lixiaowen-xw/DiffuEraser

Builds on Stable Diffusion v1.5 with a dual-branch architecture combining a denoising UNet and auxiliary BrushNet module, integrating temporal attention mechanisms and prior conditioning to enhance temporal consistency across video frames. Leverages temporal receptive field expansion and video diffusion smoothing for long-sequence inference, with post-processing mask blending to refine inpainted regions. Available on Hugging Face and ModelScope with support for multi-resolution inference (360p–720p) and two-stage training pipeline.

625 stars. No commits in the last 6 months.