StyleCap: Grounded Video Captioning
Docker agent for Track 2: Kimi K2.6 builds JSON fact sheets from video frames; GLM 5.2 writes formal, sarcastic, humorous-tech & humorous-non-tech captions. Same pipeline runs
Gemma-3-12B end-to-end on AMD MI300X via ROCm/vLLM (86s/8 clips).