---
license: mit
datasets:
- ILSVRC/imagenet-1k
tags:
- image-generation
- diffusion
- pytorch
---

# Persistence Forcing: Exploiting Feature Specialization in Pixel-Space Diffusion

**Paper:** [arXiv:2609.36014](https://arxiv.org/abs/2609.36014)  
**Project Page:** [chongwang1024.github.io/PerF](https://chongwang1024.github.io/PerF/)  
**Code Repository:** [github.com/ChongWang1024/PerF](https://github.com/ChongWang1024/PerF)

PerF exploits the persistent–active feature organization that emerges under heterogeneous refinement. Persistent-to-Active Conditioning lets persistent features guide active refinement, while Persistence Guidance strengthens their contribution during sampling.

## ImageNet Samples

![ImageNet samples generated by PerF](https://raw.githubusercontent.com/ChongWang1024/PerF/main/demo/visual.png)

## Checkpoints

Official pretrained checkpoints for ImageNet generation:

| Resolution | Model | Checkpoint |
|---|---|---|
| 256x256 | PerF-B/16 | [checkpoint-last.pth](./perf-b16-256/checkpoint-last.pth) |
| 256x256 | PerF-L/16 | [checkpoint-last.pth](./perf-l16-256/checkpoint-last.pth) |
| 256x256 | PerF-H/16 | [checkpoint-last.pth](./perf-h16-256/checkpoint-last.pth) |
| 512x512 | PerF-H/32 | [checkpoint-last.pth](./perf-h32-512/checkpoint-last.pth) |

## Usage

Download a checkpoint with:

```bash
hf download ChongWang1024/PerF \
  perf-b16-256/checkpoint-last.pth \
  --local-dir .
```

For evaluation commands and training instructions, see the [code repository](https://github.com/ChongWang1024/PerF).