| language: en | |
| pipeline_tag: image-to-image | |
| tags: | |
| - lora | |
| - street-view | |
| - image-editing | |
| # CrossTimeEdit | |
| LoRA weights for **CrossTimeEdit: A Decade-Spanning Cross-View Dataset and Reward-Guided Editing for Historical Street-View Generation**. | |
| This is the final epoch-200 checkpoint, selected on the validation set after supervised fine-tuning and multi-reward online reinforcement learning. The adapter is based on FLUX.2 [Klein] 4B and is loaded using DiffSynth-Studio; it is not a standalone base model. | |
| - [Implementation and inference instructions](https://github.com/LuHanwen67/CrossTimeEdit-release) | |
| - [Project page](https://luhanwen67.github.io/CrossTimeEdit-release/) | |
| - [VIGOR-his metadata](https://huggingface.co/datasets/LuHanwen/VIGOR-his-metadata) | |
| ## Download | |
| ```bash | |
| hf download LuHanwen/CrossTimeEdit-LoRA crosstimeedit.safetensors --local-dir weights | |
| ``` | |
| Use `weights/crosstimeedit.safetensors` with `inference/generate.py` in the code repository. Download the base model separately and observe its applicable license and terms. | |
| The generated historical appearance is an estimate, not a verified historical record. Satellite evidence may not reveal historical facade details or street-level occlusions. | |
| SHA-256 (`crosstimeedit.safetensors`): `0f6d50733becc89d23e93ded9d957dc12e514347f0f8b69e4ad394ad79a18aa6`. | |