Visual models¶
legopolitics separates specialized visual models from general vision-language models.
- Florence-2: captioning, region descriptions, grounding, detection-style tasks, and OCR.
- BLIP / BLIP-2: captioning and visual question answering.
- CLIP: image-text similarity, zero-shot labels, and agreement analysis.
- DINOv2: visual embeddings, clustering, representative crops, and re-identification support.
- Hugging Face image-text-to-text VLMs: open-ended frame, crop, and sequence interpretation.
Multiple Hugging Face models can be configured under vision.models.huggingface_vlms. Their raw outputs and provenance remain separate so the ensemble layer can compare model agreement rather than overwriting predictions. See Hugging Face VLM Ensemble.