File size: 2,760 Bytes
8fd1e59
6febba8
 
 
 
8fd1e59
 
 
6febba8
 
 
 
 
8fd1e59
 
6febba8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e835f7f
 
6febba8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
---
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.) β€” 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** β€” 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* β€” 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* β€” 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) β€” 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) β€” 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).