Analyze one video into a reconciled millisecond timeline, searchable Markdown, and structured JSON with optional visual/audio descriptions, transcription, subject tracks, and removable dead-part suggestions.
Where: Flowgen -> Add -> Video & Motion -> AI Edit -> Video Analysis
Before you start
- Connect a readable video to Video / video_in.
- Use a cloud-hosted clip for full visual analysis; a local-only f0local asset may provide transcription but can require reconnecting or uploading for visual analysis.
Key ideas
- One video, two synchronized outputs: video_in accepts one clip. md_out emits the readable editing report and json_out emits the normalized machine-readable result. Markdown is derived from the JSON after normalization, so the two outputs do not invent different timecodes.
- Segmentation strategy: Combined reconciles useful visual cuts, sound changes, speech boundaries, and dead-part edges. Auto chooses a useful strategy. Shots, Sound, and Speech isolate one type of boundary when the downstream task needs that interpretation.
- Auto or Deep: Auto performs one complete model pass. Deep performs a complete first analysis and then re-watches the entire clip in a second verification pass, replacing the candidate with corrected boundaries, speakers, boxes, and dead-part findings.
- Optional evidence: Independently include what is visible, what is audible, a speaker-labelled transcript, tracked subject positions, and dead parts. All five are on for a new node.
- Duration-based Sparks: The estimate uses clip duration plus context. Deep is two independently metered rounds, so it costs approximately twice the same Auto analysis. The card quotes the expected Sparks before the run and charges after completion.
- Video Analysis card: VideoAnalysisNode uses the authored shared-card path. Analyze a clip into a normalized editing timeline, transcript, subject tracks, and dead-part findings.
- Properties sidebar: Card-owned analysis controls; common properties shell
- Opened surfaces: Markdown preview; structured timeline preview
Steps
- Connect a video source to Video / video_in.
- Choose Combined for a general editing timeline, Auto for model-selected boundaries, Shots for visual cuts, Sound for audible sections, or Speech for speaker/thought boundaries.
- Leave only the evidence you need enabled: visual description, audio description, transcription, subject positions, and/or dead parts.
- Choose Auto for one pass or Deep when the additional verification pass is worth the extra time and Sparks.
- Click Analyze video or Analyze + verify. Use Stop analysis to cancel an in-flight run.
- Review the segment, speaker, subject, and dead counters, then open Markdown or JSON from the result strip.
- Download the Markdown report when a human editor needs the cut notes, or connect md_out/json_out to Transform, Gen Text, Brain/records, Route, or another typed text consumer.
Inputs, outputs, and card controls
- Video / video_in: one source clip.
- Markdown / md_out: the portable human report.
- JSON / json_out: pretty-printed normalized structured data carried on a text socket.
- Analysis depth: Auto - one pass; Deep - verify pass.
- Segmentation: Combined, Auto, Shots, Sound, Speech.
- Result actions: preview Markdown, preview JSON, download Markdown, stop while running.
Structured timeline fields
- durationMs, estimatedResolution, optional estimatedFrameRate, summary, mood, model, and timecodeBasis.
- Scenes with startTimeMs/endTimeMs, visible and audible descriptions, on-screen text, scene type, tags, transcript lines, subject tracks, and dead-part ids.
- Transcript lines keep stable speaker labels and integer-millisecond intervals.
- Subject samples use normalized x/y/width/height boxes from the original frame top-left.
- Dead-part kinds include freeze, silence, stall, false-start, setup, black-frame, and other.
Cut suggestions are conservative
A dead-part suggestion receives a confidence and recommendedAction. The analyzer uses cut only for high-confidence findings; lower-confidence intervals are marked review. Treat these as editorial evidence, not destructive edits: this node does not trim the source.
Timing precision
The timeline is integer milliseconds. Gemini timecodes are model-estimated observations unless the result says sample-clock; they are not decoded frame packet timestamps. Check exact cut frames in Annotate & Edit, Video Timeline, or the source-range editor before a final render.
Video Analysis controls and presets
- Add path: Flowgen → Add → Video & Motion → AI Edit → Video Analysis.
- Card: VideoAnalysisNode.
- Properties: Card-owned analysis controls; common properties shell.
- Controls: Auto/Deep; Segmentation; Descriptions; Transcription; Subject positions; Dead parts.
- Opened surfaces: Markdown preview; structured timeline preview.
Tips
- Use Combined for the first pass, then rerun with Shots or Speech only when you need a cleaner specialized index.
- Turn off expensive/noisy evidence fields that the next node will not use.
- Subject boxes are useful for reframing guidance, but verify fast motion, occlusion, and off-screen exits.
- Keep JSON for automation and Markdown for people; both describe the same normalized result.
Limitations and important notes
- The node analyzes and reports; it does not cut, reframe, caption, or render the clip.
- Model-estimated boundaries can drift from exact decoded frames.
- Local-only assets can lack the cloud-readable pixels needed for full visual analysis.
- Deep performs two complete model rounds and is not a free confidence toggle.
Troubleshooting
Analyze is disabled
Connect a video and wait for its source URL to resolve. A wire can exist while the asset is still loading.
Local analysis says transcription failed or asks for a cloud-hosted video
Reconnect the project-library asset or upload/mirror it so the multimodal analyzer can read the clip.
There are too many or too few segments
Switch from Auto/Combined to the boundary signal you actually need: Shots, Sound, or Speech.
A dead part should not be removed
Follow the confidence and review label, then inspect the interval in an editor. The node never removes it automatically.
FAQ
Does Deep only verify suspicious sections?
No. It re-watches the complete source and returns a complete corrected replacement result.
Can I use json_out as typed JSON?
The socket is text carrying valid formatted JSON. Parse/extract it with Extract / Merge or another JSON-aware consumer.
Is this the same as right-click Analyze in Assets?
They share the analysis foundation, but the Flowgen node exposes the feature matrix and reusable md_out/json_out graph outputs.