Vision as Unified Multimodal Generation

This paper frames computer vision as unified multimodal generation. SenseNova-Vision uses natural-language instructions and optional visual prompts to define tasks, regions, views, and decoding conventions. It emits text for symbolic outputs, images for dense spatial predictions, or mixed text-and-image responses for compositional tasks. The model starts from an off-the-shelf pretrained unified multimodal model and is trained mainly on the SenseNova-Vision Corpus, an instruction-response corpus converted from diverse vision annotations. It requires no task-specific prediction heads or architectural changes, and reportedly covers detection, OCR, keypoints, segmentation, depth, surface normals, point maps, and camera pose estimation.