Generative Modeling
Overview
Generative modeling studies how to learn the structure and variability of data so a model can produce new samples, fill in missing information, or propose designs. Major approaches include variational autoencoders, generative adversarial networks, autoregressive models, normalizing flows, energy-based models, and diffusion, score-based, and flow-matching methods. Researchers evaluate these methods on tasks such as image and video synthesis, text generation, molecular design, and scientific simulation, using measures of fidelity, diversity, coverage, and downstream usefulness. Current work seeks faster, more controllable generation that respects constraints and data geometry, while addressing trade-offs among quality, diversity, reliability, compute cost, and memorization.
Latest papers
546 papers38 papers in the last four weeks, down 46% on the four weeks before. 0.3% of all new papers.
Random-Bridges as Stochastic Transports for Generative Models
Emergence of Nonequilibrium Latent Cycles in Unsupervised Generative Modeling
Is Generation Required for Data-Efficient Perception?
Distribution Matching Variational AutoEncoder
Generative Anchored Fields: Endpoint Learning for Generative Modeling
GENA3D: Generative Amodal 3D Modeling by Bridging 2D Priors and 3D Coherence
CrochetBench: Can Vision-Language Models Move from Describing to Doing in Crochet Domain?
A Likely Geometry of Generative Models
AtomBench: A Benchmarking Framework for Generative Crystal Reconstruction Models in Conventional Superconductors
atombench, and we encourage community use. https://github.com/atomgptlab/atombench