🎭 UEmo-D1 — Unified Multimodal Emotion Recognition

UEmo-D1 is a calibrated decision model built on Qwen/Qwen2.5-Omni-3B. Give it an instruction, a list of candidate answers and a target — text, image, audio, or video+audio — and it returns the chosen option plus its calibrated probability distribution over all options (a bidirectional-prefix / causal-answer attention pass over the option letters).

Combine modalities freely — e.g. an image and text in one call. Video examples carry no audio track, so matched silence is supplied.

Examples — click one to run it

Examples

Weights: UEmo-D1 · Code: wdqqdw/UEmo-D1 · Base: Qwen2.5-Omni-3B (Qwen Research License). Audio examples: Thorsten-Voice TV-44kHz-Full (CC0-1.0). Image/video examples: linoyts example assets (CC0-1.0).