Research map / Image generation / Text-to-image diffusion
InstantRetouch: Personalized Image Retouching without Test-time Fine-tuning
Temesgen Muruts Weldengus, Binnan Liu, Fei Kou, Youwei Lyu, Jinwei Chen, Changqing Zou, Qingnan Fan
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 LoRA adaptation to SigLIPv2 style encoder instead of frozen (ablation) | RefRetouch encoder · Ablation validation set | PSNR (dB): 23.41 -> 32.40 | Improved |
| use SigLIPv2-LoRA instead of ResNet-50 encoder (ablation) | RefRetouch encoder · Ablation validation set | PSNR (dB): 30.00 -> 32.40 | Improved |
| use color-space MLP decoder instead of UNet decoder (ablation) | RefRetouch decoder · Ablation validation set | PSNR (dB): 29.58 -> 32.40 | Improved |
| use additive conditioning instead of AdaLN (ablation) | RefRetouch decoder · Ablation validation set | PSNR (dB): 31.93 -> 32.40 | Improved |
| use additive conditioning instead of cross-attention (ablation) | RefRetouch decoder · Ablation validation set | PSNR (dB): 22.66 -> 32.40 | Improved |
| retrieve top-3 relevant references instead of using all references (ablation) | RAR · MIT-FiveK (20 references) | PSNR (dB): 21.99 -> 22.95 | Improved |
| increase retrieved references from K=3 to K=5 (ablation) | RAR · MIT-FiveK (20 references) | PSNR (dB): 22.95 -> 23.37 | Improved |
| use LoRA-adapted SigLIPv2 instead of frozen SigLIPv2 for retrieval embeddings (ablation) | RAR · MIT-FiveK (20 references) | PSNR (dB): 21.02 -> 22.95 | Improved |
| increase number of references 1 to 4 on single-style benchmark | RefRetouch · VCIRB | PSNR (dB): 29.13 -> 29.82 | No clear effect |
| increase references 1 to 6 | RefRetouch · PPR10K-groups | PSNR (dB): 22.51 -> 25.27 | Improved |
| retrain MSM on the paper's dataset | MSM · MIT-FiveK reconstruction | PSNR (dB): 28.94 -> 26.32 | Hurt |
| compare to per-user trained AdaInt/RSFNet on PPR10K users | RefRetouch (tuning-free) · PPR10K users | PSNR (dB): 21.75 vs 22.01 (20 refs); 22.47 vs 22.82 (100 refs) | Mixed |
Closest papers by meaning
- PerTouch: VLM-Driven Agent for Personalized and Semantic Image Retouching (AAAI 2026 · similarity 0.89)
- DiffRetouch: Using Diffusion to Retouch on the Shoulder of Experts (AAAI 2025 · similarity 0.87)
- RetouchGPT: LLM-based Interactive High-Fidelity Face Retouching via Imperfection Prompting (AAAI 2025 · similarity 0.84)
- JarvisArt: Liberating Human Artistic Creativity via an Intelligent Photo Retouching Agent (NeurIPS 2025 · similarity 0.84)
- InstantBooth: Personalized Text-to-Image Generation without Test-Time Finetuning (CVPR 2024 · similarity 0.84)
- Tailored Visions: Enhancing Text-to-Image Generation with Personalized Prompt Rewriting (CVPR 2024 · similarity 0.84)
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.
