🍌 Gemini Banana Multi Image Edit

The single-image editor is great, but the thing people actually want from a Gemini image model is multi-image work: unify five product shots onto white backgrounds, apply one style across a character sheet, composite elements from several photos. That's this node - "🍌 Gemini Banana Multi Image Edit," the official-API version. It takes up to four image inputs plus a prompt and returns one edited image, and it ships a post_generation_control switch for how consistent that output should be across runs.

Mechanically it's the same Gemini generateContent image-edit call as the single-image node, just with multiple inline images in the request. The parts of a multi-image request that matter - how the model interprets "these images" and which one it edits - live in the prompt, and the built-in default is Chinese ("请根据这些图片进行专业的图像编辑", "please perform professional image editing based on these images"), which is a hint about the pack's primary audience. Rewrite it to be specific about which image does what, and you'll get dramatically better results.

The inputs that matter

• api_key - official Google key (no mirror here; that's the sibling node).
• image1 … image4 (optional) - the image wires. Wire what you need.
• prompt / negative_prompt - the instructions. Be explicit about per-image roles.
• model - gemini-3-pro-image-preview (default) or the 2.5-flash image models.
• post_generation_control - randomize (default), maintain_consistency (keep results stable across runs), or enhance_creativity (looser, more variation). This plus seed/lock_seed is how you control reproducibility.
• aspect_ratio, response_modality, output_resolution (Auto/1K/2K/4K) - output geometry and format.
• upscale_factor / gigapixel_model - Topaz Gigapixel hook.
• quality, style, detail_level, camera_control, lighting_control, template_selection - the photography presets.
• temperature / top_p / top_k / max_output_tokens / seed - sampling. Note the seed max here is 999999, not the full 32-bit range of the single-image node.
• enable_iterative_refinement - session-based refinement (keeps keep_last_turns of context, default 3); reset_conversation clears it.
• Optional: custom_additions, safety_level (5 presets), system_instruction_preset (7), custom_system_instruction.

Outputs: edited_image (IMAGE) and response_text (STRING).

Install

ComfyUI Manager → 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

Restart. Needs google-genai>=1.51.0.

Gotchas

Multi-image calls are where attribution errors bite hardest - the same multi-subject weakness that plagues every VLM (the character-consistency and llm-in-comfyui KB docs). If you ask for one style across several photos, audit what actually happened before trusting it. And each call costs like a multi-image Pro request, which is more than single-image - batch deliberately, and use maintain_consistency + a fixed seed when you're iterating toward one result you actually want to keep.