🎭 Gemini Pro Batch Refine

The usual failure mode of a good product render isn't the whole image - it's one logo that came out as mush, a label with a hallucinated letter, stitching that reads as static. Gemini Pro Batch Refine is the node for exactly that job: it takes a batch of images and re-renders each one with the details corrected against reference shots, while leaving everything else pixel-identical. It's the "refine" stage of a detect → crop → refine → stitch pipeline, and it's the most niche thing in this pack - built for e-commerce garment work, useful anywhere a close-up photo needs forensic detail repair.

It's one of the API nodes, so it needs a Google API key and it costs money per call. This is Nano Banana Pro (Gemini 3 Pro Image) under the hood, and you feel it in the pricing and in the output quality - the default system prompt is a tightly-worded "identity lock" contract that orders the model to keep the crop, framing, lighting, and fabric texture exactly as-is and change only the specific wrong detail.

How it works

The node is list-aware (INPUT_IS_LIST), so every image you feed in is refined separately against the same shared reference shots. Each iteration becomes its own API call: {image_1} is the per-iteration base crop, and {image_2} through {image_12} are optional reference images - the close-ups of the correct logo, label, or hardware that the model is told to copy "like a forensic document."

Two behaviors matter before you run it. The seed is per-batch: 0 means random each iteration, but any non-zero seed produces seed + i for the i-th item, so a fixed seed gives you reproducible refinements across the whole batch. And stop_on_error defaults to off - a failed iteration inserts a tiny (1,1,1,3) placeholder and keeps going, which sounds weird but exists to preserve index alignment with the alphas_list from Gemini Mask Crop downstream. If you'd rather the whole run abort on the first failure, flip it on.

Inputs that matter

• images - wire this from Gemini Mask Crop's images_list (native-size list) or any IMAGE/batch source. Each entry is refined separately.
• system_prompt - the default is the garment identity-lock prompt. It's editable and replaces fine - the {image_1}/{image_2} placeholders are the contract.
• image_2 … image_12 - the reference shots. They're optional but pointless to skip; the whole node exists to copy details from them.
• aspect_ratio - leave at custom so the model keeps the base image's framing, which is what you want for refinement.
• image_size - 1K / 2K / 4K resolution tier. 2K is the sane default; 4K is where the bill climbs.

Outputs are refined_images (a list, one per input), a summary string, and a metadata dict with per-iteration latency and success info. Feed the refined list straight into Gemini Mask Stitch to put the corrections back on the original.

Install and the key

Install via ComfyUI Manager (search "ComfyUI_Gemini") or clone into custom_nodes:

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

The requirements pull in google-genai - not google-generativeai, which is the classic gotcha (the pack even auto-uninstalls the wrong one at startup). For the key, put GOOGLE_API_KEY=... in the .env file the pack creates in its own folder and leave the node's key field empty, so your shared workflow JSON never embeds the secret.

Where people get burned

Cost sneaks up fast: a batch of ten crops with several reference shots is ten separate paid generations. Watch the *-image-preview model names - Google retires preview variants, so if one errors out, move to the GA name (gemini-3-pro-image). Safety filters can silently block, and if you get "API key is missing" check that the .env value isn't still the placeholder. This pack is a small, single-author project - 14 threads by name in the community corpus, nothing by author name - so don't expect a support army; the console output is actually detailed and worth reading when something fails.