Research map / Image generation / Text-to-image diffusion
StyleDoctor: Towards Specialist Reward Model for Style-centric Generation Tasks
Xilin He, Xiaole Xian, Xiangyu Yue, Muhammad Haris Khan
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add global style feature learning and unified preference learning (both) (ablation) | StyleDoctor (Qwen2.5-VL-3B) · SPRData unseen test | MM Und. Acc.: 66.54 -> 71.14 | Improved |
| add global style feature learning (contrastive stage) (ablation) | StyleDoctor · SPRData unseen test | MM Und. Acc.: 66.54 -> 69.63 | Improved |
| add unified preference learning (ablation) | StyleDoctor · SPRData unseen test | MM Und. Acc.: 66.54 -> 69.70 | Improved |
| add both components on style customization (ablation) | StyleDoctor reward · Style customization CSD | CSD: 0.69 -> 0.78 | Improved |
| use StyleDoctor reward in RL fine-tuning of OmniGen2 | OmniGen2 · Text-controlled style generation | style accuracy (StyleDoctor/GPT-4o): 65.42/58.85 -> 69.58/67.60 | Improved |
| use StyleDoctor reward in RL fine-tuning of Flux-Kontext | Flux-Kontext · Instruction-guided style transfer | GPT-4o style consistency: 74.35 -> 78.54 | Improved |
| use StyleDoctor reward in RL fine-tuning of OmniStyle | OmniStyle · Reference image-guided style transfer | CSD: 0.74 -> 0.78 | Improved |
| replace CSD regularizer with StyleDoctor regularizer | B-LoRA · Style customization (SDXL) | CSD: 0.75 -> 0.78 | Improved |
| use human-preference reward model HPSv2 instead of StyleDoctor as reward | B-LoRA style customization · Style generation | style consistency: 0.78 -> 0.45 | Hurt |
| use HPSv3 instead of StyleDoctor as reward | B-LoRA style customization · Style generation | style consistency: 0.78 -> 0.64 | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 4 of 4 on most shared results.
- OmniStyle — OmniStyle: Filtering High Quality Style Transfer Data at Scale (CVPR 2025) · better on 3 of 3 shared results
- StyleStudio — StyleStudio: Text-Driven Style Transfer with Selective Control of Style Elements (CVPR 2025) · better on 1 of 1 shared result
- CSGO — CSGO: Content-Style Composition in Text-to-Image Generation (NeurIPS 2025) · better on 1 of 1 shared result
- HPSv3 — HPSv3: Towards Wide-Spectrum Human Preference Score (ICCV 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- DRM: Diffusion-based Reward Model With Step-wise Guidance (CVPR 2026 · similarity 0.88)
- Aligning Vision Models with Human Aesthetics in Retrieval: Benchmarks and Algorithms (NeurIPS 2024 · similarity 0.87)
- Multimodal Large Language Models Make Text-to-Image Generative Models Align Better (NeurIPS 2024 · similarity 0.86)
- PrefPaint: Aligning Image Inpainting Diffusion Model with Human Preference (NeurIPS 2024 · similarity 0.86)
- UniAR: A Unified model for predicting human Attention and Responses on visual content (NeurIPS 2024 · similarity 0.86)
- Diffusion Model as a Noise-Aware Latent Reward Model for Step-Level Preference Optimization (NeurIPS 2025 · similarity 0.86)
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.
