Skip to content

Latest commit

 

History

History
117 lines (80 loc) · 6.17 KB

README.md

File metadata and controls

117 lines (80 loc) · 6.17 KB
Raw

CFG++ : MANIFOLD-CONSTRAINED CLASSIFIER FREE GUIDANCE FOR DIFFUSION MODELS

This repository is the official implementation of CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models, led by

Hyungjin Chung*, Jeongsol Kim*, Geon Yeong Park*, Hyelin Nam*, Jong Chul Ye

main figure

Project Website arXiv

🔥 Summary

Classifier-free guidance (CFG) is a fundamental tool in modern diffusion models for text-guided generation. Although effective, CFG requires high guidance scales, which has notable drawbacks:

  1. Mode collapse and saturation
  2. Poor invertibility
  3. Unnatural, curved PF-ODE trajectory

We propose a simple fix to this seemingly inherent limitation and propose CFG++ 🚀, which corrects the off-manifold problem of CFG. The following advantages are observed

  1. Small guidance scale $\lambda \in$ [0, 1] can be used with a similar effect as $\omega \in$ [1.0, 12.5] in CFG
  2. Better sample quality and better adherence to text
  3. Smooth, straighter PF-ODE trajectory
  4. Enhanced invertibility

Experimental results confirm that our method significantly enhances performance in text-to-image generation, DDIM inversion, editing, and solving inverse problems, suggesting a wide-ranging impact and potential applications in various fields that utilize text guidance.

🗓 ️News

🛠️ Setup

First, create your environment. We recommend using the following comments.

git clone https://github.com/CFGpp-diffusion/CFGpp.git
cd CFGpp
conda env create -f environment.yaml

For reproducibility, using the same package version is necessary since some dependencies lead to significant differences (for instance, diffusers). Nonetheless, improvement induced by CFG++ will be observed regardless the dependency.

Diffusers will automatically download checkpoints for SDv1.5 or SDXL. For the fast sampling, we also support SDXL-lightning. See T2I examples below.

🌄 Examples

Text-to-Image generation

  • CFG
python -m examples.text_to_img --prompt "a portrait of a dog" --method "ddim" --cfg_guidance 7.5
  • CFG++ We support DDIM CFG++ (ddim_cfg++) and DPM++ 2M CFG++ (dpm++_2m_cfgpp) at this moment. Please refer to Auto1111 reForge and ComfyUI for the other samplers, e.g. Euler-a CFG++, DPM++ SDE CFG++, etc.
python -m examples.text_to_img --prompt "a portrait of a dog" --method "ddim_cfg++" --cfg_guidance 0.6
  • CFG++ (SDXL-lightning)

First, download sdxl_lightning_4step_unet.safetensors in ckpt. Then run the test below.

python -m examples.text_to_img --prompt "stars, water, brilliantly, gorgeous large scale scene, a little girl, in the style of dreamy realism, light gold and amber, blue and pink, brilliantly illuminated in the background." --method "ddim_cfg++_lightning" --model "sdxl_lightning" --cfg_guidance 1 --NFE 4

  • You can test other NFEs with different safetensors (e.g., 8step_unet). Make sure to modify the --NFE accordingly.

  • SDXL-lightning already distilled the score with pre-fixed CFG scale. Therefore, we set --cfg_guidance 1. That said, the key difference lies in the renoising step.

  • For the Lightning with original DDIM, run above with --method "ddim_lightning"

Image Inversion

  • CFG
python -m examples.inversion --prompt "a photography of baby fox" --method "ddim_inversion" --cfg_guidance 7.5
  • CFG++
python -m examples.inversion --prompt "a photography of baby fox" --method "ddim_inversion_cfg++" --cfg_guidance 0.6

Tip

If you want to use SDXL, add --model sdxl.

🔬 Callback

We provide callback functionality to monitor intermediate samples during the diffusion reverse process. For now, the function could be called only at the end of each timestep, for the readability of scripts.

Currently, we provide two options (default: None).

  • draw_tweedie : save $\hat x_{0|t}$ to workdir
  • draw_noisy : save $x_t$ to workdir

Note that using callback may take more time due to file save. You can refer utils/callback_util.py for details.

📝 Citation

If you find our method useful, please cite as below or leave a star to this repository.

@article{chung2024cfg++,
  title={CFG++: Manifold-constrained Classifier Free Guidance for Diffusion Models},
  author={Chung, Hyungjin and Kim, Jeongsol and Park, Geon Yeong and Nam, Hyelin and Ye, Jong Chul},
  journal={arXiv preprint arXiv:2406.08070},
  year={2024}
}

Note

This work is currently in the preprint stage, and there may be some changes to the code.