Imagen Image Generation

Google runs two different image generation APIs, and they are not interchangeable. The ImageGen node in this pack talks to Nano Banana - Gemini's native image output, great at reference-image editing. This node talks to Imagen, which lives on a separate generate_images endpoint and behaves like a classic text-to-image model: prompt in, clean image out, with an explicit negative prompt and no reference-image slot. Imagen is Google's "photorealistic, no-fuss" option, and it's the one you reach for when you want a polished render and you don't need to edit anything.

The node offers the full Imagen 4 lineup: Ultra (highest quality, slowest), Standard (the default, balanced), and Fast (cheaper and quicker). Same prompt works across all three, which makes it easy to A/B your prompt before committing to the expensive pass.

How it works

It calls client.models.generate_images() with a GenerateImagesConfig - a different code path than the Nano Banana node's streaming generate_content. The response comes back as a batch of generated images, which the node decodes into a batched IMAGE tensor.

The inputs that matter:

• model - imagen-4.0-ultra-generate-001, imagen-4.0-generate-001, or imagen-4.0-fast-generate-001.
• prompt - describe the image.
• negative_prompt - things to avoid. Imagen actually honors this, unlike Gemini's image models.
• number_of_images - up to 4 in a single call; all come back as one batch.
• aspect_ratio - Imagen's narrower set: 1:1, 3:4, 4:3, 9:16, 16:9.
• seed - -1 for random; note this release honors seed=0 too (older versions silently dropped it).
• safety_filter_level - defaults to block_low_and_above. Here's the catch worth reading twice: the free Developer API tier often only accepts block_low_and_above; Vertex AI accepts all four levels. If you set a stricter option on the free tier, the API surfaces a clear error rather than quietly ignoring you.
• person_generation - whether/how to generate people (dont_allow, allow_adult, allow_all). Some models enforce stricter defaults regardless.

Output: images - a batched IMAGE tensor.

Installation

Part of the NanoBanana2 pack:

cd ComfyUI/custom_nodes
git clone https://github.com/IxMxAMAR/ComfyUI-NanoBanana2
pip install google-genai

ComfyUI Manager: search NanoBanana2. Needs a Gemini API key from aistudio.google.com - same key covers both endpoints.

Gotchas

The two-image-APIs confusion is the #1 footgun: drop an Imagen model ID into the Nano Banana node (or a Nano Banana model into this one) and you'll get a confusing 400. Keep Imagen models in this node, period. And the safety filter level is genuinely account-dependent - if you see an error about an unacceptable filter level, that's the free-tier/Vertex difference, not a bug. Finally, if a call comes back with no images at all, the pack raises a "no images returned" error that almost always means the prompt was blocked by safety filters - rephrase, and check your person_generation and filter level settings.