Research map / Image generation / Editing and style transfer
X2Edit: Revisiting Arbitrary-Instruction Image Editing Through Self-Constructed Data and Task-Aware Representation Learning
Jian Ma, Xujie Zhu, Zihao Pan, Qirong Peng, Xu Guo, Chen Chen, Haonan Lu
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 task embedding to MoE-LoRA gating (ablation) | MoE-LoRA (6 experts, rank 128) · GEdit-Bench++ CN | ImgEdit-Judge overall: 7.943 -> 8.087 | Improved |
| use MoE-LoRA instead of single LoRA (ablation) | FLUX.1 editing LoRA · GEdit-Bench++ CN | ImgEdit-Judge overall: 7.834 -> 7.943 | Improved |
| more experts with proportionally lower rank (6x128 -> 12x64) (ablation) | MoE-LoRA w/TA · GEdit-Bench++ EN | ImgEdit-Judge overall: 7.985 -> 8.084 | Improved |
| add task-aware contrastive regularization (cosine, 4th layer) (ablation) | MoE-LoRA w/TA 12x64 · GEdit-Bench++ CN | ImgEdit-Judge overall: 8.161 -> 8.253 | Improved |
| squared L2 distance instead of cosine similarity in contrastive loss (ablation) | X2Edit (4th layer) · GEdit-Bench++ CN | ImgEdit-Judge overall: 8.253 -> 8.289 | Improved |
| apply contrastive loss on all 19 layers instead of only the 4th (ablation) | X2Edit · GEdit-Bench++ EN | ImgEdit-Judge overall: 8.120 -> 8.334 | Improved |
| apply contrastive loss on all layers instead of 4th (ablation) | X2Edit · GEdit-Bench++ CN | ImgEdit-Judge overall: 8.289 -> 8.313 | Improved |
| plug X2Edit into FLUX.1 community variants (Schnell, Krea-dev, etc.) | X2Edit · GEdit-Bench++ (Table 4) | ImgEdit-Judge overall: 8.313 vs 8.254 to 8.458 (CN) | No clear effect |
| X2Edit vs full-parameter fine-tuned Bagel and Kontext | X2Edit (0.9B) · KontextBench | VIEScore (Qwen2.5-VL-72B): 7.606 vs Bagel 7.929 | Hurt |
Later papers that compared with this method
1 later accepted paper reports X2Edit as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- HP-Edit: A Human-Preference Post-Training Framework for Image Editing (CVPR 2026 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- AnyEdit — AnyEdit: Mastering Unified High-Quality Image Editing for Any Idea (CVPR 2025) · better on 1 of 1 shared result
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
- AnyEdit: Mastering Unified High-Quality Image Editing for Any Idea (CVPR 2025 · similarity 0.93)
- InsightEdit: Towards Better Instruction Following for Image Editing (CVPR 2025 · similarity 0.92)
- UltraEdit: Instruction-based Fine-Grained Image Editing at Scale (NeurIPS 2024 · similarity 0.91)
- ImgEdit: A Unified Image Editing Dataset and Benchmark (NeurIPS 2025 · similarity 0.91)
- Multi-Reward as Condition for Instruction-based Image Editing (ICLR 2025 · similarity 0.90)
- Emu Edit: Precise Image Editing via Recognition and Generation Tasks (CVPR 2024 · similarity 0.90)
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.
