Research map / Image generation / Editing and style transfer
StyO: Stylize Your Face in Only One-Shot
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes StyO, a one-shot face stylization method that fine-tunes a latent diffusion model to disentangle style and content into identifier tokens, then recombines them with cross-attention control to produce stylized faces.
Problem
One-shot face stylization from a single artistic target: existing GAN-based methods struggle to preserve source content (colors, details) while achieving geometric variation of the target style.
Approach
Disentangle style and content of source and target images into contrastive positive/negative text identifiers (rare firearm-word tokens) via triple-reconstruction fine-tuning of an LDM, then recombine identifiers across images in a prompt and transfer source cross-attention maps for content control, plus repeated-identifier prompt augmentation.
Key results
- In a user study with 150 answers over 50 source-target pairs, StyO was preferred for identity (0.42), geometry (0.47) and texture (0.38), beating StyleID (0.38/0.12/0.28), GOGA and Custom Diffusion.
- Ablations show positive-only identifiers overfit and produce haloing artifacts, while adding negative identifiers and the auxiliary set improves style-content disentanglement.
- Fine-grained content controller reduces diversity and improves identity fidelity relative to the recombined prompt alone.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use StyO instead of StyleID | One-shot face stylization · User study | Identity: 0.38 -> 0.42 | Helped |
| Use StyO instead of Custom Diffusion | One-shot face stylization · User study | Geometry: 0.27 -> 0.47 | Helped |
| Use StyO instead of StyleID | One-shot face stylization · User study | Texture: 0.28 -> 0.38 | Helped |
| Add negative identifiers to positive identifiers (ablation) | StyO's disentangled prompt template · One-shot face stylization | Content preservation | Helped |
| Add the auxiliary image set to positive-negative identifiers (ablation) | StyO's disentangled prompt template · One-shot face stylization | Style and content quality | Helped |
| Add FCC cross-attention control (ablation) | StyO · One-shot face stylization | Identity fidelity | Helped |
| Add FCC cross-attention control (ablation) | StyO · One-shot face stylization | Synthesis diversity | Hurt |
| Increase the number of style identifiers (ablation) | StyO's augmented prompt · One-shot face stylization | Style | Mixed |
| Increase the number of content identifiers (ablation) | StyO's augmented prompt · One-shot face stylization | Content faithfulness | Mixed |
Limitations
- Evaluation relies solely on a small user study (50 pairs), with no automatic metrics reported.
- Hyperparameters n_s and n_c must be tuned per example to balance style and content.
- Requires per-target fine-tuning of the diffusion model, so each new style needs its own training run.
Experimental setup
Datasets: FFHQ AAHQ celebrity photos
Metrics: user preference rate for identity user preference rate for geometry user preference rate for texture
Baselines: GOGA Custom Diffusion StyleID
Builds on: latent diffusion models DreamBooth prompt-to-prompt cross-attention control
Closest papers by meaning
- ACFun: Abstract-Concrete Fusion Facial Stylization (NeurIPS 2024 · similarity 0.88)
- Deformable One-shot Face Stylization via DINO Semantic Guidance (CVPR 2024 · similarity 0.87)
- Styl3R: Instant 3D Stylized Reconstruction for Arbitrary Scenes and Styles (NeurIPS 2025 · similarity 0.86)
- DEADiff: An Efficient Stylization Diffusion Model with Disentangled Representations (CVPR 2024 · similarity 0.85)
- One-Shot Structure-Aware Stylized Image Synthesis (CVPR 2024 · similarity 0.85)
- Morpheus: Text-Driven 3D Gaussian Splat Shape and Color Stylization (CVPR 2025 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.