What it does

DragGAN gives you a GUI where you click and drag points on a GAN-generated image to reshape it—pull a nose left, widen eyes, rotate a car. The model figures out how to move the corresponding latent-space representation so the edit stays semantically coherent. It ships with a Gradio web UI, a local GUI, and a Docker image that’ll cost you 25 GB of disk.

The interesting bit

The trick isn’t just moving pixels; it’s optimizing in the GAN’s latent manifold so the result still looks like a plausible, generated image rather than a smeared Photoshop job. The catch: it only natively edits GAN-generated images. Want to edit a real photo? You need to invert it into latent space first using something like PTI.

Key highlights

• Built on StyleGAN3; supports StyleGAN-Human and LHQ checkpoints
• Gradio demo runs cross-platform (Windows/Linux); native GUI via shell/batch scripts
• Docker setup available with GPU passthrough for CUDA acceleration
• macOS M1/M2 supported via MPS fallback, or CPU-only if you’re patient
• CC-BY-NC license for the DragGAN algorithm itself; StyleGAN3-derived code under Nvidia’s license

Caveats

• Real-image editing requires separate GAN inversion tooling (not included)
• Docker image is a 25 GB download
• Must preserve “AI Generated” watermarking in any derivative

Verdict

Grab this if you’re doing research on controllable generation or need an intuitive demo for manipulating StyleGAN outputs. Skip if you’re looking for a general-purpose real-image editor—this is a specialized latent-space tool, not a replacement for your photo workflow.