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.
Related event: NVIDIA Introduces PDD to Accelerate Image and Video Generation(3 posts)→
More from Multimodal
- Seedance 2.0 chaos prompt turns Circuit City into a surreal AI meme — techhalla · 2026-07-30
- Seedance 2.0 chaos prompt produces a mutant catfish meme — techhalla · 2026-07-30
- Testing Microsoft Mage-Flow Against Qwen and Krea Image Editors — Curious-Hurry-9149 · 2026-07-30
- Google DeepMind Launches Lyria 3.5 Music Model with Enhanced Vocals and Stems Export — jasonbaldridge · 2026-07-30
- Four months of Gabor-wavelet image generation led to a broader Claude-assisted research project — pixlpa · 2026-07-30
- Optimized Qwen Image 2512: 5x Smaller, 3x Faster Inference — enrique-byteshape · 2026-07-30