What it does

This repo implements DragGAN, the SIGGRAPH 2023 technique for manipulating StyleGAN-generated images by placing handle points and dragging them to target positions. You click a nose, drag it left, the face follows. It wraps the core algorithm in a Gradio GUI, a Colab notebook, and a pip install draggan package for Windows, macOS, and Linux.

The interesting bit

The project beat the official code to release and accumulated nearly 5K stars in the process. It also integrated into InternGPT, a larger pointing-language visual system, suggesting the authors see drag-based editing as one component of a broader interactive toolkit rather than a standalone novelty.

Key highlights

• pip install draggan with automatic checkpoint downloads
• Gradio GUI, Colab demo, and Docker image available
• Supports StyleGAN2-ada for higher-quality outputs
• Custom image upload via GAN inversion (with noted limitations)
• Multiple handle points and movable regions supported

Caveats

• The README explicitly warns that custom images may distort due to GAN inversion limitations, and manipulations can fail
• The official DragGAN code has since been released, making this unofficial implementation’s long-term maintenance unclear
• The project is now a sub-project of InternGPT; future non-DragGAN tools will land there instead

Verdict

Worth a spin if you want to test drive DragGAN without wrestling with the official codebase, or if you need the Gradio/Colab conveniences. Skip if you require production reliability or are already committed to the official implementation.