Claude Vision Analysis

You know that part of a workflow where you'd kill for someone to actually look at the image and tell you what's wrong with it? Claude Vision Analysis is that someone. It sends a ComfyUI IMAGE tensor to Claude and returns a text analysis - captioning renders, checking anatomy, describing a reference, grading lighting. The one-sentence version: feed it an image and a question, get a STRING back.

The pack frames this as its multimodal workhorse, and the headline number is 20 images at once: one primary image plus up to 19 more on additional_images, all analyzed in a single call. That makes it useful for batch review - generate eight variations, dump them in, and ask "which of these has consistent anatomy?"

How it works

The node converts the tensor to what the Anthropic API expects, sends it with your question, and returns the model's text. The model override is where it gets interesting: the client normally fixes the model (say, Sonnet 4.6), but this node lets you bump a single vision call up to Opus 4.7, which supports 2576px image resolution versus 1568px on older models. If you're analyzing fine details - text on a poster, small defects - that resolution headroom is real, not marketing.

Inputs you'll actually set

• image - required. The primary tensor. Any Load Image / generated output works.
• question - required. Defaults to "Describe this image in detail." Be specific: "List every error in this face" beats "what's wrong?"
• detail_level - high by default; low is cheaper and faster for simple checks.
• max_tokens - 2048 default. Verbose answers get truncated; raise it if the analysis cuts off.
• model - optional override; default is inherit-from-client.
• additional_images - the batch input, up to 19 extra.
• seed - cache control only.

The single output, analysis, is a STRING - wire it into Preview Anything to read it, or back into another node's prompt input for a self-correcting loop.

Install and gotchas

Part of the ERPK Collection: ComfyUI Manager → search erpk → install, or

cd ComfyUI/custom_nodes
git clone https://github.com/eRepublik-Labs/comfyui-nodes-erpk.git erpk
cd erpk && pip install -r requirements.txt

Restart, then put an Anthropic key in Settings > ERPK > API Keys (right-click canvas → ERPK Settings). No local models - every analysis is a paid API call, and 20 images in one request still bills as one conversation with a big token bill, so the cost scales with detail.

The thing that catches people: that additional_images input isn't a batch-friendly "drop a whole folder" socket - you still connect a tensor batch, and the node treats it as the extra set. And remember the model override is per-call: if your analysis looks capped at lower resolution, that's the Opus 4.7 knob you forgot to flip.