NotebookLM’s Audio Overview changed how people consume research. Upload your sources, click a button, and get a 15-minute podcast where two AI hosts discuss your material with accurate citations. Millions of people now listen to AI-generated briefings on their commute instead of reading summaries at their desk.

Video Overview is the visual counterpart. Instead of audio-only, you get a narrated slide presentation – AI-generated visuals synchronized to a voiceover script – that covers the key themes of your source material. NotebookLM launched Video Overview on the web in early 2026, followed by a higher-production “Cinematic Video Overview” variant that leans on Veo-3 for fluid animations.

Both are cloud-only. The video renders on Google’s servers. You watch it in a browser tab. You cannot download the MP4. You cannot edit the slides. You cannot change the voiceover. You cannot render it offline. You cannot use it in a presentation. The video exists inside notebooklm.google.com and nowhere else.

GeminiDesktop implements Video Overview with a fundamentally different architecture: the slides are generated by Gemini’s structured output, optional images are generated per slide (with Veo-3 motion where appropriate), and the entire video is rendered locally via Remotion. No server-side rendering. No upload. No cloud dependency for the final output. This post explains how it works and what that architectural difference means for your workflow.

TL;DR / Key takeaways

NotebookLM’s Video Overview renders on Google’s servers and plays only inside notebooklm.google.com. No MP4 download, no slide editing, no offline playback.
GeminiDesktop builds Video Overview as a three-stage pipeline: Gemini-generated script, optional per-slide image generation (Veo-3 for motion), and local Remotion rendering.
The slides JSON is editable – reorder, rewrite, swap images, adjust timing. Re-renders happen instantly in the Remotion Player.
Two visual styles: narrative (clean text-forward slides) and cinematic (full-bleed visuals with parallax). Both render from the same slides JSON.
Compares to Synthesia (corporate avatar video), HeyGen (voice cloning), Opus Clip (long-to-short clipping) – none solve “turn my research into a summary video.” Video Overview does.
Rooted in Google’s Veo line – see our Veo glossary for the model context.

How the pipeline works

Video Overview in GeminiDesktop is a three-stage pipeline. Each stage is independent and produces an artifact you can inspect and modify before the next stage runs.

Stage 1: Script generation from Audio Overview

The pipeline starts with an Audio Overview script. If you have already generated an Audio Overview for your notebook, that script is the input. If you have not, the system generates one first. The Audio Overview script is a structured document that covers the key themes of your sources in a conversational format – it already has the narrative arc, the topic transitions, and the grounding citations that a video needs.

From this script, Gemini’s structured output generates a slides JSON: an ordered array of slide objects, each with a title, body text, speaker notes (the voiceover transcript for that slide), visual description (what the slide image should depict), and timing metadata (duration in seconds). The structured output schema enforces these fields, so every slide is well-formed by construction.

A typical 10-minute Video Overview produces 15 to 25 slides, each lasting 20 to 40 seconds. The slide count and pacing adapt to the density of the source material – a technical paper with many distinct concepts produces more slides with shorter durations; a narrative essay produces fewer slides with longer, more reflective segments.

Stage 2: Optional image generation per slide

Each slide in the JSON includes a visualDescription field: a natural-language description of what the slide image should show. This field serves double duty. If you choose to generate images, it becomes the prompt for an image generation model (Nano Banana 2 / Gemini 3.1 Flash Image). If you skip image generation, it becomes alt text and a visual placeholder description that helps you understand the intended visual context.

For the cinematic style, motion-forward slides use Veo-3 to generate short animated clips synchronized to the voiceover. Veo-3 handles scene transitions, camera moves, and subject animation while preserving the grounded content of the slide. See the Veo glossary for the underlying model. Non-motion slides continue to use static image generation, which is faster and cheaper.

Image generation is optional and per-slide. You can generate images for all slides, for selected slides, or for none. When you generate images, the model produces visuals that match the content of each slide – diagrams for technical concepts, scene illustrations for narrative content, data visualizations for quantitative findings. The images are generated asynchronously and slotted into the slides JSON as base64 or local file references.

This optionality matters because image generation takes time and costs compute. For a quick internal briefing, text-only slides with clean typography are often sufficient. For a presentation you intend to share externally, generated images add significant production value. You choose based on the use case, not based on what the platform permits.

Stage 3: Remotion Player rendering

The slides JSON – with or without generated images – feeds into Remotion Player, which renders the video entirely in the browser. Remotion is a React-based framework for creating videos programmatically. Each slide becomes a Remotion composition with animated transitions, timed text reveals, and synchronized voiceover audio. The Remotion Player component provides a standard video player interface: play, pause, seek, scrub, full-screen.

The critical architectural point: Remotion Player runs as a pure client-side React component inside GeminiDesktop’s webview. There is no server, no upload, no render farm. The browser’s own rendering engine composites the frames in real time. The video is available instantly after the slides JSON is generated – no “rendering” wait time. You click play and it plays. You scrub to any point and the frame appears.

Narrative and cinematic styles

GeminiDesktop ships two visual styles for Video Overview, corresponding to two different use cases.

Narrative style

Clean, text-forward slides with clear hierarchy. Title at top, body text below, optional supporting image to the side. Simple cross-fade transitions. Optimized for internal briefings, study reviews, and contexts where the viewer is watching to learn. Works well without generated images – a well-designed text slide with proper typography communicates effectively on its own.

Cinematic style

Full-bleed images or Veo-3 motion clips with text overlays, parallax motion effects, and more sophisticated transitions. Optimized for presentations you intend to share: conference talks, social media clips, team all-hands meetings, client deliverables. This style benefits significantly from generated visuals since the full-bleed layout makes the image (or motion clip) the primary visual element.

Both styles render from the same underlying slides JSON. Switching between them is a style parameter change, not a regeneration.

Use case walkthroughs

Academic presentation prep

A postdoc has a 30-page paper accepted at a conference. Needs a 10-minute talk version. Generate a Video Overview in narrative style. Review the slides JSON – reorder two sections, tighten the methodology slide, add a reference citation to slide 14. Re-render instantly. Screen-capture the Remotion Player output for the conference slot. What used to take two days of slide design takes an afternoon.

Journalist explainer

A reporter covering a dense regulatory filing wants to drop an explainer video into the published article. Generate a Video Overview in cinematic style. Veo-3 motion clips animate the key scenes – a courtroom, a data center, a supply chain map. Swap any slide where the generated visual misrepresents the story. Export (when MP4 export lands) or screen-capture to embed in the CMS.

Internal all-hands

A product team wants to summarize user-research findings for a company all-hands. Drop the 40 interview transcripts into a notebook. Generate a 12-minute Video Overview. Narrative style keeps it business-like. Share the Remotion Player link internally; anyone on the team opens GeminiDesktop to watch.

Teacher building course modules

A teacher building an online course has 15 reading assignments. Generate a Video Overview for each. Publish the resulting videos (once MP4 export ships, or via screen capture today) as pre-class briefings. Students arrive already oriented; class time goes to discussion, not lecture.

Podcast-to-video repurposing

You already have Audio Overview MP3s of your research (see Audio Overview). Use them as input to Video Overview. The script reuses. Slides get generated. You now have both an audio artifact and a video artifact from the same source effort.

How this compares to NotebookLM’s Video Overview

NotebookLM’s Video Overview and GeminiDesktop’s Video Overview solve the same user problem: turning research into a watchable video summary. The implementations differ in architecture, and that architectural difference produces meaningful workflow differences.

Rendering location

NotebookLM renders video on Google’s cloud infrastructure. You submit the generation request and wait for the rendered video to appear. The rendering time depends on Google’s server load, the length of the video, and the complexity of the visual style. You cannot control any of these variables.

GeminiDesktop renders video locally in the browser via Remotion Player. The only network dependency is the initial generation of the slides JSON and optional image generation. Once those artifacts exist locally, playback and interaction are fully offline-capable.

Editability

NotebookLM’s Video Overview is a finished artifact. You watch it. If the content emphasis is wrong, the slide ordering is off, or a particular visual does not match the concept, you regenerate the entire video and hope the new version is better. There is no slide editor. There is no way to swap one image. There is no way to adjust the voiceover timing for a single slide.

GeminiDesktop’s Video Overview is built from a slides JSON that you can inspect and modify. Each slide is a JSON object. You can reorder slides, edit text, change timing, swap images, or remove slides entirely. The Remotion Player re-renders from the modified JSON in real time. The video is not a black box – it is a document you can edit.

Export and reuse

NotebookLM’s Video Overview lives inside notebooklm.google.com. You can share a link. You cannot download the MP4. You cannot extract individual slides. You cannot embed the video in a Notion page, a Google Slide deck, or a Keynote presentation without screen-recording it.

GeminiDesktop plans to support local MP4 export via Remotion’s rendering pipeline in a future update. The slides JSON itself is already exportable and can be used as input for other video tools or custom Remotion compositions. The intermediate artifacts – the script, the slides JSON, the generated images, the Veo-3 motion clips – are all accessible as local files.

Privacy

NotebookLM processes your source material on Google’s servers for both the content analysis and the video rendering. Your PDFs, your notes, your research data all transit through Google’s infrastructure.

GeminiDesktop sends your source material to Gemini’s API for the structured output generation (the same API call used for any Gemini interaction), but the video rendering itself is entirely local. Your slides JSON, your generated images, and your rendered video frames never leave your machine during the visualization and playback stages.

How it compares to adjacent video-AI tools

Capability	NotebookLM	Synthesia	HeyGen	Opus Clip	GeminiDesktop
Source: your documents	Yes	No (script you write)	No (script you write)	No (existing video)	Yes
Grounded citations	Yes	No	No	Source video	Yes
AI avatar presenters	No	Yes	Yes	No	No
Voice cloning	No	Yes	Yes (strong)	Source voice	No
Long video to short clips	No	No	No	Yes (the product)	No
Narrated slide summaries	Yes	No	No	No	Yes
Cinematic motion (Veo-3)	Yes	No	No	No	Yes
Local rendering	No	No	No	No	Yes (Remotion)
Editable slides JSON	No	Limited	Limited	N/A	Yes
Typical price	Free + Pro	$22+/mo	$24+/mo	$9+/mo	Free + API

Synthesia and HeyGen are excellent for avatar-led corporate training where you own the script. They don’t “summarize” your research; they narrate what you tell them.

Opus Clip takes long video and carves it into social clips. Different problem entirely.

NotebookLM is the closest peer. Same source-grounded approach, same generation from documents. Architecture differs: cloud vs. local, black box vs. editable.

Why local rendering matters more than it sounds

“Local rendering” sounds like a technical implementation detail. In practice, it changes the user experience in ways that compound over repeated use.

Instant playback. There is no render queue. You do not submit a job and check back in five minutes. The Player composes frames in real time. You click play and it plays.

Iterative refinement. Because playback is instant, you can watch the video, identify a slide that needs adjustment, edit the JSON, and watch the updated version in seconds. This tight feedback loop is impossible when rendering happens on a remote server with a multi-minute turnaround.

Offline capability. Once the slides JSON and images are generated, the video works without an internet connection. You can review and present your Video Overview on a plane, in a conference room with unreliable wifi, or anywhere else where cloud-dependent tools fail silently.

No usage limits on playback. The generation of slides JSON and images counts against your Gemini API usage, but playback, scrubbing, and re-rendering from modified JSON are unlimited local operations. You can iterate on the video indefinitely without burning API credits.

Why Google hasn’t shipped a desktop Video Overview

NotebookLM’s Video Overview is still tied to Google’s web infrastructure because that’s the only surface Google ships NotebookLM on. A desktop renderer would require NotebookLM to expose a public API for script/slides JSON extraction – which Google hasn’t done – or Google itself would need to build a cross-platform desktop app, which the Labs team has chosen not to do. This is the same structural story we covered in Why NotebookLM Has No Native App. Until Google’s incentives change, the desktop gap stays open – and GeminiDesktop fills it by doing the script and slides generation against the public Gemini API plus local Remotion rendering.

Windows context

On Windows, Google has shipped neither a NotebookLM app nor a native Gemini app. The Google App for Windows (2026-04-14) is a thin search shell. Video Overview is entirely out of reach for Windows users through any official Google surface. GeminiDesktop’s Windows binary runs the full Video Overview pipeline natively. See Why Google didn’t make a Gemini Windows app for background and Gemini Windows install guide for setup.

Limitations

Video Overview depends on Gemini API quota – longer videos and per-slide image generation burn through it faster. Veo-3 motion clips cost significantly more than static images; use them selectively. Remotion Player’s real-time composition is fast on modern hardware but may drop frames on older Intel Macs with no discrete GPU. MP4 export via Remotion’s server-side rendering is on the roadmap but not yet shipped – today’s workflow is screen capture for external sharing. And while the slides JSON is editable, there’s no WYSIWYG slide editor yet; you edit the JSON directly, and the Player re-renders.

FAQ

Does NotebookLM let me download the Video Overview as MP4? No. NotebookLM’s Video Overview plays only in the browser tab at notebooklm.google.com. No MP4 download.

Can I edit individual slides in Video Overview? In NotebookLM, no – it’s a finished render. In GeminiDesktop, yes – edit the slides JSON and the Remotion Player re-renders instantly.

What model powers the voice synthesis? The same gemini-3.1-flash-tts-preview TTS model used by Audio Overview. See Audio Overview for the TTS pipeline deep-dive.

What model powers the motion clips in cinematic style? Veo-3, Google’s text-to-video generation model. Background: Veo glossary.

Does Video Overview work offline? Playback of already-generated videos, yes. Generation requires the Gemini API.

Is Video Overview free? NotebookLM’s Video Overview is available on free and Pro tiers with generation quotas. GeminiDesktop’s generation counts against your Gemini API quota; there’s a free tier, but heavy Veo-3 use accrues cost.

Get started

Video Overview is available now in GeminiDesktop. Upload your sources, generate the overview, choose your style, and present your research as a video – rendered locally, editable at every stage, grounded in your actual sources.

Try it at geminidesktop.app/app.