Research map / Video / Video generation
VGEdit: Unlocking Video Generation Priors for Reasoning-Informed Image Editing
Haiquan Lu, Gongfan Fang, Xinyin Ma, Xinchao Wang
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 MLLM instruction transfer to video-oriented descriptions (ablation) | HunyuanVideo-1.5 · RISE-Bench Overall | accuracy (%): 6.2 -> 15.2 | Improved |
| add instruction transfer (ablation) | HunyuanVideo-1.5 · KRIS-Bench Overall | overall score: 57.71 -> 67.96 | Improved |
| add instruction transfer (ablation) | HunyuanVideo-1.5 · ImgEdit Overall | overall score: 2.87 -> 3.56 | Improved |
| add RL adaptation on top of instruction transfer (ablation) | HunyuanVideo-1.5 + instruction transfer · RISE-Bench Overall | accuracy (%): 15.2 -> 22.3 | Improved |
| add RL adaptation (ablation) | HunyuanVideo-1.5 + instruction transfer · KRIS-Bench Overall | overall score: 67.96 -> 72.48 | Improved |
| add RL adaptation (ablation) | HunyuanVideo-1.5 + instruction transfer · ImgEdit Overall | overall score: 3.56 -> 4.33 | Improved |
| score a collage of the whole trajectory instead of the final frame (logit reward) (ablation) | VGEdit · RISE-Bench Overall | accuracy (%): 20.4 -> 22.3 | Improved |
| score the whole trajectory instead of the final frame (sampling reward) (ablation) | VGEdit · RISE-Bench Overall | accuracy (%): 15.7 -> 17.5 | Improved |
| use logit-based expectation reward instead of sampled token (trajectory-level) (ablation) | VGEdit · RISE-Bench Overall | accuracy (%): 17.5 -> 22.3 | Improved |
| use logit-based expectation reward instead of sampled token (trajectory-level) (ablation) | VGEdit · ImgEdit Overall | overall score: 4.01 -> 4.33 | Improved |
| use logit reward instead of sampling (final-frame) (ablation) | VGEdit · KRIS-Bench Overall | overall score: 69.34 -> 70.43 | Improved |
| repurpose video model with VGEdit instead of image editor Qwen-Image | VGEdit vs Qwen-Image · RISE-Bench Overall | accuracy (%): 8.9 -> 22.3 | Improved |
Closest papers by meaning
- Learning Action and Reasoning-Centric Image Editing from Videos and Simulation (NeurIPS 2024 · similarity 0.89)
- Reasoning to Edit: Hypothetical Instruction-Based Image Editing with Visual Reasoning (ICML 2026 · similarity 0.88)
- Pathways on the Image Manifold: Image Editing via Video Generation (CVPR 2025 · similarity 0.88)
- VINCIE-NExT: Unlocking Video Editing from Images via In-Context Modeling (NeurIPS 2026 · similarity 0.88)
- FireEdit: Fine-grained Instruction-based Image Editing via Region-aware Vision Language Model (CVPR 2025 · similarity 0.88)
- Edit-R2: Context-Aware Reinforcement Learning for Multi-Turn Image Editing (NeurIPS 2026 · 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.
