URL Gemini图像编辑丨API

The stock GeminiEditAPI only talks to Google AI Studio. CustomURLGeminiImageEditAPI is the version that lets you point Gemini's image-editing capability at whatever base URL you like - Google's native endpoint, a Vertex-style setup, or one of the reseller gateways (the pack's own tooltip suggests https://api.apiyi.com as an example) that proxy Gemini image models with easier billing. If you've been burned by the friction of a Google Cloud project, this node is the "paste a gateway key and go" route. The cost tradeoffs - reseller undercutting, region access, looser onboarding - are the ones the KB's external-API essay documents for the whole Nano-Banana-style proxy layer.

It's part of ComfyUI-QING's API category. Mechanically it's a different animal from the OpenAI-style nodes: this one calls Gemini's generateContent REST format directly, not chat completions.

How it works. You give it a base_url (the Gemini-compatible root, e.g. https://api.apiyi.com) and a model ID (the tooltip's example is gemini-3-pro-image-preview). It builds a generateContent request: your prompt becomes the first text part, and the image(s) - up to three via image2/image3 - are attached as independent inlineData parts. Then the Gemini-specific controls kick in: aspect_ratio (ten ratios from 1:1 to 21:9), image_size (1K/2K/4K), and response_modalities (IMAGE or TEXT,IMAGE) map into generationConfig. The key-sending mode is the thing most people get wrong, and it's a dropdown: Authorization Bearer (what gateways usually want), x-goog-api-key (native Gemini style), query key, or none for endpoints with no auth.

The inputs that matter.

• prompt - the edit instruction.
• image - required; the main image to edit.
• base_url and model - the gateway root and its model ID.
• api_key_mode - pick Bearer for gateways, x-goog-api-key for native Gemini.
• aspect_ratio, image_size, response_modalities - the output geometry/format controls.
• timeout_sec (default 300 - image generation is slow).
• Optional: image2/image3, custom_headers_json, extra_body_json, api_key_override.

Outputs: image (IMAGE), raw_response, status, debug_info.

How to install. ComfyUI-QING is a one-pack install - 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

Restart after installing. The README's clone URL is typo'd (GAOSHI-QING, missing the H) - use the URL above or Manager.

Troubleshooting. api_key_mode is where almost everyone stumbles: a gateway that expects Authorization: Bearer will 401 if you leave the node in the native x-goog-api-key mode, and vice versa - switch the dropdown before you question the key itself. Second, base_url must be the Gemini-compatible root, not a /v1/chat/completions endpoint - this node doesn't speak OpenAI format. Third, if you get "modalities" errors, check response_modalities against what the gateway supports (TEXT,IMAGE needs a model that returns text alongside the image). And the honest cautions: gateways may run looser moderation than Google's first party, but the model still enforces its own refusals; your images go to a third party; and the "reseller undercutting" price is real but so is the risk of an unknown endpoint handling your prompts - only point this at a gateway you have a reason to trust.