Looped Diffusion Transformer: the same blocks run several times inside each denoising step, and loop depth can be changed at sampling time [R]

Paper: Looped Diffusion Transformer (arXiv 2609.40305, 30 Sep). Instead of making a text-to-image DiT bigger, it runs a shared group of transformer blocks several times inside each denoising step. The authors report that naive looping doesn't reliably help, and fix it with deep supervision on every loop plus self-modulating attention. Their headline claims: a 260M looped model beats a model 6.5x larger on several benchmarks with 4.9x less inference compute, and at a fixed inference budget, more loop depth helps more than more denoising steps.

The B/16 code and weights are MIT (the weights are on Hugging Face under sensenova), so I ran it on an M5 Pro: PyTorch MPS in fp32, 100 Euler steps, CFG 6.0, and one seed per prompt reused at loop depths 1 to 4. It was trained at depth 4, and the README says other depths work without retraining. The grid shows five of the prompts I tried.

Most of the change happens before depth 3, and depths 3 and 4 come out very close on every prompt shown. The interesting part to me is that depth is picked at sampling time, the way the iteration count is in looped and recurrent-depth language models, so you can trade compute for refinement per image. On this machine an image took about 23 to 24 seconds at depth 1 and 43 to 47 at depth 4.