This Tiny AI Can Tag, Find, and Describe Anything in Your Photos

A team took the win in the MUMU Track of the 8th LSVOS Challenge with Efficient Unified Multimodal Understanding (EUMU), a single efficient model that jointly performs multi-concept image tagging, open-vocabulary object detection, and image captioning. It builds on a shared pretrained multimodal model, using that model's prompt-based capabilities for detection and captioning while training lightweight heads on shared visual features to predict quality, scene, and event tags. Rather than treating the three tasks independently, EUMU applies task-aware inference refinement by reusing task outputs as cross-task cues: caption cues help recover objects missed by the initial detection, detection cues help refine the caption to better reflect the detected objects, and image statistics refine quality predictions while caption and detection cues refine scene and event predictions. EUMU has 239.169M parameters, requires 23.947 GFLOPs, uses 4.5 GB of peak inference memory, and achieved a final challenge score of 17.3409.