Gemini Vision

Claude has its vision node in this pack; this is Gemini's answer. Feed it an IMAGE tensor and a prompt, and it returns a text analysis - captioning, defect-checking, describing a reference image, evaluating a render. The batch-friendly detail: the image input takes a tensor batch, so you can dump multiple frames into one call and ask a question across all of them.

Mechanically it's the same shape as Gemini Text Generation - same model dropdown (default gemini-3.5-flash), same sampling knobs, same response_mime_type/response_schema JSON-mode pair, same thinking_level translation between 2.5 and 3.x models - but with an image input instead of a text prompt as the required field. The prompt is still required too, defaulting to "Describe this image in detail."

Inputs you'll actually set

• image - required tensor(s).
• prompt - required. Ask specifically: "list the anatomical errors in this render" beats "what do you see?"
• model - default gemini-3.5-flash; bump to a Pro model for harder analysis. Cheap for flash, so start there.
• temperature - 0.4 default. Lower = more factual; the tooltip says it straight.
• max_tokens - 8192 default, plenty for a thorough answer.
• top_p / top_k / stop_sequences / response_mime_type / response_schema / thinking_level - same semantics as the text node; JSON mode works here too if you want a structured analysis object.
• seed - cache control, -1 randomizes.

Output: analysis, a single STRING. Wire to Preview Anything, or - the power move - back into another node's prompt input to build a self-correcting loop (generate, critique, regenerate).

Install and gotchas

Part of the ERPK Collection:

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

Restart (or ComfyUI Manager → search erpk), add a Google AI Studio key in Settings > ERPK > API Keys. Paid per call, but Flash vision is cheap.

Where people stumble: forgetting that the image input is a batch socket and treating it like a single image picker (it handles both, but a batch of N sends all N in one call, which bills accordingly). And if analysis keeps coming back shallow, it's almost always the temperature or the prompt specificity - 0.4 is not 0, and "describe" invites a paragraph instead of a checklist. For consistent, comparable output across runs, fix the seed.