ReFocusEraser

This is the official code for the paper: "ReFocusEraser: Refocusing for Small Object Removal with Robust Context-Shadow Repair".

⭐ Update

• [2026.06] Release the pretrained VAE and LoRA weights on Hugging Face.
• [2026.02] Release the inference code.
• [2026.02] This repo is created.

✅ TODO

• [✅] Release the pretrained weights
• [✅] Release the inference code

Pretrained Weights

We have released the pretrained VAE and LoRA weights on Hugging Face:

👉 Zark12/ReFocusEraser

Please download the pretrained weights from the Hugging Face repository and specify the corresponding paths during inference.

Inference

Run the following command to try our model:

python inference.py \
  --img_path /path/to/imgs \
  --mask_path /path/to/masks \
  --output_path /path/to/output \
  --flux_path /path/to/FLUX.1-dev \
  --lora_path /path/to/ReFocusEraser_Checkpoint

If you use the pretrained weights released on Hugging Face, please replace /path/to/ReFocusEraser_Checkpoint with the downloaded LoRA weight path.