APNext Groq Vision Analyzer

If you're batch-captioning images and want it fast and cheap, this is the one. GroqVisionNode (APNext Groq Vision Analyzer) sends your images to Groq's API - running Meta's Llama 4 vision models on Groq's inference hardware - and returns a text description you feed into a generation. Same job as the GPT and Grok vision nodes in this pack; the draw here is speed and a generous free tier, so it's my default when I'm captioning a folder rather than a single hero image.

It's the vision half of the pack's whole premise: modern image models read prose, so a caption is a prompt, and a multimodal LLM writes that caption in a fraction of a second. dagthomas ships one vision node per provider so you can pick whichever backend you have credits (or free quota) for.

How it works

It's an API client - nothing on your GPU. Image up, description back. The groq_model dropdown lists Groq's vision options, currently the Llama 4 Scout and Maverick instruct models. Scout is the lighter, faster one; Maverick is bigger. For prompt captioning, Scout is usually plenty.

Shared style controls, same as the rest of the pack:

• happy_talk (on) - enthusiastic vs. plain tone.
• compress + compression_level (soft / medium / hard) - shorten the caption. Worth it; tight beats verbose for prompting.
• poster - poster formatting, off by default.

Optional custom_base_prompt / override replace the built-in analysis instruction. One STRING output into your positive text encode. Like the Grok vision node - and unlike GptVisionCloner - there's no image blending here; it's straight image-to-caption.

Installing it

ComfyUI Manager: search comfyui_dagthomas, install, restart. Manually:

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

The install pulls the pack's full dependency set regardless of which node you want. Then export GROQ_API_KEY before launching ComfyUI.

Common issues

Groq vs Grok, again. Groq (q) is this node - fast open-model inference, key GROQ_API_KEY. Grok (k) is xAI's separate node. Mixing the keys up is the single most common failure.

Rate limits on batches. The free tier caps requests and tokens per minute. Captioning a big folder will hit that wall - throttle the batch, or accept the occasional retry. Scout uses fewer tokens than Maverick, so it also survives the caps better.

Model retired. Groq rotates hosted models regularly. If a vision model errors, pick a current one from the dropdown or update the pack's groq_models.json and restart; the node prefers the live API list and falls back to that file.