Goldfish / MiniGPT4-video

Goldfish (with its MiniGPT4-video component) introduces a retrieval-based approach to understanding arbitrarily long videos: rather than feeding an entire clip to the model, it retrieves the most relevant segments per query, avoiding context-length blowup. The 636★ release, presented at ECCV 2024, is the origin of the Mohamed Elhoseiny / Vision-CAIR long-video line that culminates in Tempo.

Built on the MiniGPT-4 lineage and a pretrained LLM backbone rather than trained from scratch, so its scale is inherited from those bases.