AI
Improved Distributional Diffusion Models
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 28/09/2026, 20:00:00 UTC
Briefing created: 30/09/2026, 14:10:31 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 30/09/2026, 13:18:24 UTC.
Briefings
Distributional Diffusion Models (DDMs) replace mean-prediction denoisers with distributional objectives, learning stochastic approximations to image probability distributions. This study mitigates scaling obstacles by deferring particle expansion to late transformer layers and introducing time-dependent scoring rule schedules informed by Biroli2024. Combined with a DiT-based latent setup, the method enables practical training on class-conditional ImageNet-256^2, achieving 4.48 FID at 4 steps and 2.38 at 50 steps with DiT-XL/2, from a single model trained from scratch without teacher, self-distillation, or JVPs. The FID remains stable across sampling budgets, and the recipe transfers to text-to-image generation.
Key points
- DDMs replace mean-prediction denoisers with distributional objectives to learn stochastic approximations to image probability distributions.
- The method achieves 4.48 FID at 4 steps and 2.38 at 50 steps with DiT-XL/2 on ImageNet-256^2 from a single model trained from scratch.
Context and limitations
Results are limited to ImageNet-256^2 class-conditional generation; stability across sampling budgets is confirmed for 4 to 50 NFE steps, but broader generalization and text-to-image transfer specifics remain unverified.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 30/09/2026, 14:10:31 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeiezmsyalybmcbok4ybbkkvos54gcfjbl7sdlf3wuiud6geq5q5szq