xslim

KO | EN

gitlite — search

#javascript #python #hacktoberfest #react #ai #typescript #llm #go #golang #android #machine-learning #rust #deep-learning #linux

★ 12 Open GitHub ↗

Official implementation of X-Slim(xslim): Accelerating diffusion model via Extreme-slimming Caching

Download

Download Latest Version (.zip)

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

<h1>X-Slim: No Cache Left Idle</h1>
<h3>Accelerating Diffusion Models via <i>Extreme-Slimming Caching</i></h3>

<div align="center">
<strong>🚀 X-Slim is a <u>training-free</u> cache-based accelerator that jointly exploits redundancy across <u>temporal</u> (timesteps), <u>structural</u> (blocks), and <u>spatial</u> (tokens) dimensions.</strong>

<details>
<summary><b>📖 Abstract</b> (click to expand)</summary>

Diffusion models deliver strong generative quality, but inference cost scales with timestep count, model depth, and token length. Feature caching reuses nearby computations, yet aggressive timestep skipping often hurts fidelity while conservative block or token refresh yields limited speedup. We present <b>X-Slim</b> (e<b>X</b>treme-<b>Slim</b>ming Caching), a training-free, cache-based accelerator that jointly exploits redundancy across temporal, structural, and spatial dimensions.

X-Slim introduces a dual-threshold <b>push-then-polish</b> controller: it first pushes timestep-level reuse up to an early-warning line, then polishes residual error with lightweight block- and token-level refresh; a critical line triggers full inference to reset error. Level-specific, context-aware indicators guide when and where to cache, shrinking search overhead.

On FLUX.1-dev and HunyuanVideo, X-Slim reduces latency by up to 4.97× and 3.52× with minimal perceptual loss, and on DiT-XL/2 it reaches 3.13× acceleration with a FID improvement of 2.42 over prior methods.

</details>

</div>

---

## ✨ Highlights

- 🔥 **Push-then-Polish** caching with a **dual-threshold controller** (early-warning + critical reset).
- ⚡ **Up to 4.97×** latency reduction on **FLUX.1-dev** and **3.52×** on **HunyuanVideo** (minimal perceptual loss).
- 🏆 **3.13×** acceleration on **DiT-XL/2**, with **FID improved by 2.42** over prior methods.
- 🧩 **Level-specific, context-aware indicators** guide *when* and *where* to reuse vs refresh.

---

## 🎞️ Video Demo

https://github.com/user-attachments/assets/743e743b-16db-4076-85dc-5a7065ad26ab

---

## 🧭 Todo
- [ ] 🔌Polish the plug-and-play cache manager interface
- [x] 📹 Release X-Slim for HunyuanVideo.
- [x] 🖼️ Release X-Slim for FLUX.
- [x] 🔗 Release the project page
- [x] 📝 Release the arXiv paper

Our goal is to make X-Slim **truly plug-and-play**: readers can drop in the manager interface, record their own statistics, and design custom schedules.

---

## 🧠 Method at a Glance

### 1) Push-then-Polish Caching
X-Slim **pushes** step-level reuse until an **early-warning line**, then **polishes** residual error by selectively refreshing **blocks/tokens**. When a **critical line** is triggered, X-Slim performs a **full inference step** to reset accumulated error.

### 2) Level-specific Strategy
Different reuse levels follow different dynamics:
- **Step-level**: adjacent timesteps show a *U-shaped* change pattern (weakly prompt-dependent).
- **Block-level**: sensitivity varies with depth, but exhibits *consistent depth-wise patterns*.
- **Token-level**: largely *content-dependent*; refresh focuses on high-change regions.

---
## 🔖 Citation

If you find X-Slim helpful, please consider giving a star ⭐ and citing 📝

```bibtex
@article{xslimcache2025,
  title={No Cache Left Idle: Accelerating Diffusion Model via Extreme-Slimming Caching},
  author={Wen, Tingyan and Li, Haoyu and Chen, Yihuang and Zhou, Xing and Zhu, Lifei and Wang, XueQian},
  journal={arXiv preprint arXiv:2512.12604},
  year={2025}
}
```

---

## 🛡️ Disclaimer

This is the official code release of **X-Slim**.  
Demo images/videos are from community users; please contact us if you would like them removed.

---

## 💞 Acknowledgements

We thank the open-source community and upstream projects that made this work possible, including (but not limited to):
- FLUX
- HunyuanVideo
- DiT backbones
- Caching-acceleration works (TeaCache, TaylorSeer,etc.)

xslim

Download

Today's picks

Was this content helpful?

Recent Feedback

xslim

Download

Today's picks

Was this content helpful?

Recent Feedback

Download README