DMAD distills MiniMax-H3 to 4 sampling steps, demo trailer made with GPT-6 screenplay
linoy_tsaban · x · 2026-10-04
The DMAD team released a video generation model that distills MiniMax-H3 down to just 4 sampling steps, publishing a project page, Hugging Face weights, code, and a paper.
The author wrote the trailer's story with GPT-6 and generated every shot with DMAD — their first time directing. They also plugged the team's DiffHDR work and invited others to try making short films or even 4K HDR movies with the tool.
More from Multimodal
- Recreating the fabled Fable 5.5 video with Opus 5.5 — near-identical result — AlchainHust · 2026-10-04
- Runway unveils open-weight world action model Praxis-1 with 0.95 sim-to-real correlation — CyberRobooo · 2026-10-04
- AI Video Pit John Wick Against Mr. Bean and the Internet Loves It — TinfoilTricorn · 2026-10-04
- Seedance 2.5 + GPT Image combo recreates Bruce Lee meme, full prompt shared — TinfoilTricorn · 2026-10-04
- Creator uses Gemini + GPT image models to co-produce a graphic novel storyboard — aiamblichus · 2026-10-04
- A Vampire Short Film Trailer Made Entirely with Seedance — Interesting_Tone6532 · 2026-10-04