Video Understanding & Analysis via Google's Gemini
Ask questions about video content, generate summaries, or extract structured information from clips inside a CNAPS Studio flow. Four Gemini tiers let you trade off speed, cost, and reasoning depth. Runs on Google's servers via API.
🔑 Requires a Google API key
What it does
This node connects your CNAPS Studio flows to Google's Gemini models for video analysis through the Gemini API. Feed it a video and a natural-language instruction — summarize, answer questions, transcribe, extract structured data — and the node returns a text response. Unlike the generation nodes, this node analyzes existing video rather than creating new content, and its outputs are text.
Problem it solves
- Video understanding inside a pipeline — Analyze clips as a flow step: e.g., generate a video with Sora/Veo, then run Gemini analysis for automated QA or captioning, all in one flow.
- Speed/cost/depth tiers — Match the model tier to the task: fast defaults for routine analysis, Pro for complex multi-step reasoning, Flash-Lite for high-volume batch transcription.
- Reproducibility across runs — The prior-gen
gemini-3-flashremains available when consistency with earlier analysis runs matters more than latest-gen quality. - Long-form video without babysitting — Automatic soft-chunking breaks videos longer than 25 minutes into overlapping segments before analysis, so you don't have to trim manually
Configuring the Node
Basic Settings
Model name
Model | Best For |
gemini-3.5-flash | Default fast tier — latest-gen Flash, best speed/quality trade-off for most analyses |
gemini-3.1-pro | Highest-quality reasoning — complex multi-step analysis, nuanced summaries, long-context understanding |
gemini-3-flash | Prior-gen fast tier — cost-efficient legacy option, useful when reproducibility with earlier runs matters |
gemini-3.1-flash-lite | Ultra-lightweight — cheapest per call, great for high-volume batch processing or simple transcription |
Number of highlight clips — how many highlight clips the model should return from the video.
Clip length range (sec) — the minimum and maximum length (in seconds) each highlight clip should be (default 15 – 60 seconds). Gemini chooses each clip's exact length within this range based on the content.
Advanced Options
Option | Default | Description |
EDL target length (seconds) | 45 | Only used in EDL mode. The total length of the final sequenced playlist. Gemini selects and orders clips whose combined runtime fits this target. Ignored in Original mode. |
Output language | Auto | The language for the returned text (Auto lets Gemini choose based on the video's audio / on-screen text). |
Focus / guidance (optional) | (empty) | A short natural-language hint about what to look for — e.g., "product close-ups", "moments of laughter", "goal attempts", "scenes with the presenter on screen". Leave empty for a general highlight extraction. |
Max tokens | 4,096 | Maximum length of Gemini's text response (range 1 – 8,192). |
Temperature | 0.4 | How creative the response is (range 0 – 2). Lower = more focused and consistent; higher = more varied. |
Soft-chunk threshold (min) | 25 | Videos longer than this many minutes are automatically split into chunks before being sent to the API (range 5 - 20). |
Soft-chunk window (min) | 20 | Length of each chunk when soft-chunking kicks in (range 5 - 60). |
Soft-chunk overlap (min) | 2 | How much each chunk overlaps with the next, so context isn't lost at the seam (range 0 - 10). |
Notes on Soft-Chunking
For videos longer than the threshold (default 25 minutes), the studio automatically splits the video into overlapping segments before sending them to Gemini, then stitches the responses back together. This exists because:
- Very long videos can hit API length limits or context limits
- Very long videos are also slower and more expensive to process in one shot
- Overlap between chunks keeps context from getting lost at chunk boundaries (e.g., a sentence that starts at 19:58 and ends at 20:02 won't get cut in half)
Default 20-minute window with 2-minute overlap works well for most content. For dialog-heavy or fast-cutting material, a longer overlap can help. For highly repetitive or static content (surveillance, timelapse), a shorter overlap is fine.
Choosing a tier
If your flow needs… | Use |
A sensible default for most analyses | gemini-3.5-flash |
Deep reasoning over long or complex videos | gemini-3.1-pro |
Consistency with analyses run on the prior generation | gemini-3-flash |
Cheapest per-call cost at batch scale | gemini-3.1-flash-lite |
Compliance & Provenance
Provider | Google |
Provider type | GPAI — General-Purpose AI (GPAI) |
License | Google Cloud Terms |
EU AI Act risk class | Minimal Risk |
Art. 50 transparency | Not applicable |
Region availability | Available globally |
Training data summary |
For more on how we classify models and mark outputs, see our AI Policy.