Würstchen is a new framework for training text-conditional models by moving the computationally expensive text-conditional stage into a highly compressed latent space. Common approaches make use of a single stage compression, while Würstchen introduces another Stage that introduces even more compression. In total we have Stage A & B that are responsible for compressing images and Stage C that learns the text-conditional part in the low dimensional latent space. With that Würstchen achieves a 42x compression factor, while still reconstructing images faithfully. This enables training of Stage C to be fast and computationally cheap. We refer to the paper for details.
You can use the model simply through the notebooks here. The Stage B notebook only for reconstruction and the Stage C notebook is for the text-conditional generation. You can also try the text-to-image generation on Google Colab.
Würstchen is fully integrated into the diffusers library. Here's how to use it:
# pip install -U transformers accelerate diffusers
import torch
from diffusers import AutoPipelineForText2Image
from diffusers.pipelines.wuerstchen import DEFAULT_STAGE_C_TIMESTEPS
pipe = AutoPipelineForText2Image.from_pretrained("warp-ai/wuerstchen", torch_dtype=torch.float16).to("cuda")
caption = "Anthropomorphic cat dressed as a fire fighter"
images = pipe(
caption,
width=1024,
height=1536,
prior_timesteps=DEFAULT_STAGE_C_TIMESTEPS,
prior_guidance_scale=4.0,
num_images_per_prompt=2,
).imagesRefer to the official documentation to learn more.
Training Würstchen is considerably faster and cheaper than other text-to-image as it trains in a much smaller latent space of 12x12. We provide training scripts for both Stage B and Stage C.
| Model | Download | Parameters | Conditioning | Training Steps | Resolution |
|---|---|---|---|---|---|
| Würstchen v1 | Hugging Face | 1B (Stage C) + 600M (Stage B) + 19M (Stage A) | CLIP-H-Text | 800.000 | 512x512 |
| Würstchen v2 | Hugging Face | 1B (Stage C) + 600M (Stage B) + 19M (Stage A) | CLIP-bigG-Text | 918.000 | 1024x1024 |
Special thanks to Stability AI for providing compute for our research.
If you use our approach in your research or were inspired by it, we would be thrilled if you cite our paper:
@inproceedings{
pernias2024wrstchen,
title={W\"urstchen: An Efficient Architecture for Large-Scale Text-to-Image Diffusion Models},
author={Pablo Pernias and Dominic Rampas and Mats Leon Richter and Christopher Pal and Marc Aubreville},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=gU58d5QeGv}
}