gemini-pro-文件处理

Gemini File Processing is the payoff of this pack. Feed it a GEMINI_FILE handle - the kind the companion Gemini File Upload node produces - plus a prompt, and it runs Google's Gemini model over that file and hands back the response as text. It's how you get a real language model to summarize an audio file, extract the transcript of a long video, or answer questions about a document that was too big to inline into a normal API call.

It's the only node in the pack that is genuinely an output node, meaning it can stand alone at the end of a workflow and render its result. That's fitting: it's the last step. Upload on one side, questions here, answers out.

How it works

The mechanism is straightforward. It loads your API key from the pack's config.json (there's no api_key input here - you set the key once via the main Gemini Pro node, or hand-edit the config file), constructs a combined prompt of the form System: <user_prompt>\nUser: <prompt>, and calls generate_content with your uploaded file attached.

Two behaviors worth knowing:

• stream is a real toggle. With it off (default), the model answers in one shot. With it on, the node streams chunks and stitches them together - slightly faster first-token feel on long outputs.
• Retries are baked in. ResourceExhausted (i.e. rate-limited) responses trigger exponential backoff up to five attempts. This pack clearly spent its early life fighting Google's free-tier quotas.

The inputs that matter

• file - the GEMINI_FILE from Gemini File Upload. Required, and it has to be a handle the node recognizes.
• prompt - your actual question ("summarize this audio", "what's being said at 10:30?"). Default is a Chinese "analyze and summarize this file" prompt.
• user_prompt - the system-style instruction that gets prepended, defaulting to a "you are a professional file analysis assistant" persona.
• model - same family as the main node, default gemini-2.5-flash-preview-09-2025, with gemini-3-pro-preview, gemini-2.5-flash-preview-05-20, and an older gemini-2.0-flash-exp available.
• max_output_tokens - defaults to the full 65536 ceiling. For a summary you can dial this way down and save money; for a full transcript you want it maxed.
• temperature, proxy - as you'd expect from the rest of the pack.

Output is a single STRING named response. Wire it into a Show Text node, or pipe it into anything else that takes text.

Install

Same pack as the other two nodes - install once, get all three:

cd ComfyUI/custom_nodes
git clone https://github.com/penposs/ComfyUI_Gemini_Pro.git

Or search "ComfyUI Gemini Pro" in ComfyUI Manager, then restart ComfyUI. Dependencies are whatever the pack's requirements.txt pulls in (google-generativeai and friends); if the SDK conflicts with other packages, the README's advice is to uninstall and reinstall google-generativeai fresh.

Common issues

• "File not in an ACTIVE state." The most useful error message in the whole pack. Files API objects go stale - they expire, or the handle was already consumed. The node even ships a dedicated hint telling you to re-upload the file. That's the fix: re-run the upload node and feed the fresh handle.
• No key in config.json. You'll get a Chinese error about the missing API key before anything else runs. Set the key on the main Gemini Pro node first - it persists to config.json - and this node picks it up.
• Empty output. If the model returns nothing, the code checks prompt_feedback.block_reason - Gemini's safety filters can silently swallow a request, especially on audio/video with speech. The generic "no valid output" message usually means that.
• Slow. Every call starts with a deliberate delay and retries add more. Long files plus a flash model is the responsive combo; Pro models on huge inputs will test your patience.

It's a niche little node, but if you need Gemini to actually read the big media files in your ComfyUI workflow, it's the missing half of the equation.