MinusZone - Captioner(Florence-2)

Captioner(Florence-2) is the batch-mode sibling of CLIPTextEncode(Florence-2). It takes a directory instead of an image, walks every .jpg/.jpeg/.png in it, captions each with Florence-2, and writes the captions as sidecar files (image.png.caption by default) right next to the images. It's a purpose-built dataset-captioning tool, and it's the node in this pack you'd actually run overnight.

The KB's lora-training guidance is the whole rationale: Florence 2 is a community-recommended captioner for natural-language training captions, and captioning quality matters more than dataset size. This node turns that recommendation into a one-shot batch job with a progress bar in the ComfyUI UI.

The inputs that matter

• model_name - Florence-2-large-ft (default) or Florence-2-large. Use the -ft for better instruction following.
• directory - the folder to scan, recursively. This is the one that actually does the work.
• resolution - default 512, resize target before captioning.
• batch_size - default 1. How many images to caption per model forward pass. You can raise it if VRAM allows; 1 is the safe setting.
• caption_suffix - default .caption. Change to .txt if your trainer expects that.
• force_update - default False. Skip images that already have a caption file; flip on to re-caption everything (useful after changing the model or prompt).
• prompt_fixed_beginning - string prepended to every caption, e.g. a trigger word.

Output is a single debug STRING - the batch runner isn't built to return captions as tensors, so the real output is the files on disk.

How it works

Internally it wraps the Florence-2 encode path in the pack's shared batch captioner: walk the directory, skip already-captioned images unless force_update, convert each image to a tensor, run the model (kept resident across the batch), write <image><suffix>, and show progress. If a generation comes back empty, it retries with a bumped seed up to 5 times - a sane guard against the occasional empty output.

Installing

Part of ComfyUI-Prompt-MZ:

cd ComfyUI/custom_nodes
git clone https://github.com/MinusZoneAI/ComfyUI-Prompt-MZ

Restart. Needs transformers + torch; the Florence-2 model downloads into ComfyUI/models/minus_zone_models on first run (multi-GB). llama-cpp-python from the pack's requirements isn't exercised by this node but installs with the pack anyway.

Troubleshooting

The realistic complaints: (1) It's slow - Florence-2-large on CPU can take several seconds per image, so a few hundred images is an all-afternoon job. Set batch_size up if your GPU has room, and consider the smaller Florence-2-large-ft's sibling models from elsewhere if speed is the blocker. (2) The captions are natural-language sentences, which is what you want for Flux-style bases and wrong for Danbooru-tag bases like Illustrious - pick your trainer before trusting the output. (3) OOM mid-batch: keep batch_size at 1 and note this node's batch mode keeps the model loaded, so VRAM stays busy the whole run - close the sampler's windows or run captioning as a separate step from generation. (4) If a run "completes" with no files, check directory - a typo path silently scans nothing.