Expanded Gemini Text/Image

This is the node that actually does the work. The pack's other two nodes are bookkeeping - SSL_GeminiAPIKeyConfig builds your credentials and GetKeyAPI feeds it a key from a file. SSL_GeminiTextPrompt ("Expanded Gemini Text/Image") is where your prompt leaves your machine, hits Google's servers, and comes back as text or an image. It's a cloud-API node, not a local diffusion model: nothing downloads, nothing runs on your GPU, and every run costs tokens billed to your API key.

The core trick is that one node covers three jobs. Text-to-text is the default: connect your config, type a prompt, get text. Image understanding works by also wiring an image into input_image (or input_image_2 for a second one) - Gemini reads it alongside your prompt. And text-to-image is just a matter of switching the model combo to one of the image-capable models, like gemini-2.5-flash-image or the nano-banana lineup (nano-banana-pro-preview). The KB's closed-source-models essay has the full story on those; they're Google's hosted image generators with real text rendering, and they're excellent at following long, specific prompts.

The inputs you'll actually touch, most days:

• config - required, from the Configure Gemini API Key node. Everything else is a setting.
• prompt and system_instruction - your message and the system prompt (default "You are a helpful AI assistant."). Set the system instruction to "you are an image editing assistant" for editing workflows.
• model - a big combo spanning Gemini 1.5 through 3.x, plus the image models. Default is gemini-2.5-flash, a good cost/speed balance.
• temperature, top_p, top_k, max_output_tokens - the standard sampling controls. Crank temperature up for creative text, leave it low for editing tasks.
• thinking_budget - Gemini's reasoning toggle. 0 disables thinking, -1 turns on dynamic thinking, any positive number sets a fixed token budget. The related thinking_level combo (low/medium/high) only works on the newest models and takes precedence when set.

Three outputs come out: text (the model's reply), image (an IMAGE tensor), and final_actual_seed (the seed that was really used). That image output is a trap worth knowing: when the model returns text only, the node emits a small blank dark placeholder tensor instead of nothing, so a downstream preview node shows a black box that looks broken but isn't. The final_actual_seed matters because Gemini lets you pass a seed, and this node exposes it so you can lock in a result you like - pass the same seed back and you get the same image.

There's real engineering under the hood beyond the obvious API call. The node fingerprints every input - including a hash of your input images - and caches results, so re-running an identical workflow returns the cached answer instead of burning another API call. That's great for iteration and mildly surprising the first time you tweak something and the output doesn't change; turn use_seed off or change any input to bust the cache. There's also a regex retry system: put a pattern in retry_pattern, and if the response matches (say, a refusal or a known bad token), the node re-rolls the seed and tries again up to max_retries times. And the proxy settings (use_proxy, proxy_host, proxy_port, defaulting to 127.0.0.1:7890) exist because the README was written with China-based users in mind.

Install is the same as the rest of the pack:

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

Restart, or install via ComfyUI Manager by searching "ComfyUI Gemini Expanded API". The only new dependency that matters is google-genai.

The troubleshooting list, from the README and the source: the specific error [ERROR]API call error: 'NoneType' object has no attribute 'parts' means your prompt or image tripped Google's "Generative AI Prohibited Use Policy" - the request went out and got blocked, so test with a plain product or scene image before assuming your workflow is broken. A failed call with no response usually means the API key is wrong, the network can't reach Google (check your proxy), or you hit the default 30-second timeout on a slow generation - raise timeout up to 300 for image models. One caveat: Google's filters and rate limits change out from under you, and this node can't fix a policy block. It can only make the API easy to hit, which it genuinely does.