Spaces:
Running on Zero
Running on Zero
| title: RePlan Image Editing | |
| emoji: 🎯 | |
| colorFrom: blue | |
| colorTo: pink | |
| sdk: gradio | |
| sdk_version: 6.20.0 | |
| app_file: app.py | |
| short_description: Reasoning-guided region planning for image editing | |
| python_version: "3.12" | |
| startup_duration_timeout: 1h | |
| pinned: false | |
| # RePlan: Reasoning-guided Region Planning for Complex Instruction-based Image Editing | |
| RePlan couples a reasoning VLM planner with a diffusion editor to handle | |
| **complex, instruction-based image editing**. The planner | |
| ([`TainU/RePlan-Qwen2.5-VL-7B`](https://huggingface.co/TainU/RePlan-Qwen2.5-VL-7B)) | |
| performs chain-of-thought reasoning to decompose an instruction into | |
| region-specific bounding boxes + local hints and a global prompt. The editor | |
| then executes the edit using training-free **region-constrained attention | |
| injection**, keeping edits localized without spillover. | |
| - Editor backbone: [`FLUX.2-klein-4B`](https://huggingface.co/black-forest-labs/FLUX.2-klein-4B) (Apache-2.0, distilled 4-step) | |
| - Paper: https://huggingface.co/papers/2512.16864 | |
| - Code: https://github.com/dvlab-research/RePlan | |