HZSDU/Exploration_Platform

Attention Distillation: A Unified Approach to Visual Characteristics Transfer

Project Page

🔥🔥 News

• 2025/03/08: We provide a new notebook with Style-specific T2I Generation with Flux.1-dev. See Issue 1 for more details.

• 2025/03/05: We add tiling to enable seamless textures generation. See Issue 3 for more details.

• 2025/03/01: We provide a simple HuggingFace🤗 demo. Check it out here !

Setup

This code was tested with Python 3.10, Pytorch 2.5 and Diffusers 0.32.

Examples

Texture Synthesis

• See [Texture Synthesis] part of ad notebook for generating texture images using SD1.5.

Style/Appearance Transfer

• See [Style/Appearance Transfer] part of ad notebook for style/appearance transfer using SD1.5.

Style-specific T2I Generation

• See [Style-specific T2I Generation] part of ad notebook for style-specific T2I generation using SD1.5 or SDXL.

VAE Finetuning

python train_vae.py \
    --image_path=/path/to/image \
    --vae_model_path=/path/to/vae

Web UI

Run the following command to start the Web UI:

python app.py

The Web UI will be available at http://localhost:7860.

ComfyUI

We also provide an implementation of Attention Distillation for ComfyUI. For more details, see here.