LLaVA-1.6-Mistral-7B

Plain captioning models describe what's in an image. LLaVA-1.6 answers questions about it. Show it a render and ask "what's wrong with the lighting here" or "write a detailed prompt that could regenerate this," and it'll talk back. This node puts that on your canvas via Runware's cloud - no local weights, no Ollama, no VRAM budget.

How it works

It's a caption task in the Runware pack, but LLaVA is a full vision-language model rather than a one-line tagger. Your image gets uploaded, your prompt rides along as the instruction, and the model's answer comes back as a text output (STRING). The prompt field is optional - leave it empty and you get a plain description; fill it and you get a guided analysis. That's the input that turns this from a captioner into a workflow tool: "list every object and its approximate position" or "describe the material of the subject" are all fair game.

The 7B model runs on Runware's GPU cluster, which is the pack's whole deal - the "meta" in the class name is just Runware's catalog creator label, and it's a misnomer: LLaVA is the Haotian Liu / LLaVA project model, not a Meta release. Doesn't change how it behaves, but now you know.

A genuinely useful trick: chain this into a prompt-writing pass. Ask it for a detailed regeneration prompt, feed the text into a text-prompt input downstream, and you've got an automated describe-then-regenerate loop. The text output is a plain string, so anything that consumes STRING works.

Install

Part of ComfyUI-Runware, installed once:

cd ComfyUI/custom_nodes
git clone https://github.com/Runware/ComfyUI-Runware
pip install -r ComfyUI-Runware/requirements.txt

Restart ComfyUI, or ComfyUI Manager → search "Runware" → install. Dependencies are light and the model lives in the cloud, so hardware is a non-issue. You need a Runware API key: Settings → Runware API key, RUNWARE_API_KEY, or runware auth login.

Gotchas

Pay-per-run - the title bar shows cost after each call, and a chatty multi-turn exchange on a big image adds up. There's no conversation memory here: every execution is a single image + prompt round trip, so if you want back-and-forth, you structure it yourself across nodes. And while LLaVA-1.6 is a solid open VLM, it's not a frontier model - for hard visual reasoning it'll occasionally hallucinate confidently. Verify anything that matters.