Try Pro FREE for 2 weeksGet Started
Start for free

gemini-3.5-flash — LLM AI Model

Hand Gemini a video and get timecoded segments back — where the moments worth keeping are, each with a title, a hook and a score. It watches and listens rather than sampling frames. Google AI key required.

Try in Studio

What it does

Video Analysis hands a clip to Gemini and gets back timecoded segments - where the moments worth keeping are, each with a title, a hook and a score. It watches and listens rather than sampling stills, and what arrives on the Analysis port is structured data, not a paragraph of prose.

Two modes decide the shape of it. Highlights picks the best standalone moments and ranks them, one clip per moment. Single Short (EDL) instead assembles one coherent short in playback order inside a target length, which is the mode that pairs with the trim and concat nodes downstream.

When to use it

  • Finding the moments worth cutting out of a long recording
  • Assembling one short from a long video in playback order
  • Ranking candidate clips before anyone opens an editor
  • Titles and hooks written in the language the video speaks

Details

  • CategoryLLM
  • Year2026
  • LicenseGoogle Cloud Terms

Compliance & Provenance

  • ProviderGoogle · GPAI
  • EU AI Act RiskMinimal Risk
  • Art. 50 TransparencyNot applicable

Inputs & Outputs

VideoInputvideo

Video to analyze

TextInputstringoptional

Optional instruction steering which moments to pick (e.g. 'prefer surprising moments')

AnalysisOutputsegments

Timecoded highlight segments (feeds the Video Trim node)

Parameters

modelstringdefault Gemini 3.7 Flashoptions: Gemini 3.7 Flash, Gemini 3.6 Flash, Gemini 3.5 Flash, +4 more

Model variant

Which Gemini model analyses the video. 'Gemini 3.7 Flash' (default) — newest tier that reliably returns timecoded highlights on this node. '3.6 Flash' / '3.5 Flash' — older, equally reliable. 'Gemini 3.1 Pro' — highest quality, slower. Gemini 3.8 Flash is deliberately absent: measured on dev it failed 3 of 7 runs here with "analysis output had no usable timecoded highlights", while 3.7, 3.6 and 3.5 never failed. It stays available on the plain Gemini node, where there is no highlight schema to satisfy.

output_modestringdefault Highlightsoptions: Highlights, Single Short (EDL)

Output mode

Highlights (default) selects the top-N standalone viral clips ranked by virality (1 long video -> N shorts). Single Short (EDL) assembles ONE coherent short in playback order within the target length (1 long video -> 1 short); pairs with Video-Trim -> Video-Concat.

target_len_secintegerdefault 45range 10–180advanced

EDL target length (seconds)

Total target length in seconds of the assembled Single Short (EDL). A budget across all segments, not per clip. Only used when output_mode = Single Short (EDL). Default 45.

highlight_countintegerdefault 5range 1–15

Number of highlight clips

How many standalone highlight clips to select, ranked by virality (best first). Default 5.

min_clip_secintegerdefault 15range 2–180

Clip length range (sec)

Length window for each clip in seconds, as [min, max]. Drag the two handles; set both equal for a fixed length (both 5 = exactly 5s clips). Clamped so 2 <= min <= max <= 180 (180 = YouTube Shorts cap). Default 15-60.

languagestringdefault Autooptions: Auto, English, Korean, Japanese, Chineseadvanced

Output language

Language for the generated title and hook. Auto (default) follows the video's primary spoken language; set explicitly to force one.

menu_whiteliststringdefault advanced

Menu / product list (optional, EDL)

Optional list of dish or product names (comma or newline separated) shown in the video, e.g. a mukbang menu. When given, the Single Short (EDL) analysis must pick each segment's food_label FROM this list, copied exactly, and omit it when a segment's food isn't listed — so a downstream food-name label track shows the right dish instead of a plausible guess. Leave blank to let the model name foods freely.

promptstringdefault advanced

Focus / guidance (optional)

Optional guidance that steers WHICH moments to pick (e.g. 'prefer emotional or surprising moments', 'focus on the cooking steps'). The output format is enforced by the system regardless, so this cannot break the result — leave blank for default highlight selection.

max_tokensintegerdefault 4096range 1–8192advanced

Max tokens (1-8192)

Maximum tokens for the response. Default 4096 — a long video yields many segments. Raise up to 8192 when a long video is expected to produce a lot of them.

temperaturefloatdefault 0.4range 0–2advanced

Temperature (0-2, not used by 3.6 Flash and later)

Sampling temperature. 0 = deterministic (the same clip picks the same moments). 0.4 (default) = mostly faithful with light phrasing freedom. >1 = more adventurous titles and hooks. Only the older tiers use it (3.5 Flash, 3.1 Pro, 3.1 Flash-Lite, 3 Flash). Not sent for Gemini 3.6 Flash and later, including the default 3.7 Flash: Google fixes sampling on those models, so changing it has no effect.

chunk_threshold_minintegerdefault 25range 5–120advanced

Soft-chunk threshold (min)

Videos longer than this (minutes) are analyzed in overlapping windows instead of one pass, because highlight coverage degrades on long videos. Default 25.

chunk_window_minintegerdefault 20range 5–60advanced

Soft-chunk window (min)

Length (minutes) of each analysis window when soft-chunking a long video. Default 20. Keep at or below the threshold.

chunk_overlap_minintegerdefault 2range 0–10advanced

Soft-chunk overlap (min)

Overlap (minutes) between adjacent analysis windows so a highlight straddling a boundary isn't lost. Default 2.

Limitations

  • The output is a fixed segment structure. The optional prompt steers which moments get picked, not the shape of what comes back.
  • A clip can be set anywhere from 2 to 180 seconds, 180 being the cap short-form platforms accept.
  • Videos over 25 minutes are read in overlapping windows rather than one pass, because coverage degrades over a long one.
  • Runs on your own Google AI API key. Without one configured in the workspace the node cannot run.

Tags

  • llm
  • multimodal
  • video-understanding
  • transcription
  • reasoning

Alternatives in Image/Video Models

  • GPT Image

    OpenAI GPT-Image text-to-image / edit. Needs API key.

  • Nano Banana

    Google Gemini-based image gen/edit with multimodal reasoning. Needs API key.

  • Omni

    Google Gemini Omni: text/image/video in, video out, with conversational editing. Which input ports apply depends on Task — Text to Video: Text only. Image to Video: + Image 1. Reference to Video: + Image 1–3. Edit / Extend: + Video (aspect ratio and resolution are ignored). Unused ports are dropped even if wired. Needs API key.

  • VEO

    Text/image to video with native audio, now run by Gemini Omni 1.1 Flash (Google retires VEO 3.1 on 2026-10-22). Needs API key.

Resources

All AI models