Reproduce “Constrained Generative Adversarial Networks for Interactive Image Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Eric Heim
CONstrained GAN (CONGAN), a generative adversarial network framework, allows iterative, human-in-the-loop image generation guided by relative pairwise constraints. This framework maintains image quality comparable to unsupervised GANs while effectively satisfying over 90% of constraints on MNIST and approximately 87% on CelebA, enabling interactive refinement of generated images.
Key insights
- Interactive image generation can be effectively guided by relative pairwise assessments, offering a more intuitive and flexible user interface than fixed attributes or latent space manipulation.
- A generator architecture combining CNNs, attention-based LSTMs for set processing, and transpose convolutions can successfully handle variable numbers of order-invariant constraints.
- It is possible to integrate human-in-the-loop control into GANs without substantially degrading the quality of the generated images compared to unsupervised models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.