over structures from PDB. But how can we make them aware of new X-ray crystallography data?
Introducing CrystalBoltz: end-to-end X-ray structure determination via experiment-guided diffusion.
1/4 🧵
💫AsymFlow🤩
x0- and u-prediction: two ends of the same spectrum.
Pick the middle, stop wasting capacity on full-rank noise. 1.57 FID on a plain transformer and the first viable 9B latent-to-pixel lift!
Joint work with @HanshengCh, @jan_on_x, @GordonWetzstein, @GuibasLeonidas
New paper: AsymFlow🔥
JiT x0-prediction is not enough for pixel generation. Better keep velocity in a low-rank subspace:
- 1.57 FID on ImageNet (best pixel flow model)
- Finetunes FLUX.2 klein into pixel space, beats the original on HPSv3/DPG/GenEval (#1 overall on HPSv3)
1/7
@HanshengCh The reframing of x0- and u-prediction as two ends of the same spectrum is the part I keep coming back to. 1.57 FID on a plain transformer and the first viable 9B latent-to-pixel lift feel like the start of what this parameterization opens up, not the ceiling 🙌