OneThingAI Gemini Image

Google's Gemini image models are the current frontier for prompt adherence and resolution - the "Nano Banana" family, rebranded from Gemini's image generation. Gemini 3 Pro Image outputs natively at 4K and reasons about the prompt before it generates; 3.1 Flash is the speed-quality hybrid. None of it runs on your machine. OnethingAIImageGemini ("OneThingAI Gemini Image") is the node that hands that family to ComfyUI through OneThing AI's API, and it's the node in this pack with the most model choices.

How it works. Same OpenAI-compatible client as the pack's other nodes: POST to OneThing AI's /v1/images/generations with your key, get base64 JSON back, and it becomes an IMAGE tensor ready to wire into Save Image or anything downstream. Wire a reference_image in and it switches to the image-edit endpoint for guided generation. The Gemini node differs from its siblings in two ways: no width/height inputs at all, and a model-gating system that only shows inputs a given model actually supports.

Inputs that matter. The model dropdown has seven entries: gemini-3.1-flash-image (and its -2k / -4k tiers), gemini-3-pro-image (and -2k / -4k), and gemini-2.5-flash-image - that last one is the original Nano Banana, still in the list for legacy workflows. Instead of image size, you get aspect_ratio with the usual suspects (1:1, 2:3, 3:2, 3:4, 4:3, 4:5, 5:4, 9:16, 16:9, 21:9) plus 默认 (default), which sends no ratio and lets the API pick. Then:

• seed - set it for reproducible runs; at -1 the node picks a random one. Note the gating: the seed input is only honored on the 3-pro and 3.1-flash lines. On gemini-2.5-flash-image it's silently dropped.
• retries / timeout - 3 and 120 by default; the 4K tiers can be slow, so expect to nudge timeout up.
• reference_image (optional) - flips the node into edit mode.
• extra (optional) - JSON string merged into the payload, your escape hatch for params the node doesn't expose.

Install. Standard for the pack:

cd ComfyUI/custom_nodes
git clone https://github.com/OneThingAI/ComfyUI_Onething_Image.git
pip install Pillow requests

Restart ComfyUI, or grab "ComfyUI OneThing AI Node" from ComfyUI Manager. No model downloads - the model runs on OneThing AI's servers.

Troubleshooting. Three things to know. First, every Gemini image carries Google's invisible SynthID watermark; that's the model, not the node, and you can't toggle it off. Second, Gemini's content filtering is strict and got stricter - the community's own verdict on ImageFX was "powerful, unimaginative, massive hidden biasing," and a January 2026 policy tightening added restrictions on celebrity and IP. If a prompt errors out, it's usually that wall. Third, remember the pack-wide quirk: the node re-runs on every queue, so each pass is a fresh billed generation. If you're on the 2.5 model and your seed seems to do nothing, that's the gating, not a bug - switch to a 3.x line if reproducibility matters.