🅐 Image Question Answering

"Describe what's in this image," asked of a model that runs on your machine, with no API key. That's ACE_ImageQA - a vision-language QA node that takes an image, a question, and a model choice, and returns a text answer. It's the "auto-caption my generations" node, or the "is there a hat in this?" checker, or the first half of an automated feedback loop that re-prompts based on what the model saw.

It's part of 🅐 Ace Nodes (hay86/ComfyUI_AceNodes), the usual one-author grab-bag - and it's one of the heavier nodes in it, because it runs real multimodal models locally.

How it works

You pick a model, both real open-weight vision-language models loaded through transformers:

• moondream2 (default) - a ~1.8B lightweight VLM from vikhyatk/moondream2. Small, fast, runs on modest hardware, good enough for general captioning and simple questions. This is the one to start with.
• MiniCPM-V-2 - OpenBMB's older 8B vision-language model. Notice the name: this is MiniCPM-V 2, not the newer 2.6. It's a substantially bigger download and heavier at inference, with better answers - the kind of jump you make when moondream2's summary is too shallow.

The first use of either model downloads it into ComfyUI/models/prompt_generator/. moondream2 is a few hundred MB; MiniCPM-V-2 is multiple GB and will look frozen while it pulls. Both are cached after that.

Inputs: image (a single IMAGE), text (your question, multiline), and model. Output is STRING - the answer. One image at a time; don't feed it a big batch and expect per-frame answers.

What people actually hit

• First-run downloads are the whole setup cost. Budget for it, especially if you pick MiniCPM-V-2 sight unseen on a small drive.
• It's a QA node, not a captioner-in-a-box. You still have to ask a good question. "Describe this image" works; "count the people" is where small VLMs start hallucinating - verify answers before you build automation on them.
• Device handling is automatic (CUDA → MPS → CPU), with bf16 on capable cards, but an 8B model on CPU will be painfully slow. moondream2 is the realistic choice on anything without a GPU.

Installing

ComfyUI Manager → search ComfyUI_AceNodes, or:

cd ComfyUI/custom_nodes
git clone https://github.com/hay86/ComfyUI_AceNodes

Then restart. This node needs transformers and huggingface_hub from the pack's requirements.txt - which also pulls the usual heavy tail (rembg, insightface, soundfile, openai, …). The model downloads are separate from the pack install and land in models/prompt_generator.

The honest comparison: there are dedicated captioning/QA packs with more models and fancier pipelines. What this one gets you is a competent, local, single-node "ask the image a question" with zero setup beyond the download. If your use is occasional auto-captioning, that's plenty. If you're building a serious eval or captioning pipeline, you'll outgrow it - but it's a genuinely decent place to start.