Image Captioner (API: edit config.json)

The display name tells you the important part: (API: edit config.json). This node does not contain a model. It sends your image to an OpenAI-compatible endpoint you configure, and the endpoint is the whole story - specifically, it has to be a vision model, because a text-only LLM behind that URL will just get confused.

What it's for is the two jobs VLM captioning has always had: generating captions for a LoRA training set, and seeding a prompt from an image you already have for img2img or i2v. If you're captioning a big set, this is the automation; if you're seeding one prompt, it's a faster start than staring at the image and typing.

What you set

The image input is a dropdown of files in ComfyUI/input/, with an upload button - it is not an IMAGE socket. This is the first thing that trips people: you can't wire your freshly decoded latent into it. Save or upload the image to input/ first, then pick it. The node resolves the path and reads the file from disk.

instruction is the contract. The default asks for exactly one JSON object with a chinese key and an english key, and it's worth reading rather than skimming, because the node parses those keys back out of the reply: chinese becomes one output, english the other. If the model replies in prose instead of JSON, the raw text lands in the first output and the second is empty - a signal you need to tighten the instruction or drop the temperature.

max_side (default 1024) caps the longest side before upload. The image is re-encoded as JPEG, scaled to fit, and base64'd into a data URL, so this is your token/cost/VRAM dial. temperature (0.7) and max_tokens (1024) do what they look like. thinking_mode has three options - Ours (8-step) (default), Model native, Both - and it's a real behavioural switch: Model native ignores your instruction box entirely and uses a built-in minimal instruction so the model's own reasoning runs, with enable_thinking set true. Ours (8-step) uses your box with thinking off.

Outputs are three strings: chinese, english, thinking. Wire english into a CLIP Text Encode and you've turned a reference image into a prompt.

Setting the endpoint

Everything lives in config.json in the pack folder - there is no api_base widget and, notably, no environment variable is read anywhere in this pack, not OPENAI_API_KEY, not proxy vars. There's a per_node block, and the Captioner is the one that usually needs it, since your vision model is often on a different host than your text LLM:

{
  "api_base": "http://127.0.0.1:8080",
  "model": "",
  "per_node": {
    "FeiFeiImageCaptioner": { "api_base": "http://127.0.0.1:11434", "model": "qwen2.5vl" }
  }
}

Leave api_key empty for local llama.cpp. The file is re-read on every run, so no restart. Requests go through the openai SDK when it's installed and fall back to urllib when it isn't - a missing package degrades, it doesn't kill the node.

One honest note about this category. The most damaging security incident in this ecosystem's history was ComfyUI_LLMVISION, a captioner that shipped credential-stealing malware and ended in a federal prosecution; a vision node is tempting to weaponise precisely because it's expected to touch the network and load weights. That's a reason to read a pack before installing, and a reason the no-env-vars design here is the right call: the only credential that can travel is what you wrote in config.json.

Install

ComfyUI Manager → ComfyUI-FeiFei, or:

cd ComfyUI/custom_nodes
git clone https://github.com/aifeifei798/ComfyUI-FeiFei.git
pip install -r ComfyUI-FeiFei/requirements.txt   # openai>=1.40; optional, urllib fallback exists

No local model files - the model is on the other end of the URL.

Common issues

• "API Error: api_base is empty in config.json". Exactly what it says; the top-level value is empty and there's no per_node override.
• The text LLM answers with a description of nothing useful. You pointed a text-only model at an image payload. Use Qwen-VL, MiniCPM-V, or whatever vision model your server is serving, and set model accordingly on vLLM/Ollama-style servers.
• 400 on enable_thinking. Some servers reject the field; the pack retries once without it rather than failing.
• Long captions truncated. Raise max_tokens. Bigger max_side means a better description and a bigger request - the two costs move together.
• Two-subject images get their attributes crossed. Not this node's fault; it's the known failure mode of every VLM captioner. Audit your captions before training on them, especially at small dataset sizes.