PuLID: Pure and Lightning ID Customization via Contrastive Alignment

Authors: Zinan Guo, Yanze Wu, Chen Zhuowei, Lang chen, Peng Zhang, Qian HE

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that Pu LID achieves superior performance in both ID fidelity and editability.
Researcher Affiliation Industry Zinan Guo Yanze Wu Zhuowei Chen Lang Chen Peng Zhang Qian He Byte Dance Inc. guozinan.1@bytedance.com wuyanze123@gmail.com
Pseudocode No The paper does not contain structured pseudocode or algorithm blocks.
Open Source Code Yes Codes and models are available at https://github.com/To The Beginning/Pu LID.
Open Datasets Yes As a supplementary resource, we also used a recent open-source test set, Unsplash-50 [8], which comprises 50 portrait images uploaded to the Unsplash website between February and March 2024.
Dataset Splits No The paper does not provide specific dataset split information (exact percentages, sample counts, citations to predefined splits, or detailed splitting methodology) for its main training dataset.
Hardware Specification Yes Training is performed with Py Torch and diffusers on 8 NVIDIA A100 GPUs in an internal cluster.
Software Dependencies No The paper mentions "Py Torch and diffusers" but does not provide specific version numbers for these or any other ancillary software components.
Experiment Setup Yes We set the λalign-sem to 0.6, λalign-layout to 0.1, and λid to 1.0. In the Lightning T2I training branch, we set the resolution of the generated image to 768 × 768 to conserve memory.