Research map / Image generation / Text-to-image diffusion
Implicit Style-Content Separation using B-LoRA
Yarden Frenkel, Yael Vinker, Ariel Shamir, Danny Cohen-Or
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 |
|---|---|---|---|
| train LoRA only on blocks 4 and 5 (B-LoRA) instead of all blocks to separate style and content (ablation) | B-LoRA on SDXL · image stylization | storage reduction: 70% less storage | Improved |
| optimize blocks {4,5} instead of {2,5} (ablation) | B-LoRA on SDXL · single-image reconstruction | qualitative reconstruction and content capture | Improved |
| use the prompt 'A [v]' instead of 'A [v] <class-name>' during B-LoRA training (ablation) | B-LoRA on SDXL · style transfer between objects | qualitative content preservation and style transfer | Improved |
| inject 'A [c]' into content block W4 and 'A [c] in [s] style' into other blocks at inference (ablation) | B-LoRA on SDXL · complex scenes | qualitative generation quality | Improved |
| scale style adapter W5 by alpha 0.4-0.5 (ablation) | B-LoRA on SDXL · identity preservation | qualitative color preservation | Improved |
| use B-LoRA instead of DB-LoRA + ControlNet | style transfer · Style transfer (multiple content images) | DINO style similarity: 0.863 -> 0.881 | Improved |
| use B-LoRA instead of ZipLoRA on content preservation | style transfer · Content preservation (multiple content images) | DINO content similarity: 0.841 -> 0.790 | Hurt |
| use B-LoRA instead of SWAG (non-SDXL) for style transfer | style transfer · Style transfer (single image) | DINO style similarity: 0.883 -> 0.881 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 2 on most shared results.
- ZipLoRA — ZipLoRA: Any Subject in Any Style by Effectively Merging LoRAs (ECCV 2024) · better on 1 of 2 shared results
- StyTr2 — StyTr2: Image Style Transfer With Transformers (CVPR 2022) · better on 1 of 2 shared results
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
- LoRAtorio: An intrinsic approach to LoRA Skill Composition (NeurIPS 2026 · similarity 0.88)
- Cached Multi-Lora Composition for Multi-Concept Image Generation (ICLR 2025 · similarity 0.88)
- K-LoRA: Unlocking Training-Free Fusion of Any Subject and Style LoRAs (CVPR 2025 · similarity 0.88)
- Conditional Balance: Improving Multi-Conditioning Trade-Offs in Image Generation (CVPR 2025 · similarity 0.87)
- SigStyle: Signature Style Transfer via Personalized Text-to-Image Models (AAAI 2025 · similarity 0.87)
- CSGO: Content-Style Composition in Text-to-Image Generation (NeurIPS 2025 · 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.
