Grok 4.6

What it is

Grok 4.6, packaged as one node in KIE Next / LLM / xAI / Grok. Text in, text out - with one twist that changes what the node is for: it has an images socket. Connect an IMAGE and you're no longer looking at a chat model, you're looking at a VLM. The pack uploads your image, sends it alongside the prompt, and hands back a STRING.

That makes this the API-side answer to the captioning half of our KB's LLM-in-the-graph write-up: describe a shot so you can feed the description to img2img or image-to-video, build alt text, or build a prompt from a reference picture you didn't generate. If you're captioning a LoRA dataset, be honest with yourself about the tradeoff - a hosted model is convenient and filtered, and a local captioner like JoyCaption is what the dataset crowd actually uses for coverage on the material that gets auto-moderated off a commercial endpoint.

How it works

On run, the node uploads the connected image batch as PNG(s) to KIE's upload endpoint, takes the returned URLs, and builds a Responses-style request: prompt as text content, images as image content, system_prompt as instructions, history_json as earlier input items. Then it polls the task and extracts text from the response JSON.

Nothing runs locally beyond the upload and a JSON parse. That's the whole node: the visible widgets are just a friendly face on an HTTP call, and the schema was generated from KIE's published request documentation for this model version.

The knobs

• prompt (required) - what you want done with the image. "Describe this image in one paragraph of photographic language, no preamble" works a lot better than "what is this".
• images - the IMAGE socket. Accepts a batch; feed a whole folder through it and you're captioning at API speed.
• system_prompt - the standing instruction. For captioning, this is where you say what kind of caption you want: tags, natural language, one sentence, whatever your target model was trained on.
• reasoning_effort - low/medium/high/xhigh, default high. For "look at this and describe it", low is fine and cheaper; raise it when you're asking it to compare frames or spot inconsistencies between two shots.
• max_output_tokens - 4096 default. Long captions plus thinking both count.
• web_search - off by default. Irrelevant for captioning; relevant when you're asking about something current.
• history_json and expert_override_json - plumbing. [] and {} unless you know why you're changing them.

Outputs

text is what you forward - into CLIPTextEncode, a "save text" node, or another LLM node for a second pass. raw_json is the full response for debugging, usage_json is token accounting, and credits_consumed / credits_left tell you what the run cost and what's left, both of which are also printed on the node after a success.

One workflow shape worth stealing from the KB: caption first, then reshape. A VLM describes the image, a second LLM node (a different Grok version in this pack is fine) rewrites the description into the prompt format your generator wants. Two narrow jobs drift less than one open-ended "enhance my prompt" request, and it keeps your captioning and your prompting independent.

Install

Manager → KIE.ai Nodes Next, or git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next into ComfyUI/custom_nodes:

cd ComfyUI/custom_nodes
git clone https://github.com/felipederosilva/ComfyUI-KIE-Nodes-Next
python -m pip install -r ComfyUI-KIE-Nodes-Next/requirements.txt

requests, Pillow, numpy, PyYAML - that's the list, and there are no weights to download. Restart ComfyUI, then save your KIE.ai key once under Settings → KIE.ai Nodes Next → Connection (or set KIE_API_KEY on a headless box).

Note that the pack generates its node library from KIE's live docs index on first launch and re-checks it about once a day, occasionally reloading the frontend. If a node is missing, run the catalog refresh rather than reinstalling.

Gotchas

• Captioners bleed subjects. This is the failure every vision model shares, hosted or local: with two people in frame it'll happily swap who wears what. Audit the first batch by hand before trusting a set of 200.
• Your images leave the machine. They're uploaded, and moderation follows the provider's policy, not yours. For sensitive material, caption locally.
• High effort on a description task is the most common way to turn a cheap run into a slow, expensive one. Set low.
• Empty text almost always means the answer wasn't where the scraper looked - check raw_json before assuming the model refused.
• No key yet? The error is No KIE API key configured.... That's the Settings step, nothing more.