Research map / Image generation / Editing and style transfer
PairEdit: Learning Semantic Variations for Exemplar-based Image Editing
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 PairEdit, a LoRA-based exemplar editing method that learns semantic variations from one or a few source-target image pairs without text, using a guidance-direction target noise, a content-preserving noise schedule, and separate content and semantic LoRAs on FLUX.1-dev.
Problem
Text prompts struggle to precisely specify certain editing semantics, and existing exemplar-based editing methods rely on text descriptions or implicit instructions and often fail to preserve content identity while learning complex semantics from paired examples.
Approach
Encode the semantic difference between paired images as a CFG-style guidance direction in the target noise prediction, use a content-preserving noise schedule x_t = x_0 + t*beta*eps to make the guidance weight time-independent, and jointly train a content LoRA and a semantic LoRA so the semantic LoRA learns the variation while the content LoRA handles reconstruction.
Key results
- On the quantitative comparison, PairEdit obtains lower LPIPS than SDEdit, Textual Slider, and Visual Slider at comparable CLIP editing magnitude across age, smile, chubbiness, and glasses (e.g., age LPIPS 0.1359 vs 0.1631 for the best baseline).
- In a user study with 720 responses from 24 participants, PairEdit was preferred over VISII (93.8%), Analogist (98.7%), and Slider (96.2%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use PairEdit instead of SDEdit, Textual Slider, or Visual Slider | FLUX.1-dev exemplar-based editing · Age | LPIPS distance: 0.1956 / 0.1631 / 0.1716 -> 0.1359 | Helped |
| Use PairEdit instead of SDEdit, Textual Slider, or Visual Slider | FLUX.1-dev exemplar-based editing · Smile | LPIPS distance: 0.1419 / 0.1749 / 0.1380 -> 0.1120 | Helped |
| Use PairEdit instead of SDEdit, Textual Slider, or Visual Slider | FLUX.1-dev exemplar-based editing · Chubbiness | LPIPS distance: 0.2173 / 0.1423 / 0.1747 -> 0.0815 | Helped |
| Use PairEdit instead of SDEdit, Textual Slider, or Visual Slider | FLUX.1-dev exemplar-based editing · Glasses | LPIPS distance: 0.1370 / 0.1602 / 0.1706 -> 0.0911 | Helped |
| Replace PairEdit semantic loss with visual concept loss (ablation) | PairEdit · Exemplar-based image editing | semantic fidelity | Hurt |
| Remove the content LoRA (ablation) | PairEdit · Exemplar-based image editing | identity preservation | Hurt |
| Replace the content-preserving noise schedule with a standard noise schedule (ablation) | PairEdit · Exemplar-based image editing | generalization capability | Hurt |
| Use guidance-based LoRA fusion instead of linear combination of LoRA weights | PairEdit real-image editing · Real image editing | identity preservation | Helped |
| Train with multiple image pairs instead of a single image pair | PairEdit · Exemplar-based image editing | generalization capability | Helped |
Limitations
- Requires paired source-target images and cannot directly use unpaired image sets.
- Real image editing relies on a simple per-image LoRA reconstruction rather than a proper inversion method.
- Quantitative evaluation uses only simple semantics and a CLIP-based editing magnitude proxy.
Experimental setup
Datasets: image pairs created with SDEdit and manual region transfer web-collected image pairs
Benchmarks: in-house editing semantics (age, smile, chubbiness, glasses)
Metrics: LPIPS CLIP similarity user preference
Baselines: VISII Analogist Edit Transfer Visual Concept Slider Textual Concept Slider SDEdit GPT-4o
Builds on: Concept Slider Flux Classifier-free guidance SDEdit Pair Customization
Closest papers by meaning
- Improving Editability in Image Generation with Layer-wise Memory (CVPR 2025 · similarity 0.88)
- InsightEdit: Towards Better Instruction Following for Image Editing (CVPR 2025 · similarity 0.88)
- RelationAdapter: Learning and Transferring Visual Relation with Diffusion Transformers (NeurIPS 2025 · similarity 0.88)
- ImgEdit: A Unified Image Editing Dataset and Benchmark (NeurIPS 2025 · similarity 0.87)
- Instruct-CLIP: Improving Instruction-Guided Image Editing with Automated Data Refinement Using Contrastive Learning (CVPR 2025 · similarity 0.87)
- Concept Lancet: Image Editing with Compositional Representation Transplant (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.