Gemini编辑丨API

"Change the background to a blue sky," "turn this into watercolor" - Gemini's image-editing models take a natural-language instruction and a picture, and hand you back an edited picture. GeminiEditAPI is the QING pack's node for exactly that, and it's the closest thing in this pack to what the KB's external-API essay calls the closed-model frontier: no local weights, no VRAM cost, just an API key and a bill. The default model is gemini-2.5-flash-image-preview, which the tooltip describes as a professional image editor supporting up to three images at once, style transfer, and content edits.

It's part of ComfyUI-QING's API category, platform fixed to Google AI Studio. You wire in up to three IMAGE inputs, describe the edit, and out comes edited_image plus conversation info and token usage.

How it works. The node sends your instruction and images to the Gemini editing endpoint and returns the edited image as an IMAGE tensor - so it plugs into the same downstream graph (preview, save, further processing) as any locally generated image. The framework keeps a short multi-turn history (default 4 rounds) so you can iterate: "make it more dramatic" after the first edit, within the same node. The tooltip's own best practice is low temperature (0.2 default) to keep edits consistent - high randomness makes image edits wander, which is a genuinely useful hint baked in by the author.

The inputs that matter.

• text_input - the edit instruction. Concise and explicit ("将背景改为蓝天" / "make the background a blue sky") gets better results than vague wishes.
• image1 - required; the primary image to edit.
• model - gemini-2.5-flash-image-preview (the only choice).
• max_tokens (default 2048) and history (default 4).
• Optional: temperature (0.2), top_p, image_quality, reasoning_effort (none/low/medium/high), clear_history, plus image2 and image3 for multi-image edits.

Outputs: edited_image (IMAGE), conversation_info, total_tokens.

How to install. ComfyUI-QING installs as one pack - search "ComfyUI-QING" in ComfyUI Manager, or:

cd ComfyUI/custom_nodes
git clone https://github.com/GAO-SHIQING/ComfyUI-QING
cd ComfyUI-QING
python install_dependencies.py

Set your Google AI Studio key in Settings → QING → API Configuration and restart. The README's clone URL is typo'd (GAOSHI-QING, missing the H) - use the URL above or Manager.

Troubleshooting. This is a paid API, and image editing is a slow, token-heavy job - if the call times out, lower image_quality from high to auto and keep max_tokens sane. Second: don't crank temperature for "creativity"; the author's own advice (low temperature, 0.2) is right, and you'll notice edits getting inconsistent as you push it up. Third, remember the filter: whatever Google's image model refuses, this node refuses too - there's no local bypass for a closed model, so plan your requests accordingly. And edited images come back as tensors in memory; wire edited_image into a Save/Preview node or it's gone when the queue clears.