Stable-Hair: Real-World Hair Transfer via Diffusion Model
Authors: Yuxuan Zhang, Qing Zhang, Yiren Song, Jichao Zhang, Hao Tang, Jiaming Liu
AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments demonstrate that our approach achieves state-of-the-art performance compared to existing hair transfer methods. ... Through extensive experiments, Stable-Hair has demonstrated its superior performance, significantly surpassing existing state-of-the-art hair transfer methods in terms of fidelity and fine-grained detail. ... Experiments demonstrate that our method significantly surpasses existing SOTA hair transfer methods in terms of fidelity and fine-grained detail. ... Qualitative Comparison. As illustrated in Fig. 4, we conducted qualitative comparison experiments on a variety of hairstyles. ... Quantitative Comparison. The experiment uses the Celeb A-HQ dataset (Karras et al. 2017) as experimental data... User Study. Considering the subjective nature of the hairstyle transfer task, we conducted a comprehensive user study involving 30 volunteers. ... Ablation Study. To thoroughly investigate the role of each module in our method, we conducted systematic ablation. |
| Researcher Affiliation | Collaboration | 1Shanghai Jiao Tong University 2Tiamat AI 3Shenyang Institute of Automation Chinese Academy of Sciences 4National University of Singapore 5Ocean University of China 6Peking University |
| Pseudocode | No | The paper describes the methodology using textual explanations and diagrams (e.g., Fig. 2), but does not include any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | No | The paper does not contain any explicit statements about releasing source code or provide links to a code repository. |
| Open Datasets | Yes | In the first stage, we trained the Bald Converter using the Non-Hair FFHQ dataset (Wu, Yang, and Jin 2022). ... The experiment uses the Celeb A-HQ dataset (Karras et al. 2017) as experimental data, 2500 face images are randomly selected as input from the Celeb A-HQ dataset with an equal number of reference images from the remaining Celeb-HQ dataset. |
| Dataset Splits | No | In terms of image data, we used every single image as original images to generate reference images and bald proxy images, resulting in a total of 60,000 images. For video data, we sampled two frames from the same video. One frame was used to generate bald proxy images, and another frame was used to generate reference images. Resulting in a total of 90,000 images. ... The experiment uses the Celeb A-HQ dataset (Karras et al. 2017) as experimental data, 2500 face images are randomly selected as input from the Celeb A-HQ dataset with an equal number of reference images from the remaining Celeb-HQ dataset. |
| Hardware Specification | Yes | This training was conducted on a single H800 GPU with a batch size of 16 and a learning rate of 5e-5, over a total of 8,000 steps. In the second stage, we trained our Hair Extractor and Latent Identity Net using the prepared triplet data. This stage was performed on 8 H800 GPUs with a batch size of 8 and a learning rate of 5e-5, over a total of 100,000 steps. |
| Software Dependencies | No | We employed Stable Diffusion V1-5 as the pre-trained diffusion model. ... The pipeline uses Chat GPT to generate text prompts, the Stable Diffusion Inpainting model for creating reference images... |
| Experiment Setup | Yes | This training was conducted on a single H800 GPU with a batch size of 16 and a learning rate of 5e-5, over a total of 8,000 steps. In the second stage, we trained our Hair Extractor and Latent Identity Net using the prepared triplet data. This stage was performed on 8 H800 GPUs with a batch size of 8 and a learning rate of 5e-5, over a total of 100,000 steps. During inference, we followed the same two-stage approach as used in training. Both stages utilized the DDIM sampler with 30 sampling steps and the classifier-free guidance scale setting of 1.5. |