Finegan

The strongest thing about FineGAN is how cleanly its three-factor story is demonstrated. In the qualitative results every three rows share the same parent code p, so shape stays consistent while each row's child code c changes the appearance; shape is retained down a column and appearance across a row, with background code b held constant per column. That is a legible experiment, not just a claim, and it makes the paper a usable teaching artefact for stagewise conditional generation. The architecture is simple to describe and hard to beat for clarity: the background stage generates B from z and b; the parent stage uses B as a canvas, conditioned on z and p, to produce a hollow P with only the object's shape; the child stage conditions on c and uses P as a canvas to paint appearance into the shape outline. Background is retained across parent and child stages, which is why the factor separation holds up visually. The second result often gets less attention than the images: clustering real images using FineGAN's automatically learned features, presented as a first attempt at unsupervised fine-grained object category discovery. If your interest is labels-free category discovery rather than sample quality, that is the part to read. Weaknesses are the ordinary ones for research code. Experiments are birds, dogs and cars — three fine-grained domains, not a general-purpose generator. You supply the GAN and deep learning expertise to train and adapt it to a new domain. The framing is stagewise generation, not real-time generation. The site is a paper page with a demo video, code and approach diagram, not a product surface, so plan on reading the BibTeX and the paper rather than a support desk. Treat it as a baseline and an idea source; pair it with a modern generator when you need output quality.