ROMS-IMLE: A Minimalist Approach to Competitive Single-Step Generative Modelling
The paper says a single-step IMLE model can reach competitive ImageNet sample quality without iterative denoising.
Vashist and Li strip the setup down to Implicit Maximum Likelihood Estimation and a moderately sized convolutional network. They avoid variational inference, adversarial training, numerical integration, transformers, and diffusion-style multi-step refinement. The resulting ROMS-IMLE model reports an FID of 2.56 on ImageNet 256, with good precision and recall. Source: ArXiv · AI/CL/LG's note
Vashist and Li strip the setup down to Implicit Maximum Likelihood Estimation and a moderately sized convolutional network. They avoid variational inference, adversarial training, numerical integration, transformers, and diffusion-style multi-step refinement. The resulting ROMS-IMLE model reports an FID of 2.56 on ImageNet 256, with good precision and recall. Source: ArXiv · AI/CL/LG's note
score 4