Gemini Image (Symbiotica)

Gemini Image (Symbiotica) is an API-wrapper node that turns a prompt and up to fourteen reference images into an actual render using Google's Gemini image models - Nano Banana and friends. This is one of the closed models you literally cannot download, so an API node is the only door, and this is the category the KB calls out plainly: the model isn't on your machine, your prompt and references leave it, and you pay per call. If that's the trade you want (and for Nano Banana it's the only trade available), this node does it well - with one output that's genuinely unusual.

That output is thought_image. When the model thinks at HIGH level, it produces an interim sketch before the final image, and this node hands it to you on its own socket. That's gold for an order/revision workflow: you can show the client the rough before the final, or debug why a render went sideways. It only arrives with thinking_level HIGH and response_modalities on IMAGE+TEXT, so if you never see it, those two settings are why.

How it works

The prompt and reference images go to Google's Gemini image API. References are described by the prompt, not replaced by it - wire your references into the image slots and tell the model what to do with them in text. resolution runs through Gemini's own upscaler at 2K/4K, and aspect_ratio can match your references automatically (auto), which is the setting you want for edit workflows. The seed input isn't sent to Google - Gemini takes no seed - it exists so re-queueing re-runs the node instead of serving ComfyUI's cached output.

Billing follows the same two-path shape as the pack's Claude node: with the Cloudflare AI Gateway env vars set (SYMBIOTICA_AIG_BASE and friends) every call routes through the gateway on the studio's own key with spend tagging; anywhere else it calls Google directly with a key from the node, the Settings UI, or GEMINI_API_KEY / GOOGLE_API_KEY. Beginners: skip the gateway, paste a Google AI Studio key, render.

The inputs that matter

• prompt - what to draw. The default system prompt pushes hard for an image ("You must ALWAYS produce an image"), and the tooltip warns you: empty it and a conversational prompt may come back as prose instead of a picture. That's a real, documented failure mode.
• model - dynamic combo. The Lite model is the cheap lever for drafts and renders at 1K only. Pick Nano Banana 2 Lite and the node offers exactly one resolution, because that's all it renders - no phantom settings.
• seed - fake, as above; use it to bust the cache.
• response_modalities - IMAGE+TEXT (default) or IMAGE alone. IMAGE suppresses the model's commentary and the thought image, so if you want the sketch, leave this on IMAGE+TEXT.
• temperature / top_p - exposed at ComfyUI's own defaults. Lower temperature for more repeatable output.
• images - up to 14 reference slots, filled as you wire them in.

Outputs: image (the render), text (the model's commentary, when IMAGE+TEXT), and thought_image (the interim sketch, only under HIGH thinking + IMAGE+TEXT).

Install

ComfyUI Manager → search Symbiotica → install, or:

cd ComfyUI/custom_nodes
git clone https://github.com/symbiotica-ai/comfyui-nodes.git symbiotica
pip install -r symbiotica/requirements.txt

Restart ComfyUI. No downloads - it's a hosted call. You need a Google key; the pack loads without one.

Common issues

• No image comes out, just a refusal. When Gemini declines, that sentence is the error. Read text; the node doesn't hide refusals behind a fake success.
• No thought_image. Check thinking_level is HIGH and response_modalities is IMAGE+TEXT. Both are required; HIGH costs latency on every render, which is why the pack's own order pipeline runs MINIMAL.
• Prompt that should be a picture returns prose. Your system_prompt was emptied or overridden. Put the image-generation instruction back.
• Re-queueing returns the identical render. That's ComfyUI's cache doing its job; bump the seed control to force a real call.

The security note is the same one for every node of this type: it's arbitrary Python that holds your key and phones home by design, from a pack with essentially no community footprint yet. MIT-licensed and open - read it before you trust it, and remember the whole category was weaponized once.