Load Florence2 Model

Let's get the elephant out of the room first, because it's the first thing you'll see: the README for this pack is one line - "Use https://github.com/kijai/ComfyUI-Florence2/ instead." The author, spacepxl, retired the pack in July 2025 and points everyone to Kijai's maintained fork. If you're starting from zero, you genuinely should install Kijai's instead. But this loader still works, it's the node behind the popular Civitai "Image to Text Prompt (Florence2)" LoRA-captioning workflow, and its model cache is shared with Kijai's pack, so old workflows keep running fine. Knowing that up front saves you a confused afternoon.

What the node does is simple: it loads Microsoft's Florence-2 - a small vision-language model that captions, detects objects, and segments - into ComfyUI. Florence-2 has been a quiet workhorse of the ecosystem since mid-2024: it's the default captioner in Fluxgym, and the standard "good enough" alternative to JoyCaption when you're building LoRA training sets (the KB's LoRA training essay rates it as the tool to reach for when you want natural-language captions for Flux).

How it works under the hood is worth understanding because it explains two behaviors that confuse people. The loader has a version dropdown with four choices - base, base-ft, large, large-ft - and on first load it downloads the weights and the processor from HuggingFace (microsoft/Florence-2-*) into ComfyUI/models/LLM/, then keeps them cached in the node instance. That cache is why re-runs are instant and why changing the version mid-session triggers a fresh download and a ~30-second reload. The code also patches out the flash_attn import from Florence-2's modeling file, which is the good news: you don't have to fight FlashAttention to get this running.

The inputs are the whole story, so here they are:

• version - the only input. base (~0.23B params) is fast and plenty for captioning. large (~0.77B) is slower, more VRAM, marginally smarter. The -ft suffix is the fine-tuned variant, and base-ft is the one most people actually run.

The single output, FLORENCE2, is a bundle (model, processor, version, device) that wires straight into the pack's Florence2 node - it's not an image or a string, so don't expect to preview it.

Install is the standard two-step: ComfyUI Manager → search "ComfyUI-Florence-2" → Install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/spacepxl/ComfyUI-Florence-2

One real gotcha: this pack has no requirements.txt. It needs transformers and matplotlib from Python, and nothing installs them for you. If you hit ModuleNotFoundError: No module named 'transformers', that's the fix - pip install transformers matplotlib.

Where people get burned: first load needs internet (the download can silently stall on a flaky connection - retry, and the partial files live in models/LLM/), and the model stays in VRAM once loaded, so large plus a checkpoint on a 6–8GB card is how you get OOMs. But the single most useful thing about this loader is that it uses the models/LLM/ folder - so when you eventually migrate to Kijai's fork, it reuses every download. No double-dipping.