Parallel Decoding Distillation Adopted by Alibaba, NVIDIA, Amazon Weeks After Release
ArashVahdat · x · 2026-10-22
- Parallel Decoding Distillation (PDD), released end of July, has already been adopted by six teams: Alibaba's accelerated LoRAs for MiniMax-H3, WorldPlay2's causal distillation, FastVideo's FastH3 Omni, NVIDIA's continuous diffusion LLM, Amazon's generative forecasting model, and NVIDIA/MSR extensions on parametric student trajectories.
- PDD's appeal is its simplicity: a single regression-based MSE loss with one student and one detached teacher forward — no data, no adversarial loss, no JVP or finite differences, no extra trainable modules or stabilization tricks. It claims SOTA quality with variable NFE and is the first trajectory distillation to scale to few-step video generation.
- Official 8-step acceleration LoRAs (rank=64, BF16) for MiniMax-H3 (FL2VA and Ref2VA) are live on Hugging Face, generating 768p video at 4/8 NFE; the authors plan to open-source their implementation soon.
More from Multimodal
- AI-generated 'Dark Knight' images nail cinematic blur and film grain — egeberkina · 2026-10-22
- MOSS-Image ecosystem roundup: 33,937-character offline catalog, low-VRAM LoRA training and more — optimisticalish · 2026-10-22
- Analyst predicts $8-10B Skydance-SpaceX deal to fuel Grok Imagine's AI films — beffjezos · 2026-10-22
- GPT-Image 2 still unbeaten 6 months on, Google's Nano Banana 2.1 fumbles — flowersslop · 2026-10-22
- FIBO 2.0 Edit Turbo Does Text-to-Image Too: 4 Steps, No CFG, Tested on 22 Prompts — TimeTruth2490 · 2026-10-22
- Atlas full fly-through demo stuns viewers: "the matrix might be real" — XRarchitect · 2026-10-22