// the find
kijai/ComfyUI-Florence2
Inference Microsoft Florence2 VLM
A ComfyUI custom node that wraps Microsoft's Florence-2 vision-language model, letting you run captioning, object detection, segmentation, and DocVQA as nodes in a ComfyUI workflow. Useful if you're already deep in the ComfyUI ecosystem and want VLM capabilities without leaving it.
Auto-downloads model weights from HuggingFace directly into ComfyUI's models directory, so setup is one node drag. Supports the full Florence-2 family including fine-tuned variants like the SD3/Flux captioners, which are practically useful for generating training captions. Ships its own model implementation files rather than relying entirely on transformers internals, giving you more control over inference behavior. DocVQA support via HuggingFaceM4's fine-tune is a genuine addition over the stock model.
The repo bundles custom model implementation files (davit.py, model.py, processing.py) that duplicate or diverge from the upstream transformers implementation — when Microsoft updates Florence-2, this fork will quietly fall behind and you'll get subtle inference differences with no warning. No tests whatsoever. The README is sparse; there's no explanation of what each node outputs or how to wire them together beyond two screenshots. Last meaningful activity was May 2026 but Florence-2 itself peaked in 2024 — newer VLMs (Qwen2.5-VL, InternVL2) have largely overtaken it on every benchmark, so you're adopting a model that's already dated.