Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]
StyO: Stylize Your Face in Only One-Shot
Authors: Bonan Li, Zicheng Zhang, Xuecheng Nie, Congying Han, Yinhan Hu, Xinmin Qiu, Tiande Guo
AAAI 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The extensive evaluation shows that Sty O produces high-quality images on numerous paintings of various styles and outperforms the current state-of-the-art. Extensive experiments show that the proposed Sty O model can produce surprisingly good quality of facial images on various styles in the one-shot manner. It also outperforms state-of-the-art models quantitatively. We conducted three user studies on the results in terms of the identity, geometric and texture. We receive 150 answers on 50 source-target pairs in total for each study. As shown in Table 1, over 0.38% of our results are selected as the best in both three metrics, which proves a significant advantage in stylization. Ablation Study Effect of Contrastive Disentangled Prompt Template. |
| Researcher Affiliation | Collaboration | 1 University of Chinese Academy of Sciences 2 MT Lab, Meitu Inc. EMAIL, EMAIL, EMAIL |
| Pseudocode | No | The paper describes methodologies using mathematical equations and textual explanations of steps (e.g., 'Inference with FCC. On this basis, the process of generation with FCC can be summarized as follows...'), but it does not include any clearly labeled 'Pseudocode' or 'Algorithm' block or structured code-like procedures. |
| Open Source Code | No | Baselines. We compare Sty O with two recent representative state-of-the-arts of GOGA (Zhang et al. 2022b), Custom Diffusion (Kumari et al. 2023) and Style ID (Chung, Hyun, and Heo 2024) to show the advantage of Sty O. All these methods are implemented by their official codes. We refrain from comparing with Instant ID (Wang et al. 2024) due to its lack of capability in extracting style from target images. There is no explicit statement about Sty O's code being open-source or provided, nor any link to a repository for the described methodology. |
| Open Datasets | Yes | In addition to the task-relevant data xsrc and xtgt, we introduce a no-cost auxiliary image set xaux to further enhance the disentanglement, in which about 200 natural faces are randomly sampled from FFHQ (Karras, Laine, and Aila 2019) and labeled with a same prompt a drawing with [Ssrc][not Stgt] style of portrait . The artistic portraits are token from AAHQ dataset (Liu et al. 2021b), while the natural faces employ photos of celebrities. |
| Dataset Splits | No | In addition to the task-relevant data xsrc and xtgt, we introduce a no-cost auxiliary image set xaux to further enhance the disentanglement, in which about 200 natural faces are randomly sampled from FFHQ (Karras, Laine, and Aila 2019) and labeled with a same prompt a drawing with [Ssrc][not Stgt] style of portrait . We receive 150 answers on 50 source-target pairs in total for each study. The paper specifies quantities for an auxiliary dataset and user study pairs, but it does not provide explicit training, validation, or test dataset splits (e.g., percentages, counts, or references to standard splits). |
| Hardware Specification | No | The paper does not provide any specific details about the hardware used to run the experiments, such as GPU models, CPU types, or memory specifications. |
| Software Dependencies | No | The paper does not list specific software dependencies with version numbers (e.g., programming languages, libraries, or solvers with their versions) that would be needed to replicate the experiment. |
| Experiment Setup | Yes | It is important to note that two hyper-parameters, ns and nc, are introduced to control the degree of stylization and identity preservation. Compared with the single identifier as used during training, a balance of ns and nc will result in better effects. Actually, it is not complicated to determine the optimal values for the two hyper-parameters. Through experiments, we find setting ns and nc to small values ( 3) is adequate to produce satisfactory results. |