Installing and Testing DragGAN Locally: Matthew Berman’s 2023 Walkthrough

DragGAN lets you select points on an image and move features toward target positions. The code had just been released when I made this July 2023 video, and I found the installation confusing and buggy. These steps describe that early version, not a guarantee that the current repository has the same commands or dependencies. The video is sponsored by Rosebud.

The demo shows how moving a point can reshape the subject. The examples I run here use images generated by pretrained GAN models. Although I initially describe working with any image, the project explains that editing a real photograph first requires a separate GAN inversion step; arbitrary photos are not simply interchangeable with these generated examples.

I use Conda to manage the Python environment. I demonstrate the process on a Mac and say I got a similar setup running on Windows. The documentation was changing during my testing, so I copy the repository URL, clone it, enter the DragGAN directory, and try `conda env create -f environment.yml`.

The initial environment creation cannot resolve two packages on my machine. In the then-current `environment.yml`, I remove the CUDA toolkit dependency for my Mac and move SciPy from the Conda dependencies into the pip section, changing its version separator from `=` to `==`. CUDA is an NVIDIA technology; requirements on a Windows machine with an NVIDIA GPU differ.

For Windows I mention that the documented CUDA version was hard to find and that I tried another version, but compatibility depends on the actual GPU, driver and package versions. After saving the file, I rerun environment creation and activate the environment with `conda activate stylegan3`. These workarounds are specific to the release and machine I tested.

I install the requirements using `pip install -r requirements.txt`, correcting the filename I saw in the documentation. I see an error and think I can continue, but do not establish that it is harmless in every environment. On my Mac I set `PYTORCH_ENABLE_MPS_FALLBACK=1` so unsupported MPS operations can fall back to CPU; that setting does not guarantee that the whole application runs on the GPU.

Next I run `python scripts/download_model.py` to download the pretrained models. The downloads can be large, and the available checkpoints correspond to different subjects, including lions, dogs, horses, elephants and faces. Choosing an appropriate model matters for the type of image being generated and edited.

The native GUI launch script, `sh scripts/gui.sh`, produces a segmentation fault on my machine. My workaround is `python visualizer_drag_gradio.py`, which starts a local web interface that I open in the browser. This is an observed workaround for my setup, not a universal fix for every crash.

The interface provides a pretrained-model selector, settings, point controls and start and stop buttons. I choose a lion, select where its head is and where I want it to go, and start the edit. The head visibly moves, showing the point-based effect.

The display seems to stop updating its step count after 100 even though the terminal continues reporting steps. I cannot tell from that alone whether useful changes are still happening, so I stop the process and download the result. In my Windows test the CPU appeared to be doing the work; I had not got GPU acceleration working there or Metal working as intended on my Mac.

I reset the points and move the lion again, then switch to the cat model and open the cat’s eye. Finally I select the FFHQ human-face model and try to make the subject smile and turn her head. The smile changes, but the attempted head turn does not visibly work in the test.

After resetting the points, I move the lower lip instead; the smile becomes much larger, if a little uncanny. The early release can produce striking changes, but my experiments also show that installation, acceleration and some edits were unreliable. I plan to revisit the tool once it is more stable.