|
--- |
|
license: cc-by-nc-4.0 |
|
library_name: diffusers |
|
base_model: PixArt-alpha/PixArt-XL-2-1024-MS |
|
tags: |
|
- lora |
|
- text-to-image |
|
inference: False |
|
--- |
|
# ⚡ FlashDiffusion: FlashPixart ⚡ |
|
|
|
|
|
Flash Diffusion is a diffusion distillation method proposed in [ADD ARXIV]() *by Clément Chadebec, Onur Tasar, Eyal Benaroche, and Benjamin Aubin.* |
|
This model is a **66.5M** LoRA distilled version of Pixart-α model that is able to generate 1024x1024 images in **4 steps**. See our [live demo](https://huggingface.co/spaces/jasperai/FlashPixart). |
|
|
|
|
|
<p align="center"> |
|
<img style="width:700px;" src="assets/hf_grid_pixart.jpg"> |
|
</p> |
|
|
|
# How to use? |
|
|
|
The model can be used using the `StableDiffusionPipeline` from `diffusers` library directly. It can allow reducing the number of required sampling steps to **2-4 steps**. |
|
|
|
```python |
|
import torch |
|
from diffusers import PixArtAlphaPipeline, Transformer2DModel, LCMScheduler |
|
from peft import PeftModel |
|
|
|
# Load LoRA |
|
transformer = Transformer2DModel.from_pretrained( |
|
"PixArt-alpha/PixArt-XL-2-1024-MS", |
|
subfolder="transformer", |
|
torch_dtype=torch.float16 |
|
) |
|
transformer = PeftModel.from_pretrained( |
|
transformer, |
|
"jasperai/flash-pixart" |
|
) |
|
|
|
# Pipeline |
|
pipe = PixArtAlphaPipeline.from_pretrained( |
|
"PixArt-alpha/PixArt-XL-2-1024-MS", |
|
transformer=transformer, |
|
torch_dtype=torch.float16 |
|
) |
|
|
|
# Scheduler |
|
pipe.scheduler = LCMScheduler.from_pretrained( |
|
"PixArt-alpha/PixArt-XL-2-1024-MS", |
|
subfolder="scheduler", |
|
timestep_spacing="trailing", |
|
) |
|
|
|
pipe.to("cuda") |
|
|
|
prompt = "A raccoon reading a book in a lush forest." |
|
|
|
image = pipe(prompt, num_inference_steps=4, guidance_scale=0).images[0] |
|
``` |
|
<p align="center"> |
|
<img style="width:400px;" src="assets/raccoon.png"> |
|
</p> |
|
|
|
# Training Details |
|
The model was trained for 40k iterations on 4 H100 GPUs (representing approximately 188 hours of training). Please refer to the [paper]() for further parameters details. |
|
|
|
**Metrics on COCO 2014 validation (Table 4)** |
|
- FID-10k: 29.30 (4 NFE) |
|
- CLIP Score: 0.303 (4 NFE) |
|
|
|
## License |
|
This model is released under the the Creative Commons BY-NC license. |
