☁️Mini FLUX image prompt retractor

Feed this an image and it hands back a prompt - the reverse of what you're normally doing in ComfyUI. It's a captioning node, but tuned specifically for producing FLUX-style tags rather than generic image descriptions, which makes it genuinely useful for the "I like this image, get me a variation of it" workflow: caption an existing image, feed that caption straight into a FLUX generation node, and you've got a starting point instead of writing a prompt from scratch by staring at a reference image.

Like Mini Long Text Error Corrector, this is one of the pack's "mini-party" nodes - self-contained, single-node utilities that configure their own LLM connection right on the node rather than pulling from a separate loader chain. Here, that LLM needs to be vision-capable, since the whole job is looking at an image and describing it, which is why the default model is gpt-4o-mini - a genuinely solid, inexpensive choice for image tagging that doesn't need a frontier-tier vision model to do well.

Inputs that matter. image is required - the picture you want tagged. model_name defaults to gpt-4o-mini; swap it for any vision-capable model your endpoint supports. base_url and api_key are optional and, same as the pack's other API nodes, need to point at a real OpenAI-compatible endpoint if you're not relying on config.ini defaults. The one that's easy to overlook is imgbb_api_key - and it matters more than it looks. A lot of OpenAI-format vision APIs expect an image URL, not raw image bytes, which means this node needs to get your image somewhere publicly fetchable before the vision model can see it at all. The README documents exactly this: the pack integrates with imgbb as an image-hosting service for precisely this reason, and it's the recommended option for any node in this pack that sends an image to a vision LLM. Leave imgbb_api_key blank and don't expect image uploads to work reliably - it's not a cosmetic optional field, it's load-bearing for how this node actually gets your image to the model. seed (default 42) is passed along for repeatability, though how much it actually pins down the output varies by provider - some honor it more strictly than others, so don't assume identical tags run to run even with the same seed. is_enable is the standard toggle. The single output, tags, is a STRING ready to feed straight into a FLUX prompt input.

Installing it. This ships with the full comfyui_LLM_party pack - no separate install. Get it via ComfyUI Manager (search "comfyui_LLM_party") or cd ComfyUI/custom_nodes && git clone https://github.com/heshengtao/comfyui_LLM_party, then pip install -r requirements.txt from inside the project folder using your ComfyUI Python environment, and restart. Get an imgbb API key (or set up the pack's other supported image host) if you plan to actually use this node, not just install it.

Troubleshooting. If tags comes back empty or errors, check imgbb_api_key first - a missing or invalid image-hosting key is the single most likely reason a vision call in this pack fails before it even reaches the model. Next, confirm base_url ends in /v1/ for any OpenAI-format endpoint, a recurring gotcha across every API node in this pack. And confirm whatever model_name you set actually supports vision input - plenty of cheap text models don't, and pointing this node at one will fail in a way that looks like a config problem but is really a model-capability mismatch. Separately, real users have reported the whole comfyui_LLM_party pack failing to import after a fresh install - a dependency conflict in the project's large bundled requirements.txt, unrelated to this specific node, with the README's requirements_fixed.txt as the documented fallback.