Multi-Caption Combiner

Captioning a video frame-by-frame gives you detail - every still described individually. What it doesn't give you is a story. Ten captions about ten frames say nothing about what happened. Multi-Caption Combiner closes that gap the pragmatic way: it takes your per-frame captions, throws them at Gemini, and asks for a single coherent description with a style you choose. It's the finishing step of the pack's local-captioning chain.

Inputs: captions is the raw material - comma-separated or newline-separated captions, one per frame. gemini_api_key and gemini_model (default models/gemini-2.0-flash-exp) are entered right on the node, which is a deliberate difference from the pack's settings-based Gemini path: this node takes its key in the widget, so it works standalone. combination_style gives you four merge strategies:

• Action Summary (default) - what happened, compressed
• Chronological Narrative - the events in order
• Movement Flow - focused on how things moved
• Custom - you supply the instruction in custom_prompt

Then the presentation knobs: optional timestamps (comma-separated seconds matching each caption, which the node can weave in for a time-coded narrative), and output_format - Paragraph, Bullet Points, or Scene Description. If you want a chunk of prose to drop into a prompt builder, Paragraph is the one; Scene Description nudges it toward the pack's Wan-flavored descriptive style.

Two outputs: combined_caption (the merged text) and gemini_status (a status string telling you the model, style, caption count, and format - handy confirmation that the run used what you think it did). Under the hood it builds a system prompt from your style choice, sends the captions (plus timestamps) to Gemini, and returns the text - with the same retry-on-overload logic the rest of the pack uses for its API calls.

Installing

Part of Swiss Army Knife:

• ComfyUI Manager: search "Swiss Army Knife", install, restart.
• Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/sammykumar/ComfyUI-SwissArmyKnife.git
cd ComfyUI-SwissArmyKnife
pip install -e .

The flow it belongs in

Frame Extractor → (LLM Studio Picture Describe per frame) → Multi-Caption Combiner. Frames become captions become a story. The timestamps input is the sleeper feature: feed it the frame_timestamps output from Frame Extractor and your combined narrative gains a time axis.

Gotchas

• Key is on the node here, not in settings. If you're used to the pack's settings-based Gemini config, this one will bite - a blank key means an auth error even when everything else works.
• Caption order matters. The node assumes input order = chronological order. Scramble the captions and the "chronological narrative" will be fiction.
• Timestamps are optional and must match the caption count. Mismatched counts cause the merge to behave oddly - when in doubt, leave timestamps out.
• gemini_status is your diagnostic friend: if the model came back with the wrong style or format, that's the first place to look.

It's a small node with a clear job: turn a pile of frame captions into the one paragraph you can actually use. In the pack's ecosystem, it's how per-frame detail becomes a usable prompt.