Microsoft’s Florence‑2 vision foundation model is empowered by FLD‑5B, housing 126 M images and 5.4 B annotations.
Florence‑2 unlocks image‑level, region‑level, and pixel‑level recognition by jointly handling spatial hierarchy and semantic granularity.
The model adopts a unified architecture that integrates spatial, temporal, and multi‑modal facets seamlessly.
Meta introduced ImageBind on 2023‑11‑01 15:30 UTC, a game‑changing multimodal model that connects six diverse data modalities – text, audio, 3D measurements, motion data, images and video – without requiring any paired data, by learning a single embedding space that aligns all modalities.
I announced that a “text‑to‑anything” model will soon be a reality (timestamp 2024‑02‑16 08:00 UTC).
A new multimodal model—referred to as a “new multi modal kid”—entered the Gen AI ecosystem, indicating a fresh capability for handling multiple data modalities.
New multi modal kid in the Gen AI town!
Google introduced Synth2, which creates synthetic image‑text pairs by first prompting an LLM to generate captions and then feeding those captions into a text‑to‑image model to produce matching images.
The Synth2 pipeline “enhances VLM performance, improves data efficiency, and scales dataset creation”, especially where manual labeling is costly.
- Caption Generation: prompt LLM → caption
- Image Synthesis: caption → text‑to‑image model → image
- Pair Assembly: combine caption+image
- VLM Training: use synthetic pairs alone or mixed with real data