Generation of semantic segmentation masks based on conditional discrete flow matching

Recent approaches to generate semantic segmentation masks based on image generative models typically obtain discrete masks via continuous latent relaxations. We propose Conditional Discrete Flow Matching, an extension of Discrete Flow Matching that directly models probability distributions over discrete data to generate masks conditioned on images. We further improve accuracy with fewer sampling by balancing conditional and unconditional distributions using Classifier-Free Guidance (CFG). Experiments show that our method reaches accuracy comparable to existing approaches, and that CFG effectively boosts accuracy with reduced sampling.