A novel unified segmentation MLLM that extends any-segmentation from images to videos, supporting conversational instructions and visual prompts through Mask Memory for temporally consistent pixel-level perception.

A novel unified multimodal large language model (MLLM) framework, which extends the segmentation from segment anything to any segmentation, enhancing pixel-level perceptual understanding.

A novel unified open-vocabulary detection method, which is pre-trained on diverse large-scale datasets with language-aware selective fusion in a unified framework.

A novel self-attention and temporal memory mechanism to capture long-range temporal relations between frames, avoiding the computational cost of optical flow prediction.

A framework that fuses OCR and ASR information using a Transformer model with mask/crop strategies and multi-level identity embeddings to generate comprehensive video subtitles.