Research map / Image generation / Editing and style transfer
CLeaR: A Unified Framework for Resolving the Leakage–Degradation Dilemma in Style Transfer
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 CLeaR, a training-free style transfer framework that uses orthogonal subspace projection, ensemble inversion over multiple VFMs, and energy-guided calibration to reduce content leakage while preserving style.
Problem
Style transfer with a style reference and content prompt suffers from content leakage (reference objects/layouts appearing in output), and methods that suppress leakage often degrade style fidelity, a leakage-degradation dilemma.
Approach
Define content-reduced style targets by orthogonally projecting reference features away from content-proxy features in each VFM space, then optimize a shared pixel-space style anchor matching all VFM targets, and calibrate diffusion sampling with a VFM-ensemble style-alignment energy gradient.
Key results
- On StyleBench, CLeaR reaches CSD 0.682 vs 0.630 for the best baseline (Attention Distillation), and style loss 0.034 vs 0.097 (StyleShot).
- DINO content leakage 0.066 (lower is better) vs 0.168 for the second best (CSGO); KID 0.358 vs 0.237.
- AI scoring 4.16 vs 3.85 for StyleAligned; CLIP-Aesthetic 6.445 is below the best baseline RB-Modulation at 6.677.
Surprising findings
- Aesthetic score is slightly lower than several baselines, which the authors attribute to the inversion process prioritizing disentanglement over aesthetic smoothing.
- Attention Distillation attains the second-best CSD but has the worst AI-scored leakage (1.84).
Limitations
- Added test-time computation from Ensemble Inversion and energy-guided calibration makes it slower than feed-forward methods.
- Incomplete or ambiguous content descriptions, degenerate style anchors, and adapter misinterpretation of inversion noise cause leakage or artifacts.
- Evaluation uses VFMs and Qwen3 overlapping with the optimization/description pipeline, partly addressed by held-out checks.
Experimental setup
Datasets: StyleBench
Benchmarks: StyleBench
Metrics: CSD style loss CLIP DINO KID AI scoring CLIP-Aesthetic SigLIP Nomic
Baselines: Attention Distillation CSGO DEADiff InstantStyle MaskST RB-Modulation StyleAligned StyleShot
Builds on: IP-Adapter DDIM StyleBench
Closest papers by meaning
- StyleSSP: Sampling StartPoint Enhancement for Training-free Diffusion-based Method for Style Transfer (CVPR 2025 · similarity 0.91)
- 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 0.89)
- Less is More: Masking Elements in Image Condition Features Avoids Content Leakages in Style Transfer Diffusion Models (ICLR 2025 · similarity 0.88)
- Conditional Balance: Improving Multi-Conditioning Trade-Offs in Image Generation (CVPR 2025 · similarity 0.88)
- Semantix: An Energy-guided Sampler for Semantic Style Transfer (ICLR 2025 · similarity 0.88)
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.