JoyCaption Beta One Simple (External LLaMA)

The big JoyCaptionNodeBetaOne node in this pack gives you a preset menu, a length picker, a giant options block, three outputs. This one gives you a single prompt text box. That's the whole difference - JoyCaption Beta One Simple is the same model, the same external llama.cpp server, the same base64-encode-and-POST mechanism, just without the decision fatigue.

Reach for it when you know exactly what you want to say. Maybe you want captions in a very specific format for a niche LoRA, or a fixed instruction that the preset list doesn't express ("describe the outfit, ignore the background, one sentence"). Drop that in prompt - default is Write a detailed description for this image. - and every image gets captioned with exactly your wording, word for word. That predictability is genuinely useful for training, where the caption style is part of the recipe and you want zero drift between runs.

The inputs

• prompt - your custom captioning instruction. This is the whole point; the node sends it verbatim.
• temperature - 0.6 default, same as the full node. Lower for more deterministic captions.
• max_tokens - 512 default; raise it if your custom prompt asks for long output and you're hitting the ceiling.
• llama_url - http://127.0.0.1:58080/chat/completions, same server and path as its sibling. Point it elsewhere if your server moved.

One output: caption, a single STRING you can feed straight into a text file, a save node, or a tagging workflow.

The server setup is identical

There's no separate install. Same llama.cpp server as the full JoyCaption node:

llama-server.exe -m "llama-joycaption-beta-one-hf-llava.Q8_0.gguf" \
  --mmproj "llama-joycaption-beta-one-llava-mmproj-model-f16.gguf" \
  -ngl 33 -c 4096 --port 58080 --image-min-tokens 2048

The GGUF + mmproj pair is concedo's JoyCaption Beta One conversion on Hugging Face. The --mmproj flag is what lets the server see images at all. Install the pack via Manager (search "MiraSubPack") or:

cd ComfyUI/custom_nodes
git clone https://github.com/mirabarukaso/ComfyUI_MiraSubPack

Restart, and you're running. The pack has no requirements.txt; this node only needs requests, which ComfyUI already ships.

When NOT to use it

If you're captioning a big varied dataset and want tags, SD-prompt style, or the full JoyCaption option block, the preset versions in the full node are much less typing - the Simple node assumes you can write a better prompt than the menu can. And keep in mind the same gotchas: one image at a time, no batching, and connection problems come back as Error: strings in the caption output rather than clean exceptions. If your caption comes back as "Error: Connection failed: ...", the server isn't reachable - check port 58080 before you suspect the node.

For a beginner, honestly, this might be the better first stop: install once, paste one prompt, and you've learned exactly how the pack's external-server captioning works before you go anywhere near the preset circus.