Research map / Image generation / Text-to-image diffusion
CRAFT-LoRA: Content-Style Personalization via Rank-Constrained Adaptation and Training-Free Fusion
Yu Li, Yujun Cai, Chi Zhang
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 time-dependent scaling of LoRA weights in ACFG (ablation) | ACFG · ACFG ablation | quality score: 0.78 -> 0.86 | Improved |
| add low-rank constraint on temporal adjustments (ablation) | ACFG · ACFG ablation | quality score: 0.82 -> 0.86 | Improved |
| anchor the unconditional path to the rank-limited backbone (asymmetric CFG) (ablation) | ACFG · ACFG ablation | quality score: 0.75 -> 0.86 | Improved |
| replace fixed weights with ACFG | LoRA fusion · complex composition pairing | composite quality score: 0.65 -> 0.80 | Improved |
| replace fixed weights with ACFG | LoRA fusion · distant semantics pairing | composite quality score: 0.68 -> 0.79 | Improved |
| replace fixed weights with ACFG | LoRA fusion · similar semantics pairing | composite quality score: 0.81 -> 0.87 | Improved |
| add rank-limited trunk fine-tuning plus layer-selective LoRA | Standard LoRA · Content-style disentanglement | cross-influence Sx: 0.42 -> 0.18 | Improved |
| use hierarchical rank schedule (128 to 4) instead of uniform rank 64 (ablation) | Rank-limited backbone fine-tuning · content-style personalization | Content Sim / Style Sim / cross-influence: Content Similarity by 4.8%, Style Similarity by 5.2%, and reduces cross-influence by 11.3% | Improved |
| add Rank-FT to direct merging (ablation) | CRAFT-LoRA · component ablation | Content Sim / Style Sim: +0.08 Content Sim, +0.10 Style Sim | Improved |
| increase contrastive pairs from 50 to 200 (ablation) | Rank-limited backbone fine-tuning · content-style personalization | overall performance: 50 pairs already achieve 95% of full performance, while increasing to 200 pairs yields less than 2% additional improveme | No clear effect |
| combine rank-limited updates with contrastive pairs | CRAFT-LoRA backbone · content-style personalization | CLIP content-style similarity: 12% reduction | Improved |
Closest papers by meaning
- NP-LoRA: Null Space Projection for Subject-Style LoRA Fusion (NeurIPS 2026 · similarity 0.90)
- LoRAtorio: An intrinsic approach to LoRA Skill Composition (NeurIPS 2026 · similarity 0.89)
- LoRACLR: Contrastive Adaptation for Customization of Diffusion Models (CVPR 2025 · similarity 0.88)
- Cached Multi-Lora Composition for Multi-Concept Image Generation (ICLR 2025 · similarity 0.88)
- FineStyle: Fine-grained Controllable Style Personalization for Text-to-image Models (NeurIPS 2024 · similarity 0.88)
- K-LoRA: Unlocking Training-Free Fusion of Any Subject and Style LoRAs (CVPR 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.
