vLLM-Omni adds text-to-image and reference-image editing

vLLM says vLLM-Omni now supports text-to-image generation and editing from a reference image.

The update adds multi-GPU VAE tiling, CUDA Graph replay and a faster image-response path, according to the project's X post on November 10, 2026. The post is the only report so far, and it gives no benchmark figures for the speed changes.