verl-omni

Why we included this project

VeRL-Omni is the project verl spun out of its multimodal work into its own home for reinforcement learning training of diffusion and omni-modality models, and it stays built on top of verl and vLLM. Where verl mainly trains autoregressive text reasoning, this framework handles models that emit whole outputs, from text-to-image and text-to-video to audio-video generation, with support for Qwen-Image, Qwen3-Omni, and LTX. It ships trainers for GRPO, DPO, GSPO, and FlowDPPO, and one optimization worth knowing about scores clean diffusion latents to skip VAE decoding, which cuts latency and memory on multimodal runs. Teams already on verl or vLLM get a single stack for generative-model alignment instead of hand-rolling denoising rollouts and VLM-as-judge reward scoring.