Gemini-Multimodal

Most nodes in this pack generate images. This one understands them. Gemini-Multimodal is the analysis node: wire in up to four images, a video, or an audio clip, type a question, and Gemini responds with text. "请对所有提供的媒体进行详细的综合分析" is the default prompt - "please provide a detailed comprehensive analysis of all provided media." You'd use it to caption a reference frame, sanity-check a video clip before deciding on prompts, transcribe audio context, or feed an image's description into a prompt-generation chain.

It's the same Gemini generateContent machinery as the pack's image nodes, but with media going in and only text coming out (plus diagnostics). And like the Banana2 node, it got the unified endpoint/auth upgrade - base_url, version, auth_mode - so it works against Google's official API or any of the mirrors/proxies in the pack.

How it works

Each media input is base64-encoded and sent as inline_data parts alongside your prompt. The model input is the thing to watch: gemini-3-pro-preview (default), gemini-3-pro-preview-thinking (labeled for which providers carry it), and the 2.x models. The thinking_level switch (high/low) only does anything on models that actually have a thinking mode - on a plain gemini-2.0-flash it's cosmetic.

The inputs that matter

• prompt - what you're asking about the media.
• model - pick the thinking model if you want reasoning quality; it's slower and costlier.
• api_provider - google, comet, the T8/Comfly mirrors, aabao, or custom. The [all] / [comet/T8...] tags on the model names tell you which providers carry each model.
• api_key, base_url, version, auth_mode - the unified endpoint/auth controls (Auto version picks v1beta on Google domains, v1 on mirrors).
• max_output_tokens - up to 8192.
• temperature - 0–2, default 1.
• media_resolution - Auto / low / medium / high. Matters for video and high-detail images; higher is more accurate and more expensive.
• image_1 … image_4 (optional IMAGE), video (optional VIDEO), audio (optional AUDIO), system_instruction (optional).

Outputs: response (STRING, the analysis), thought_signature (STRING, present when a thinking model runs), and usage_info (STRING, token/cost diagnostics).

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. Video/audio support needs the pack's opencv-python/pydub/scipy deps, which requirements.txt installs.

Gotchas

The practical trap is media_resolution staying on Auto/low while you're analyzing detailed images - the model sees a downscaled version and misses fine detail; bump it up for real work. And on the thinking model, thought_signature is a nice debugging artifact but means you're paying for reasoning tokens on every call. If the model label has a [comet/...]-style tag, don't pair it with a provider outside that list - the node will 404 or return "model not found."