Gen-Verse/MMaDA

Block diffusion sampling with semi-autoregressive text decoding enables unified handling of text reasoning, multimodal understanding, and image generation within a single modality-agnostic architecture. The framework integrates mixed chain-of-thought fine-tuning with UniGRPO, a policy-gradient RL algorithm designed for diffusion models, supporting diversified reward modeling across both reasoning and generation tasks. Models are distributed via Hugging Face with multiple checkpoints (Base, MixCoT, Parallel variants) and leverage the TraceRL training infrastructure for end-to-end optimization.

1,611 stars.