---
title: PixelMoE
emoji: 🧩
colorFrom: green
colorTo: red
sdk: gradio
sdk_version: 6.31.0
app_file: app.py
short_description: Class-conditional 512x512 pixel-space MoE diffusion
python_version: "3.12"
startup_duration_timeout: 30m
---

# PixelMoE

Interactive demo for **PixelMoE: Scaling Pixel-Space Diffusion with LLM-Style
Mixture-of-Experts** ([arXiv 2610.18857](https://arxiv.org/abs/2610.18857)).

Pick any of the 1000 ImageNet-1K classes and the Space samples a 512×512 image
with the released **PixelMoE-B-E32A4/16** base model from
[KlingTeam/PixelMoE](https://huggingface.co/KlingTeam/PixelMoE), using the
authors' official sampling settings (Heun solver, 25 steps per hierarchical
stage, CFG 3.7 applied while t < 0.8, second-EMA weights).

PixelMoE runs diffusion entirely in pixel space: every image token activates
4 of 32 routed experts plus a shared expert in each MoE block, and images are
generated by a two-stage hierarchical flow — coarse structure first, then
details at the target resolution.

## Advanced settings

- **Steps per stage** — denoising steps per hierarchical stage (25 in the
  official `scripts/sample.sh`; each Heun step is two model evaluations).
- **CFG scale / CFG interval max** — classifier-free guidance and the global
  timestep below which it is active.
- **Sigma min** — lower bound clamped into the velocity denominator.

The model code is vendored from the MIT-licensed official repository
[KlingAIResearch/PixelMoE](https://github.com/KlingAIResearch/PixelMoE)
(`torch.compile` decorators and the deterministic-scatter toggle are adapted
for the ZeroGPU runtime, otherwise inference is unchanged).