gandissect

Why we included this project

If you train or fine-tune image generation GANs and wonder which internal units actually fire for trees, windows, or doors, GANDissect answers that directly. It maps the neurons of a generative network onto human-interpretable concepts, then lets you inspect the results as a static summary or through the interactive GANPaint demo, where you can draw new objects into a scene and watch the model respond. The same unit-level control lets you diagnose artifacts and remove or insert features, like erasing windows from a room, by manipulating the units that produce them. This is research code tied to a 2018 paper, so expect a conda setup and a CUDA GPU rather than a polished application, but for anyone doing careful analysis of why a GAN renders the way it does, it is still the clearest option around.