PuLID: Pure and Lightning ID Customization via Contrastive Alignment
About
We propose Pure and Lightning ID customization (PuLID), a novel tuning-free ID customization method for text-to-image generation. By incorporating a Lightning T2I branch with a standard diffusion one, PuLID introduces both contrastive alignment loss and accurate ID loss, minimizing disruption to the original model and ensuring high ID fidelity. Experiments show that PuLID achieves superior performance in both ID fidelity and editability. Another attractive property of PuLID is that the image elements (e.g., background, lighting, composition, and style) before and after the ID insertion are kept as consistent as possible. Codes and models are available at https://github.com/ToTheBeginning/PuLID
Zinan Guo, Yanze Wu, Zhuowei Chen, Lang Chen, Peng Zhang, Qian He• 2024
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Safe Generation Rate | I2P | GPT-4o Score82.47 | 9 | |
| Prompt-image Alignment | Sneakyprompt | CLIPScore0.7211 | 8 | |
| Prompt-image Alignment | I2P | CLIPScore0.8195 | 8 | |
| Prompt-image Alignment | MMA-Diffusion | CLIPScore0.7228 | 8 | |
| Prompt-image Alignment | Misbinding | CLIPScore0.8722 | 8 | |
| Safe Generation Rate | Sneakyprompt | GPT-4o0.7868 | 8 | |
| Safe Generation Rate | MMA-Diffusion | GPT-4o0.7146 | 8 | |
| Safe Generation Rate | Misbinding | GPT-4o Score0.6466 | 8 | |
| Identity-Preserving Text-to-Image Generation | IBench 41 prompts 100 IDs | Aesthetic Score68.3 | 7 | |
| Identity Customization | IBench ChineseID editable long prompts | Aesthetic Score0.683 | 6 |
Showing 10 of 19 rows