synthetic-data-generation

MMDiff extends frozen diffusion transformers into multi-modal generative systems using lightweight decoders, achieving significant improvements in semantic segmentation and other perceptual tasks through multi-timestep feature fusion.