Research map / Image generation / Editing and style transfer
OmniStyle: Filtering High Quality Style Transfer Data at Scale
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.
Introduces OmniStyle-1M, a million-triplet paired style transfer dataset filtered by the OmniFilter quality framework, and OmniStyle, a FLUX-dev-based DiT model for instruction- and image-guided style transfer.
Problem
Style transfer lacks large-scale, high-quality paired (content, style, stylized) data, forcing slow optimization- or inversion-based methods and limiting generalization to arbitrary styles.
Approach
Generate a million stylized triplets by running six SOTA style transfer models on FLUX/ChatGPT-generated content images and Style30K style images, then filter them with a multi-dimensional MLLM/CLIP/DINOv2-based score (content preservation, style consistency, aesthetic appeal); fine-tune FLUX-dev DiT end-to-end for feed-forward stylization.
Key results
- On instruction-guided style transfer, OmniStyle reaches content preservation 0.5128 and style consistency 0.6441, beating OmniGen (0.4918, 0.5487) and DiffStyler (0.4816, 0.5127).
- On image-guided style transfer, OmniStyle achieves the highest style consistency (0.7483) and aesthetic appeal (5.7913) and the lowest style loss (0.1086) among compared methods.
- In a user study with 30 participants (1,800 votes), OmniStyle received 86.90% rank-1 for instruction-guided and 41.22% for image-guided tasks.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use OmniStyle instead of OmniGen | Instruction-guided style transfer · Instruction-guided style transfer benchmark | content preservation: 0.4918 -> 0.5128 | Helped |
| Use OmniStyle instead of OmniGen | Instruction-guided style transfer · Instruction-guided style transfer benchmark | style consistency: 0.5487 -> 0.6441 | Helped |
| Use OmniStyle instead of HQ-Edit | Instruction-guided style transfer · Instruction-guided style transfer benchmark | aesthetic appeal: 5.7777 -> 5.7512 | Hurt |
| Use OmniStyle instead of HQ-Edit | Instruction-guided style transfer · Instruction-guided style transfer benchmark | style loss: 0.3108 -> 0.2873 | Helped |
| Use OmniStyle instead of CAST | Image-guided style transfer · Image-guided style transfer benchmark | content preservation: 0.6165 -> 0.5450 | Hurt |
| Use OmniStyle instead of StyleShot | Image-guided style transfer · Image-guided style transfer benchmark | style consistency: 0.7347 -> 0.7483 | Helped |
| Use OmniStyle instead of StyleShot | Image-guided style transfer · Image-guided style transfer benchmark | aesthetic appeal: 5.7818 -> 5.7913 | Helped |
| Use OmniStyle instead of AesPANet | Image-guided style transfer · Image-guided style transfer benchmark | style loss: 0.1128 -> 0.1086 | Helped |
| Use OmniStyle instead of OmniGen | Instruction-guided style transfer · Instruction-guided style transfer user study | Rank 1: 5.16% -> 86.90% | Helped |
| Use OmniStyle instead of StyleShot | Image-guided style transfer · Image-guided style transfer user study | Rank 1: 18.63% -> 41.22% | Helped |
Surprising findings
- The paper's own evaluation metrics (OmniFilter-based scores) are used both to filter the training data and to evaluate models, which may favor OmniStyle.
Limitations
- Evaluation relies on OmniFilter's own scoring models, which are also used to filter training data, so the metrics are not independent.
- The paper notes the lack of ground-truth paired references, so objective evaluation of style accuracy is limited.
- Training data is generated by existing style transfer models, so quality is bounded by those teachers.
Experimental setup
Datasets: OmniStyle-1M Style30K AVA BAID
Benchmarks: OmniStyle evaluation benchmark (20 content x 100 style images)
Metrics: content preservation score style consistency score aesthetic appeal score style loss user study rank-1 proportion
Baselines: InstructPix2Pix HQ-Edit UltraEdit OmniGen DiffStyler ArtBank StyleID ArtFlow StyleShot AesPANet CSGO CAST
Builds on: FLUX OmniGen CLIP DINOv2 InternVL2
Closest papers by meaning
- OmniConsistency: Learning Style-Agnostic Consistency from Paired Stylization Data (NeurIPS 2025 · similarity 0.89)
- Style Injection in Diffusion: A Training-free Approach for Adapting Large-scale Diffusion Models for Style Transfer (CVPR 2024 · similarity 0.89)
- SigStyle: Signature Style Transfer via Personalized Text-to-Image Models (AAAI 2025 · similarity 0.89)
- HSI: A Holistic Style Injector for Arbitrary Style Transfer (CVPR 2025 · similarity 0.89)
- StyleSSP: Sampling StartPoint Enhancement for Training-free Diffusion-based Method for Style Transfer (CVPR 2025 · similarity 0.89)
- CLeaR: A Unified Framework for Resolving the Leakage–Degradation Dilemma in Style Transfer (NeurIPS 2026 · similarity 0.89)
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.