--- title: MPEG-Diffusion emoji: 🎞️ colorFrom: purple colorTo: green sdk: gradio sdk_version: 6.32.0 app_file: app.py python_version: "3.12" startup_duration_timeout: 1h short_description: Mixed-resolution text-to-video with a self-selected region models: - johndpope/MiniMax-H3-MPEG-Diffusion --- # MPEG-Diffusion: Mixed-Resolution Video Generation Text-to-video demo for **[MPEG-Diffusion](https://huggingface.co/papers/2611.12069)** (Menn et al.) — a model that generates video the way video codecs think about it: 1. a **low-resolution track** denoises the whole frame cheaply, 2. a tiny linear **region head** self-selects the region that deserves detail, 3. a **high-resolution track** refines only that region through dual-resolution attention, 4. the two tracks **merge in pixel space** — full-res detail where it matters, at roughly half the compute of rendering everything at high resolution. This Space runs the paper's [MiniMax-H3 port](https://github.com/scrya-com/mpeg-diffusion) (Apache-2.0, vendored under `lot/` + `third_party/`) with the published research weights from [johndpope/MiniMax-H3-MPEG-Diffusion](https://huggingface.co/johndpope/MiniMax-H3-MPEG-Diffusion): the ridge-fitted KV up-maps and the v3 region head applied to the int8 ConvRot base DiT ([Comfy-Org/MiniMax-H3](https://huggingface.co/Comfy-Org/MiniMax-H3)), at the port's *quality preset* — 20 steps, the region chosen at step 16, 6 high-resolution steps, pixel-space merge, on a short reference-length clip. By default the model picks its own detail region; a manual box can be forced under *Advanced*. Because the 62 GiB Qwen3-VL text encoder does not fit beside the transformer on one GPU, prompt conditioning runs on the [qwen3vl-conditioner](https://huggingface.co/spaces/multimodalart/qwen3vl-conditioner) Space and this Space runs the sampler. ## Credits - **Paper**: *MPEG-Diffusion: Mixed-Resolution Video Generation* — Menn et al., arXiv 2611.12069. - **Code**: the paper's MiniMax-H3 port, [scrya-com/mpeg-diffusion](https://github.com/scrya-com/mpeg-diffusion) (Apache-2.0). - **Weights**: [johndpope/MiniMax-H3-MPEG-Diffusion](https://huggingface.co/johndpope/MiniMax-H3-MPEG-Diffusion) — fitted up-maps + region head on the int8 ConvRot base. - **Base model**: MiniMax-H3, subject to the [MiniMax H3 Community License](https://huggingface.co/Comfy-Org/MiniMax-H3) — outputs are machine-generated and must be disclosed as such; the license does not permit use in the EU, UK, Republic of Korea or USA, and commercial use by companies above US$20M annual revenue needs written authorization from MiniMax. - **Example prompts**: portrait descriptions in the style of Pexels image alt-texts (Pexels content is free to use).