PDD accelerates image and video diffusion by predicting multiple denoising steps at once
ArashVahdat · x · 2026-07-29
Parallel Decoding Distillation speeds up video and image generation
A new paper, Parallel Decoding Distillation (PDD), proposes a simpler trajectory-based distillation method to accelerate diffusion and flow-matching models for fast image and video generation.
- The method predicts multiple denoising steps per network evaluation, instead of merging consecutive ODE steps into one larger step.
- It is designed to be compatible with any pre-trained model and supports variable NFE sampling.
- The authors say it avoids the optimization difficulties of VSD/adversarial losses, which can suffer from mode collapse and reduced diversity.
- Reported results reach state of the art at 4–8 NFE on LTX-2.3 Text-to-Video/Audio, Wan 14B Text-to-Video, and Qwen-Image Text-to-Image.
- The paper also claims a meaningful improvement in video diversity.
More from Multimodal
- Seedance 2.5 workflow: character reference sheets + structured prompt yields a 30s single-take fake tourist video — techhalla · 2026-10-16
- AI couple moves into a castle ruin: full Magnific + Seedance 2.5 prompt thread behind the viral clip — techhalla · 2026-10-16
- FLUX.3 Klein shines at consistent high-res photo editing on RTX 5090: 3000×4000 in ~230s — herbertseabra · 2026-10-16
- Smoky Velvet Noir: a reusable image-gen prompt template with three moody examples — LudovicCreator · 2026-10-16
- VideoGen-Agent: An 8B Multimodal Agent Turns Video Generation Into a Tool-Use Loop — jiqizhixin · 2026-10-16
- Redditor builds impressive seaplane sim project with Sonnet 5.5 + Blender — Wise_Shame_2853 · 2026-10-16