🍌 Gemini Banana 图片拼接 (最多30张)

Here's a node with a twist: it's in an API-heavy pack, but it's pure local image processing. No network, no key, no model. What it does is take up to 30 images and stitch them into a single numbered grid - cell 1, cell 2, cell 3 - and hand you a text guide saying which number sits where. Sounds trivial until you've tried pointing a Gemini/Nano-Banana image model at five separate reference images and asking it to "edit the second one." Multi-image models have a much easier time when they can say "change cell 7" than when they're guessing which of your five photos you meant.

That's the whole design intent: the pack's multi-image edit nodes (KenChenLLMGeminiBananaMultiImageEditNode etc.) work far better with a numbered collage as input, because the position guide gives the model a stable reference frame. Use this node as the front of that workflow.

How it works

It converts each input tensor to PIL, computes a grid layout (2×1 for two images, 2×2 up to four, 3×2, 3×3, … 6×5 for 26–30), scales each image into its cell, optionally draws a big number on top, and returns the composite plus a position_guide string like "1: top-left, 2: top-right…" written in a way the model can read. The layout always grows toward a square-ish shape, which models recognize better than a long strip.

The inputs that matter

• image1 … image30 - the optional IMAGE wires. Whatever you don't wire simply isn't in the grid.
• max_cell_size - each cell's maximum edge (default 512, range 256–1024). Bigger cells mean a higher-res collage but a heavier request to the API.
• resize_mode - keep_aspect_ratio (default, scales to fit within the cell), fit_to_cell (fills the cell, may distort), original_size (no upscaling of small images).
• add_numbers - on by default, and you want it on: the numbers are what let the model target cells.
• number_size, number_position, background_color - cosmetic; white/black/gray background fills the leftover space.

Outputs: collage_image (IMAGE) and position_guide (STRING). Feed both into a multi-image edit node - the collage into the image input, the guide into the prompt or a text input so the model knows the layout.

Install

ComfyUI Manager (search ComfyUI_LLM_Banana) or:

cd ComfyUI/custom_nodes
git clone https://github.com/xuchenxu168/ComfyUI_LLM_Banana
cd ComfyUI_LLM_Banana
pip install -r requirements.txt

It only needs Pillow/numpy, which the pack already pulls in. Restart, done.

Gotchas

Two practical ones. First, grid cell count and aspect ratio can make your subjects tiny - with 20 images at 512px cells you're effectively showing the model thumbnails, and detail-heavy edits will suffer; raise max_cell_size for important work (it'll cost more tokens/bandwidth). Second, the position_guide is a string you actually have to wire into the edit node for the numbers to mean anything to the model - forgetting it defeats the whole point.