NVIDIA's PDD Distills Diffusion Models to 4-8 Steps for Faster Image/Video Generation
NVIDIA and others propose Parallel Decoding Distillation (PDD), a minimal trajectory distillation method to accelerate image and video generation with diffusion and flow matching models. PDD predicts multiple ODE steps in a single network evaluation, breaking the conventional approach of merging consecutive ODE steps into one large step, achieving state-of-the-art (SOTA) quality and diversity in just 4-8 steps. This offers a new path for speeding up generative AI inference, especially valuable for compute-intensive image and video generation tasks.
Confirmed
- Core innovation: PDD directly predicts multiple ODE steps in a single network evaluation, rather than predicting denoising steps step-by-step, simplifying the distillation process.
- Performance: PDD achieves SOTA generation quality and diversity under variable network forward passes (NFE).
- Implementation advantage: PDD requires only 4-8 steps for distillation acceleration, and the authors emphasize it is easier to implement than consistency-model-based methods.
Why it matters
- Significantly reduces inference time costs for diffusion and flow matching models, offering high practical value for compute-intensive image and video generation tasks.
- Provides a simpler alternative to traditional consistency models, facilitating broader adoption of related technologies.
2026-07-29 ~ 2026-07-30 · 7 related posts
Primary sources
- [source] NVIDIA’s PDD speeds image and video generation with 4–8-step distillation — nvidia · 2026-07-29
- PDD’s follow-up reiterates a faster path for image and video generation — ArashVahdat · 2026-07-29
- PDD: Ultra-Simple Distillation for Few-Step Video Generation — ArashVahdat · 2026-07-30
- PDD: Parallel Decoding Distillation Predicts Multiple ODE Steps in One Eval — ArashVahdat · 2026-07-30
- NVIDIA's PDD Accelerates Diffusion Models by Predicting Multiple ODE Steps — ArashVahdat · 2026-07-30
- NVIDIA's PDD Paper: Parallel Decoding Distillation for Faster Image/Video Generation — Scriabinical · 2026-07-30
1 near-duplicate retellings: ArashVahdat