Research map / Large language models / Alignment and preferences
EditReward: A Human-Aligned Reward Model for Instruction-Guided Image Editing
Keming Wu, Sicong Jiang, Max Ku, Ping Nie, Minghao Liu, Wenhu Chen
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 |
|---|---|---|---|
| train Qwen2.5-VL-7B on EditReward-Data with the reward model recipe | Qwen2.5-VL-7B · GenAI-Bench | pairwise accuracy: 40.48 -> 63.97 | Improved |
| use pair-wise uncertainty ranking loss instead of point-wise regression (ablation) | EditReward · GenAI-Bench | pairwise accuracy: 49.62 -> 63.97 | Improved |
| use multiple independent reward heads instead of a shared head (ablation) | EditReward · GenAI-Bench | pairwise accuracy: 60.17 -> 63.97 | Improved |
| use mean aggregation instead of min or sum (ablation) | EditReward · GenAI-Bench | pairwise accuracy: 59.96 (min), 59.63 (sum) -> 63.97 | Improved |
| use sum aggregation rather than mean (ablation) | EditReward · EditReward-Bench | accuracy: 36.78 -> 37.60 | Mixed |
| increase backbone from Qwen2.5-VL-3B to 7B (ablation) | EditReward · GenAI-Bench | pairwise accuracy: 62.79 -> 63.97 | Improved |
| use MiMo-VL-7B backbone instead of Qwen2.5-VL-7B (ablation) | EditReward · AURORA-Bench | pairwise accuracy: 59.50 -> 63.62 | Improved |
| scale data from 130k to 200k and add tie disentanglement (ablation) | EditReward · ImagenHub | Spearman correlation: 32.45 -> 36.18 | Improved |
| scale data from 130k to 200k and add tie disentanglement (ablation) | EditReward · EditReward-Bench | accuracy: 37.81 -> 36.78 | Hurt |
| fine-tune on EditReward-selected Top 20K subset instead of full noisy 46K set | Step1X-Edit · GEdit-Bench-EN | overall score G_O: 6.780 -> 7.086 | Improved |
| select Top 30K or Top 10K rather than Top 20K (ablation) | Step1X-Edit curation · GEdit-Bench-EN | overall score G_O: 7.086 -> 6.962 (30K), 6.938 (10K) | Hurt |
| fine-tune Step1X-Edit on full unfiltered ShareGPT-4o-Image | Step1X-Edit · GEdit-Bench-EN-I | perceptual quality G_PQ: 6.962 -> 6.838 | Hurt |
Later papers that compared with this method
1 later accepted paper reports EditReward as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Auto-Rubric as Reward: From Implicit Preference to Explicit Generative Criteria (NeurIPS 2026 · 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
- HIVE: Harnessing Human Feedback for Instructional Visual Editing (CVPR 2024 · similarity 0.90)
- Multi-Reward as Condition for Instruction-based Image Editing (ICLR 2025 · similarity 0.89)
- ImgEdit: A Unified Image Editing Dataset and Benchmark (NeurIPS 2025 · similarity 0.89)
- RetouchIQ: MLLM Agents for Instruction-Based Image Retouching with Generalist Reward (CVPR 2026 · similarity 0.88)
- Towards Scalable Human-aligned Benchmark for Text-guided Image Editing (CVPR 2025 · similarity 0.88)
- AnyEdit: Mastering Unified High-Quality Image Editing for Any Idea (CVPR 2025 · similarity 0.88)
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.
