Research map / Image generation / Editing and style transfer
CSGO: Content-Style Composition in Text-to-Image Generation
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.
Builds IMAGStyle, a 210K content-style-stylized triplet dataset via a LoRA-decomposition generation pipeline with automatic CAS-based cleaning, and trains CSGO, an end-to-end SDXL-based style transfer model with independent content and style injection.
Problem
Image style transfer lacks large-scale content-style-stylized triplet data, so existing methods rely on training-free inversion or per-image LoRA fine-tuning, which limit controllability, generalization, and speed.
Approach
Generate stylized triplets by combining content LoRAs and style LoRAs separated implicitly (B-LoRA), filter them with an automatic content alignment score (CAS) computed on DINOv2 features after AdaIN-style normalization, and train an end-to-end SDXL model with decoupled content injection (Tile ControlNet plus a CLIP-based cross-attention) and style injection (Resampler-projected tokens in up-bloc
Key results
- CSGO achieves the highest CSD style similarity of 0.5146 versus 0.4522 for StyleShot and 0.4274 for Style-Aligned.
- CSGO attains CAS of 0.8386 (lower is better), beating StyTR^2 (0.9699) and StyleShot-lineart (1.0750), though StyleID reaches a lower CAS of 0.4873.
- Ablations show Resampler style projection captures more detailed style features than a linear layer while avoiding content leakage.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CSGO instead of StyleShot | Image-driven style transfer · IMAGStyle test set | CSD: 0.4522 -> 0.5146 | Helped |
| Use CSGO instead of StyleID | Image-driven style transfer · IMAGStyle test set | CAS: 0.4873 -> 0.8386 | Hurt |
| Add ControlNet content injection to cross-attention-only content control (ablation) | CSGO · Image-driven style transfer | content retention | Helped |
| Add cross-attention content injection to ControlNet content control (ablation) | CSGO · Image-driven style transfer | content semantics | Helped |
| Pre-inject style features in the ControlNet branch rather than injecting style only into the base UNet (ablation) | CSGO · Image-driven style transfer | style control | Helped |
| Use a Resampler instead of a linear style projection layer (ablation) | CSGO · Image-driven style transfer | style feature detail | Helped |
| Increase the style projection token number (ablation) | CSGO · Image-driven style transfer and text-driven stylized synthesis | style control | Helped |
| Increase content injection scale δc (ablation) | CSGO · Image-driven style transfer | content retention | Helped |
| Increase content injection scale δc to large values (ablation) | CSGO · Image-driven style transfer | style control: 0.9 and 1.2 | Hurt |
| Introduce classifier-free guidance (ablation) | CSGO · Image-driven style transfer | style transfer effect | Helped |
| Increase style scale λs to 1.0 or above (ablation) | CSGO · Image-driven style transfer | style control: 1.0 | Helped |
| Set content scale λc near 1.0 (ablation) | CSGO · Image-driven style transfer | content retention: 1.0 | Helped |
Surprising findings
- StyleID achieves much lower content-alignment CAS (0.4873) than CSGO but a very low CSD style score (0.0992), indicating over-preserved content and weak style transfer.
Limitations
- Dataset limited to 210K triplets due to time and compute constraints.
- CSGO is described as a basic version; style and content feature fusion could be improved.
- Implicit LoRA-based decomposition is unstable, so generated triplets do not always match the content image, requiring filtering.
Experimental setup
Datasets: IMAGStyle MSRA10K MSRA-B ImageNet-Sketch WikiArt
Benchmarks: IMAGStyle test set (248 content images x 206 style images)
Metrics: CSD score CAS (content alignment score)
Baselines: StyTR^2 Style-Aligned StyleID InstantStyle StyleShot StyleShot-lineart DEADiff
Builds on: B-LoRA ControlNet IP-Adapter Stable Diffusion XL AdaIN
Closest papers by meaning
- SigStyle: Signature Style Transfer via Personalized Text-to-Image Models (AAAI 2025 · similarity 0.89)
- OmniStyle: Filtering High Quality Style Transfer Data at Scale (CVPR 2025 · similarity 0.88)
- StyleStudio: Text-Driven Style Transfer with Selective Control of Style Elements (CVPR 2025 · similarity 0.88)
- StyleSSP: Sampling StartPoint Enhancement for Training-free Diffusion-based Method for Style Transfer (CVPR 2025 · similarity 0.88)
- Style Injection in Diffusion: A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer (CVPR 2024 · similarity 0.87)
- CLeaR: A Unified Framework for Resolving the Leakage–Degradation Dilemma in Style Transfer (NeurIPS 2026 · similarity 0.87)
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.