Faster Than Playback — On MiniMax H3

UC Berkeley and Impossible, Inc. just dropped Video DeltaNet (VDN) — a hybrid attention stack that keeps MiniMax H3’s look while racing past the old dense-attention tax. Their VDN-H3 checkpoint is the proof: livestream-speed video generation that still feels like full Softmax quality.

On the OpenVDN reference stack with 8-step distillation and SGLang Diffusion, an 8×B200 setup finishes DiT denoising for a 14.3–14.4s, 768p clip in about 6.7–6.9 seconds — roughly 14.5× versus a 50-step dense H3 baseline on the same GPU count. After warm-up, the finished video lands around **9 seconds**. That’s not “faster than coffee.” That’s faster than playback.

What Video DeltaNet Actually Changes

Attention is the tax on long video diffusion. Softmax over huge spatiotemporal sequences eats the clock; pure linear attention is quicker but often loses identity, layout, and temporal glue. VDN refuses the false choice.

It splits video–video attention into two complementary paths:

  • Local Softmax — sliding-window exact attention over nearby frame chunks (plus boundary anchors) so fine detail and short-term stability stay sharp
  • Bidirectional Video Delta Attention — a frame-wise linear memory that carries long-range context without quadratic cost

Text and audio keep Softmax where it matters; the hybrid targets the video–video bottleneck. Learnable gates and separate output projections keep the two branches from fighting each other on scale.

Modes Creators Can Drive

The same VDN-H3 checkpoint supports live generation modes you’d actually ship with:

  • T2VA — text to video (+ audio)
  • I2VA — image-conditioned
  • FL2VA — first-and-last-frame conditioned (plus related first/last variants on the project page)

Qualitative claims from the team: near-lossless versus dense MiniMax H3, with stronger quality and instruction following than MiniMax FastH3 in their comparisons. Treat demos as the vibe check — then verify on your own prompts.

Why This Matters for Gen Creators

If you already think in MiniMax H3 workflows, VDN-H3 is the acceleration lane: same creative surface, far less wait. Weights are up on Hugging Face as OpenVDN/vdn-minimax-h3, with the paper on HF Papers and the full story on the project site. Pair that with ArtRealmAI Gen when you want to stay in the H3 family without babysitting a research cluster.

Try this on Gen → https://artrealmai.com/gen?utm_source=magazine&utm_campaign=gen&utm_content=video-deltanet-vdn-h3

Original Source

https://openvdn.github.io/

https://huggingface.co/papers/2609.20744

Conclusion

Video DeltaNet is hybrid attention done the video-native way — local Softmax for the frames that need fidelity, linear memory for the long arc — and VDN-H3 shows MiniMax H3 can outrun its own playback clock without throwing quality overboard. Bookmark the project page, grab the weights, and time a 14-second clip for yourself.

—Aurelia ♡