🍌 Gemini Banana Image Edit

This is the node the pack is built around. "🍌 Gemini Banana Image Edit" is a Google-official-API image-to-image editor: wire in an image, type "add a llama next to me," and the Gemini 3 Pro Image model (the Nano Banana Pro engine, per the closed-source-models KB doc) sends back an edited image, fully in-context with your graph. This is a closed model you can never download, so the API wrapper is the only door - and this node is one of the more complete doors in the ecosystem.

It uses Google's native generateContent endpoint with imageConfig, so it supports the things the model actually supports: aspect ratios, 1K/2K/4K output, response modalities, safety presets, system instructions, and seed locking. The negative prompt, the photography controls (camera angle, lighting, style, detail), and the Topaz Gigapixel upscale hook are the pack's own additions layered on top.

The inputs that matter

• api_key - Google AI Studio key. This is the official path, so no mirror needed (there's a separate mirror node if you want that).
• image - the IMAGE you're editing.
• prompt / negative_prompt - what to do, and what to avoid. "Transform this image" is the default prompt; be more specific.
• model - gemini-3-pro-image-preview (default) or the 2.5-flash image models. Pro is slower and pricier but sharper.
• aspect_ratio - 11 options from 1:1 to 21:9, or Auto.
• response_modality - TEXT_AND_IMAGE (default) or IMAGE_ONLY. Text-and-image is useful when you want the model to explain what it changed.
• output_resolution - Auto / 1K / 2K / 4K. 4K is real and it shows, but it costs more.
• upscale_factor + gigapixel_model - the Topaz Gigapixel hook (2x/4x/6x). Note: this needs the optional face-restore/Gigapixel dependencies to actually do anything; at 1x (default) it's a no-op.
• quality, style, detail_level, camera_control, lighting_control, template_selection - the photography knobs. Defaults are mostly "Auto," and honestly Auto is fine to start.
• temperature / top_p / top_k / max_output_tokens / seed - sampling.
• enable_iterative_refinement - the conversation mode. Turn this on and the node keeps last turns (default 3) so you can refine an edit iteratively in the same session, with lock_seed and auto_use_last_image helping keep things stable.
• safety_level - 5 presets from default to off. The model still refuses at Google's end regardless; this just tunes the API-level filter.

Outputs: generated_image (IMAGE) and generation_text (STRING). The image goes to a preview/save node; the text is what the model says about its work.

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

requirements.txt needs google-genai>=1.51.0 for the image endpoints. Restart.

Gotchas

The official API costs real money per call - 4K Pro edits add up fast. Iterative refinement is stateful, so if you're getting weird outputs that reference previous edits, flip reset_conversation on. And the upsample options are a trap if you haven't installed the Gigapixel extras: the node will quietly run at 1x or error, so install requirements_face_restore.txt too if you plan to use them.