VeRL-Omni
VeRL-Omni is a reinforcement-learning training framework for diffusion, image, video, audio and other multimodal generative models, built as a dedicated extension of the verl RL library.
VeRL-Omni is a general reinforcement-learning training framework focused on multimodal generative models. It grew out of multi-modal generation RL work inside the verl project and now has its own home so it can evolve around the compute and I/O patterns of diffusion and omni-modality training rather than text-only LLM RL.
It targets three families of models: diffusion generative models for image, video and audio; unified multimodal understanding-and-generation models; and omni-modality models that jointly handle text, image, audio and video. Supported techniques include a trainer built on a transfer queue and replay buffer, on-policy distillation, and integration with a vLLM-based multimodal rollout backend.