MLLMSeg: Referring Expression Segmentation with Multimodal LLMs

Upload (or paste) an image and describe the object or region you want to segment. MLLMSeg predicts a pixel-level mask for the referred target using an InternVL3.5 backbone and a lightweight mask decoder.

This demo runs MLLMSeg_InternVL3_5_1B_RES. Paper · Code (MIT).