AI images of yourself in LESS THAN 10 seconds Similar to InstantID (but it’s FLUX) PuLID: Pure and Lightning ID Customization via Contrastive Alignment

Introduction

PuLID is an innovative tuning-free ID customization method designed for text-to-image generation. By incorporating a Lightning T2I branch alongside 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 demonstrate that PuLID achieves superior performance in both ID fidelity and editability. Additionally, a notable feature of PuLID is that the image elements (e.g., background, lighting, composition, and style) remain as consistent as possible before and after ID insertion.

Try PuLID para FLUX Free

Whats happenning on the twitter about Pulid Flux?

Methods

  • Contrastive Alignment: Through contrastive alignment loss and ID loss, PuLID inserts ID information without affecting the original model’s behavior.
  • Lightning T2I Branch: Introduces a Lightning T2I branch that uses fast sampling techniques to generate high-quality images from pure noise.
  • Optimizing ID Loss: Optimizes ID loss in a more accurate setting to enhance ID similarity.

Experiments

  • Quantitative Comparison: Evaluates ID fidelity using ID cosine similarity, showing that PuLID outperforms existing methods across all test sets and base models.
  • Qualitative Comparison: PuLID achieves high ID similarity while causing less disruption to the original model, accurately reproducing the original model’s lighting, style, and layout.

Contributions

  1. Proposes a tuning-free method, PuLID, which preserves high ID similarity while mitigating the impact on the original model’s behavior.
  2. Introduces a Lightning T2I branch alongside the regular diffusion branch, incorporating contrastive alignment loss and ID loss to minimize the contamination of ID information on the original model while ensuring fidelity.
  3. Experiments show that PuLID achieves state-of-the-art performance in terms of both ID fidelity and editability and is less invasive to the model, making it more flexible for practical applications.

FAQs