JoyCaption Beta One (External LLaMA)

If you've captioned any sizeable dataset in the last couple of years, you already know JoyCaption: it's the open, uncensored vision-language model built specifically for writing training captions, and it's still the community favorite for it - our lora-training notes count 133 mentions in just the first half of 2026. What this node does is give you JoyCaption Beta One through a llama.cpp server you run yourself, with the full preset menu bolted on.

The "External LLaMA" in the display name is the whole architecture in two words: ComfyUI doesn't load the model. You run llama-server with the JoyCaption GGUF plus its vision projector, the node base64-encodes your image and POSTs it to http://127.0.0.1:58080/chat/completions, and a caption comes back. No key, no hosted API, no extra VRAM inside the graph. It's your GPU doing the work, and the model lives wherever your server is.

What the presets give you

The node ships the full JoyCaption Beta One CAPTION_TYPE_MAP, so one node covers a surprising amount of ground:

• Descriptive / Descriptive (Casual) - the classic natural-language captions for general training.
• Straightforward - concise, definite, no speculation; good for photoreal data.
• Stable Diffusion Prompt / MidJourney - caption shaped like an actual diffusion prompt.
• Danbooru / e621 / Booru-like tag lists - comma-separated tags with proper 1girl, artist:, character: prefixes. This is the one you reach for if you train on booru-tagged style data.
• Art Critic / Product Listing / Social Media Post - narrower, occasionally useful extras.

caption_length takes any through very long, or an exact number (20–260) for a word budget. And extra_options defaults to the famous JoyCaption instruction block - "refer to people as {name}", "include lighting, camera angle, watermark info", the works. Leave it as-is for stock behavior or trim it to taste. The third output, extra_options_examples, just hands you that block as a string so you can edit a copy.

The model and the server

The GGUF and mmproj come from concedo's conversion on Hugging Face:

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

Q8_0 of this model is essentially fp16 at half the size - if it fits, just use it. The --mmproj flag is non-negotiable; without the projector, the server can't see the image. And note this node's default URL is /chat/completions with no /v1/ prefix, unlike the Qwen3VL sibling in the same pack - if you move servers around, check that path.

The inputs you'll actually touch day to day are caption_type, caption_length, extra_options, and maybe max_tokens (default 512). temperature (0.6 default) and top_p (0.9) are fine where they are.

Install

Manager search "MiraSubPack", or:

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

Restart, and you're done - the pack carries no requirements.txt, and requests is already in ComfyUI. The real install is the llama.cpp server, which is on you.

Troubleshooting

• Error: text in the caption output. The node returns connection failures as caption strings. Check the server's up, the port matches, and the model loaded with --mmproj.
• Captions that ignore image content. Almost always a vision-context problem - bump --image-min-tokens and check -c 4096 isn't being squeezed.
• Slow. One request at a time, no batching, and each image waits. Batch your captioning overnight and walk away.

A word on expectations: it's a "test node" from the ComfyUI_Mira author's side pack, so it's functional but unpolished - no parallel requests, errors-as-text, and the default extra_options includes the stock NSFW-style option list, which is exactly what JoyCaption's model card ships. If you want the stripped-down version with just a custom prompt, the pack's JoyCaptionBetaOneSimple node is right next to it.