gatilin/Youtu-VL-demo

Youtu-VL-4B-Instruct Multi-Task Demo

Interactive demo for Youtu-VL-4B-Instruct by Tencent Youtu Lab.

Supported Tasks (14)

• —Visual QA — Image description and question answering
• —Object Detection — Bounding box detection with labels
• —Visual Grounding — Text-to-region localization
• —Image Classification — Category recognition with attributes
• —Object Counting — Count specific object types
• —OCR — Text recognition and transcription
• —Semantic Segmentation — Pixel-level semantic labeling
• —Referring Segmentation — Language-guided segmentation
• —Depth Estimation — Monocular depth prediction
• —Pose Estimation — Human keypoint detection (16 MPII keypoints)
• —GUI Agent — UI understanding and interaction
• —Multimodal Reasoning — Logical analysis and reasoning
• —Multi-Image Understanding — Cross-image comparison
• —Custom — Free-form prompts

Architecture

Youtu-VL uses the VLUAS architecture with Multi-Latent Attention (MLA, DeepSeek-V2 style), supporting 2048x2048 normalized coordinate space and RLE-encoded dense predictions.

Notes

• —Model is loaded on first inference (~4GB download from HuggingFace Hub)
• —GPU runtime recommended for acceptable speed (4B parameters)
• —Visualization overlays are rendered on the input image after inference