T-Rex-Omni

model paper
IDEA Lab
Integrates negative visual prompts into generic object detection. Introduces unified visual prompt encoder for positive and negative prompts, Negating Negative Computing module for dynamic suppression, and NNH loss for discriminative margins. Achieves strong zero-shot performance across COCO, LVIS, ODinW, and Roboflow100, with 51.2 APr on LVIS-minival for long-tailed scenarios.
Paper (arXiv)GitHub

Outputs 2

T-Rex-Omni

model

Extends visual-prompted detection with negative visual prompts for suppressing hard negative distractors in open-set detection.

GitHub

T-Rex-Omni: Integrating Negative Visual Prompt in Generic Object Detection

paper

Proposes negative visual prompting with NNC module and NNH loss for discriminative open-set detection.

Paper (arXiv)

arXiv: 2511.08997

Venue: AAAI 2026

vision