Research map / Image generation / Editing and style transfer
LoRAShop: Training-Free Multi-Concept Image Generation and Editing with Rectified Flow Transformers
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 LoRAShop, a training-free method that extracts per-subject masks from FLUX cross-attention and blends multiple LoRA adapters' residual features only inside each subject's region for multi-concept generation and editing.
Problem
Composing multiple independently trained LoRA adapters for personalized concepts causes cross-talk and identity distortion, and existing methods target generation rather than region-controlled multi-concept editing of real images.
Approach
Derive non-overlapping subject masks from double-stream cross-attention in a short pseudo-denoising pass, then blend each LoRA's residual features only within its mask at every transformer block, with no weight changes or training.
Key results
- Single-subject generation: ID 0.740 vs DreamBooth 0.755 and InfiniteYou 0.683; best CLIP-T (0.439 tied) and HPS (0.321).
- Multi-subject generation: ID 0.532 vs next best OmniGen 0.453; HPS 0.260 best among FLUX-based methods.
- Face swapping: ID 0.709 vs ReFace 0.330, but LPIPS 0.050 worse than ReFace 0.031; user study Q1 3.762 vs 3.340 for next best.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use LoRAShop instead of DreamBooth | Single-subject generation with FLUX.1-dev · Single-subject generation | ID: 0.755 ± 0.089 -> 0.740 ± 0.066 | Hurt |
| Use LoRAShop instead of InfiniteYou | Single-subject generation with FLUX.1-dev · Single-subject generation | CLIP-T: 0.439 ± 0.039 -> 0.439 ± 0.047 | No clear effect |
| Use LoRAShop instead of Omni-Gen | Single-subject generation with FLUX.1-dev · Single-subject generation | HPS: 0.311 ± 0.030 -> 0.321 ± 0.028 | Helped |
| Use LoRAShop instead of Omni-Gen | Single-subject generation with FLUX.1-dev · Single-subject generation | Aesthetics: 6.514 ± 0.448 -> 6.499 ± 0.529 | Hurt |
| Use LoRAShop instead of OmniGen | Multi-subject generation with FLUX.1-dev · Multi-subject generation | ID: 0.453 ± 0.09 -> 0.532 ± 0.12 | Helped |
| Use LoRAShop instead of OmniGen | Multi-subject generation with FLUX.1-dev · Multi-subject generation | CLIP-T: 0.256 ± 0.08 -> 0.252 ± 0.08 | Hurt |
| Use LoRAShop instead of DreamBooth | Multi-subject generation with FLUX.1-dev · Multi-subject generation | HPS: 0.259 ± 0.04 -> 0.260 ± 0.04 | No clear effect |
| Use LoRAShop instead of OmniGen | Multi-subject generation with FLUX.1-dev · Multi-subject generation | Aesthetics: 6.264 ± 0.26 -> 6.124 ± 0.29 | Hurt |
| Use LoRAShop instead of ReFace | Face swapping · Face swapping | ID: 0.330 ± 0.091 -> 0.709 ± 0.101 | Helped |
| Use LoRAShop instead of ReFace | Face swapping · Face swapping | DINO: 0.982 ± 0.012 -> 0.970 ± 0.019 | Hurt |
| Use LoRAShop instead of ReFace | Face swapping · Face swapping | CLIP-I: 0.940 ± 0.038 -> 0.926 ± 0.037 | Hurt |
| Use LoRAShop instead of ReFace | Face swapping · Face swapping | LPIPS: 0.031 ± 0.033 -> 0.050 ± 0.019 | Hurt |
Surprising findings
- DreamBooth obtains the highest single-subject ID score (0.755), above LoRAShop's 0.740, despite LoRAShop being training-free.
Limitations
- Masks inherit biases of the underlying diffusion model and may mislocate underrepresented regions or concepts.
- Mask extraction is tuned to Flux attention patterns and may need re-tuning of thresholds for other backbones such as SDXL-Turbo.
- Face-swap LPIPS is worse than the inpainting baseline ReFace, and DreamBooth has higher single-subject ID.
Experimental setup
Datasets: publicly available LoRA adapters from Hugging Face (15 identity LoRAs, plus object/style LoRAs)
Benchmarks: single-subject generation benchmark (15 identities x 15 prompts, 520 images) multi-subject generation benchmark (two-subject pairs) face swapping benchmark vs ReFace
Metrics: ID (ArcFace) CLIP-T HPS Aesthetics DINO CLIP-I LPIPS user study Likert ratings
Baselines: DreamBooth IP-Adapter-FLUX InfiniteYou OmniGen UNO OMG MS-Diffusion MIP-Adapter ReFace
Builds on: FLUX.1-dev LoRA RF-Solver DreamBooth
Closest papers by meaning
- LoRACLR: Contrastive Adaptation for Customization of Diffusion Models (CVPR 2025 · similarity 0.90)
- Cached Multi-Lora Composition for Multi-Concept Image Generation (ICLR 2025 · similarity 0.88)
- LoRAtorio: An intrinsic approach to LoRA Skill Composition (NeurIPS 2026 · similarity 0.87)
- Enhancing Creative Generation on Stable Diffusion-based Models (CVPR 2025 · similarity 0.86)
- Concept Lancet: Image Editing with Compositional Representation Transplant (CVPR 2025 · similarity 0.86)
- Conditional Balance: Improving Multi-Conditioning Trade-Offs in Image Generation (CVPR 2025 · similarity 0.85)
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.