OG Gemini Orchestrate

You know the mess this node is for. The background came from one generation, the subject from another, the jacket from a product photo. Compositing them properly is an afternoon of masking and light-matching. GeminiOrchestrate ("OG Gemini Orchestrate") hands all of them to Gemini with a sentence about how they fit, and gets back one photograph where the light agrees with itself.

Same underlying model as OG NanoBanana Generate - the Gemini 3.x image line, aka Nano Banana - different premise. NanoBanana is "here's a prompt, maybe some refs, make me a picture." This one is "here's the environment, here are the things that go in it." And that distinction is enforced by the node, not by your prompting discipline.

The mechanism is labels and size normalisation

The implementation is quietly clever, and it's worth knowing because it explains both where this node shines and where it fights you.

First, background is a required input, not another optional reference. The node builds its own text block around your prompt - naming image 1 as the background/environment and each following image as "element to include (reference N)" - then appends your prompt underneath. The model gets a labelled manifest, not a soup of pictures.

Second, every reference is resized to the background's exact pixel dimensions with Lanczos before upload. That's why the README says any resolution works: the normalisation happens on your machine. It also makes the background the reference frame for lighting, horizon and perspective.

The inputs that matter

background sets the scene and the scale. prompt is where you do the art direction - the author's tooltip says "be specific about positions and style," and it's right: "couple on the left third, red jackets prominent, low winter sun behind them" gets you somewhere; "couple skiing" doesn't. ref_image_1 through ref_image_5 are the elements, and order is meaning - the first reference is element 1 in the manifest.

model gives you two options, gemini-3.1-flash-image and gemini-3-pro-image - no lite model here, unlike NanoBanana. resolution (1K/2K/4K) is only transmitted on a gemini-3-pro model; on flash it's silently ignored, so if 4K matters, pick pro. temperature defaults to 0.5 - drop it to 0.3–0.4 while you're iterating on placement and want the layout to stop wandering. seed goes to the API but isn't a reproducibility guarantee.

aspect_ratio gets its own paragraph because it's required and independent of your background bitmap. It defaults to 16:9 and offers nine values (no 21:9). Connect a square background, leave the field at 16:9, and the model reframes - you asked for a wide picture, so it composes one rather than letterboxing your square. Match the field to the background's shape and the background becomes the frame it's meant to be.

Output is a single image: save it, wire it in as Veo's image for a first frame, or feed it to OG Omni Flash as an aesthetic reference.

Making it behave

Clean before you compose. The pack's own docs say each input should be free of unwanted objects and point at OG LaMa Refill to do it - that's the difference between a scene and a collage of other people's leftovers. Leave a ghost in the background and Gemini will faithfully composite around it.

Then talk like a photographer: where, how lit, what style. There's no mask, no depth pass, no ControlNet here. It respects your prompt's spatial language and roughly honours your references, and it will reinvent your background's geometry if you leave it latitude. If the background's exact architecture has to survive, that's a local ControlNet job.

Cost-wise, know which model you're on: Pro at high resolution is the expensive rung of the Gemini image line, flash is the cheap one. Draft placement on flash, re-run the keeper on pro.

Install

ComfyUI Manager → search OnlyGemini → Install Custom Nodes → restart. Or:

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

You need a Gemini key: export GEMINI_API_KEY="your-key" before launching ComfyUI, a .env inside the pack folder, or the api_key field. That field also takes the name of an env var, which is the tidier habit - a literal key in the widget ends up saved in your workflow file.

Troubleshooting

Gemini returned no image content with a safety rating attached. The filter - Google's image models are aggressively moderated, and the January 2026 update added strict limits on famous IP and celebrities, which is exactly what composition prompts reach for. The error names the category.

The scene doesn't look like my background. Check aspect_ratio against the background's shape first - a mismatch tells the model to reframe. Then check the prompt: references guide content, the prompt guides layout.

Subjects land in the wrong size or position. Reference order is element order, and the manifest calls everything after image 1 "an element to include" without saying where. That's your sentence's job.

API key required. Nothing resolved: no env var, no .env, empty field. Restart ComfyUI after exporting so the process actually sees it.