Open-source Video Delta Net hits near-lossless 75-90x speedup, generating 14s of 768p video in 11s on 8 B200s
cHHillee · x · 2026-09-03
The open-source Video Delta Net (VDN) project makes video generation faster than playback: built on a MiniMax H3 backbone, it produces a 14.4-second 768p clip in 11.23 seconds on 8× NVIDIA B200 GPUs with 8 denoising steps — a 75-90x speedup at near-lossless quality.
- Hybrid attention: a highly efficient frame-wise linear attention branch paired with a softmax branch that preserves the backbone's visual quality and consistency
- Plug-and-play: delivered as a new linear attention branch plus two small LoRA adapters merged into the backbone at inference, leaving backbone weights untouched
- Fully open-source: weights, optimized inference stack, training code, and a technical blog are all released
The reposter also praised the repo's engineering (PyTorch 2.13 + FlashAttention 4 via FlexAttention's Flash backend). A notable milestone for open-source video generation approaching real-time generation.
Related event: Open-Source Video Delta Net Pushes Video Generation Faster Than Playback(8 posts)→
More from Multimodal
- Claude Code directed a full music video: 6.5 hours, 765M input tokens — colinmcnamara · 2026-10-21
- Kling 4.0 debuts native 30-second AI video with 10 keyframes, extendable to 2 minutes — lmoroney · 2026-10-21
- Mistral's Voxtral transcribes an 1891 Eiffel phonograph recording at 6.5% WER — IsForAt · 2026-10-21
- Redditor remakes SD clown dance video at 25 steps with motion boost and fixed prompts — roychodraws · 2026-10-21
- Fibo-2, the #1 open image editing model, releases weights on Hugging Face — linoy_tsaban · 2026-10-21
- Nous Research teases generative media plugin for Desktop, already links Krea/Fal — NousResearch · 2026-10-21