Top AI Repos — open-source AI, indexed and scored
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Top AI Repos tracks AI repositories on GitHub and answers two different questions about each one: is it moving right now, and would you bet a product on it.
Flash Diffusion — accelerating conditional diffusion models (AAAI 2025 Oral)
| Date | Stars |
|---|---|
| 2026-07-24 | 662 |
| 2026-07-25 | 662 |
| 2026-07-28 | 662 |
| 2026-07-30 | 662 |
| 2026-07-31 | 662 |
| 2026-08-06 | 662 |
Today
— stars today
This week
— stars this week
This month
— stars this month
Momentum
0.0
growth rate 0.00%/day
# ⚡ Flash Diffusion ⚡ (AAAI 2025 Oral)
This repository is the official implementation of the paper [Flash Diffusion: Accelerating Any Conditional Diffusion Model for Few Steps Image Generation](http://arxiv.org/abs/2406.02347).
<p align="center">
<a href="http://arxiv.org/abs/2406.02347">
<img src='https://img.shields.io/badge/Paper-2406.02347-green' />
</a>
<a href="https://gojasper.github.io/flash-diffusion-project/">
<img src='https://img.shields.io/badge/Project-page-blue' />
</a>
<a href='https://creativecommons.org/licenses/by-nd/4.0/legalcode'>
<img src="https://img.shields.io/badge/Licence-CC.BY.NC-purple" />
</a>
<br>
<a href="https://huggingface.co/spaces/jasperai/flash-diffusion">
<img src='https://img.shields.io/badge/%F0%9F%A4%97%20Demo-FlashPixart-orange' />
</a>
<a href="https://huggingface.co/spaces/jasperai/flash-sd3">
<img src='https://img.shields.io/badge/%F0%9F%A4%97%20Demo-FlashSD3-orange' />
</a>
<a href="https://huggingface.co/spaces/jasperai/flash-lora">
<img src='https://img.shields.io/badge/%F0%9F%A4%97%20Demo-FlashLoRAs-orange' />
</a>
<br>
<a href="https://huggingface.co/jasperai/flash-sd">
<img src='https://img.shields.io/badge/%F0%9F%A4%97%20Ckpt-FlashSD-yellow' />
</a>
<a href="https://huggingface.co/jasperai/flash-sdxl">
<img src='https://img.shields.io/badge/%F0%9F%A4%97%20Ckpt-FlashSDXL-yellow' />
</a>
<a href="https://huggingface.co/jasperai/flash-pixart">
<img src='https://img.shields.io/badge/%F0%9F%A4%97%20Ckpt-FlashPixart-yellow' />
</a>
<a href="https://huggingface.co/jasperai/flash-sd3">
<img src='https://img.shields.io/badge/%F0%9F%A4%97%20Ckpt-FlashSD3-yellow' />
</a>
</a>
<a href="https://huggingface.co/jasperai/flash-sdxl/tree/main/comfy">
<img src='https://img.shields.io/badge/Comfy-FlashSDXL-black' />
</a>
</p>
<figure>
<p align="center">
<img style="width:600px;" src="assets/flash_grid.jpg">
<figcaption>
<p align="center">
<b>Images generated using 4 NFEs</b>
</p>
</figcaption>
</p>
</figure>
In this paper, we propose an efficient, fast, versatile and LoRA-compatible distillation method to accelerate the generation of pre-trained diffusion models: *Flash Diffusion*. The method reaches state-of-the-art performances in terms of FID and CLIP-Score for few steps image generation on the COCO 2014 and COCO 2017 datasets, while requiring only **several GPU hours of training** and fewer trainable parameters than existing methods. In addition to its efficiency, the versatility of the method is also exposed across several tasks such as text-to-image, inpainting, face-swapping, super-resolution and using different diffusion models backbones either using a UNet-based denoisers (SD1.5, SDXL) or DiT (Pixart-α), as well as adapters. In all cases, the method allowed to reduce drastically the number of sampling steps while maintaining very high-quality image generation.
## Quick access
- [Method overview](#method)
- [Results overview](#results)
- [Installation 🛠️](#setup)
- [Text2Image model distillation](#distilling-existing-t2i-models)
- [Distilling a custom model 🚀](#example-of-a-distillation-training-with-a-custom-conditional-diffusion-model)
- [Inference with 🤗 Hugging Face pipelines](#inference-with-a-huggingface-pipeline-)
- [Using Flash with ComfyUI](#using-flash-in-comfyui)
- [Flash for training-free LoRAs acceleration 🎨](#combining-flash-diffusion-with-existing-loras-)
- [Citing this repository](#citation)
## Method
Our method aims to create a fast, reliable, and adaptable approach for various uses. We propose to train a student model to predict in a single step a denoised multiple-step teacher prediction of a corrupted input sample. Additionally, we sample timesteps from an adaptable distribution that shifts during training to help the student model target specific timesteps.
<p align="center">
<img style="width:600Excerpt of 14,745 characters
Read on GitHub
10
2
Would you bet a product on this? Bounded 0–100 and slow moving.
matched fp:416de463d556a91b, topic:diffusion-models, topic:text-to-image, topic:inpainting