Autonomous editing brain

Upload raw footage.
Get a finished film.

Cutframe is not a template editor or a pile of AI buttons. It watches the whole take, decides what matters, builds the timeline, renders it server-side, then critiques and re-cuts its own work.

Create AI EditRead the technical assessment
Live WebGL render: reel, filmstrip and waveform reacting in real time.
The studio

A whole post house, running itself

Footage in, finished film out — with your own asset library, sound design, motion graphics and colour handled inside one place.

Ingest any length

Multi-hour files, resumable chunked uploads, phone gallery or a pasted link.

Your asset library

Logos, LUTs, overlays, fonts and past exports — reusable across every project.

Sound & score

Licence-free music, SFX and sound design tucked under the edit automatically.

Motion & 3D

Title cards, lower-thirds, hooks and 3D elements rendered into the cut.

Colour & look

Exposure, balance and film looks with protected skin tones.

Export up to 4K

1080p by default, 2K/4K on demand, plus 9:16, 1:1 and 16:9 in one run.

8h+
Max source length
4K
Export ceiling
3
Ratios per render

One pipeline. No fake progress bars.

Every stage emits real state — a job either ran, is running, or failed. Nothing is simulated in the UI.

Ingest

Validation, proxies, waveform + frame index.

01
Understand

Scenes, shots, faces, speakers, quality scoring.

02
Plan

Story beats, take selection, pacing curve.

03
Edit

Deterministic timeline JSON, server-rendered.

04
Review

Self-critique pass scores the cut, then revises.

05

Editorial decisions, stated honestly

Each capability carries its real build state. Nothing ships labelled as working before the pipeline actually produces it.

MVP

Smart cutting

Silence, filler words, dead takes and repeats removed — natural pauses preserved.

MVP

Word-level captions

Speaker-aware transcripts with emphasis and safe-area placement.

MVP

Audio repair

Levelling, denoise, speech clarity and music ducking in the render graph.

MVP

Story engine

Hook, setup, main story, payoff and ending derived from the actual content.

MVP

Context-aware pacing

No fixed formula: emotional beats breathe, fast content cuts tighter.

In build

Auto reframe

9:16, 1:1, 4:5 and 16:9 with subject tracking across the cut.

In build

Beat-aware scoring

Track choice from mood and pacing; cuts snapped to detected beats.

In build

Colour pass

Exposure, white balance and contrast with protected skin tones.

In build

Multi-cam switching

Active-speaker detection picks the camera; punch-ins for single-cam.

In build

Contextual transitions

Clean cuts by default; dissolves and match cuts only when earned.

Research

B-roll reasoning

Only inserted where a shot genuinely needs visual support.

Research

Sound design

Restrained impacts and ambience tied to editorial emphasis.

ProfessionalCinematicYouTubeShorts / ReelsPodcastVlogAdvertisementEducationalGamingDocumentaryCustom
ProfessionalCinematicYouTubeShorts / ReelsPodcastVlogAdvertisementEducationalGamingDocumentaryCustom

One source. Every deliverable.

The plan is format-agnostic, so the same understanding pass produces a landscape cut, a vertical short and a podcast version without re-analysing the footage.

Aspect ratios
16:99:161:14:5
Caption styles
MinimalModernBoldCreatorCinematicPodcastShort-form
Colour looks
NaturalCinematicCleanWarmCoolLuxuryDocumentary
Versions
ProfessionalCinematicHigh retentionSocial mediaShort version
AI assistant

Direct the edit in plain language

After a render you don't reopen a timeline — you say what's wrong. The assistant patches the existing editing plan and re-renders only the affected ranges.

Cutframe assistantPreview
I've watched all 42 minutes. Three usable takes of the intro — take 2 is the cleanest. Want me to lead with it?

The edit gets graded, then re-cut

After rendering, a separate evaluator inspects the result for awkward cuts, clipped audio, black frames, caption drift and story coherence. If a dimension fails its threshold, the plan is patched and only the affected ranges are re-rendered.

EditCheckDiagnosePatch planRe-renderCheck again
Professional score91
Story92
Pacing88
Audio95
Visual86
Captions97
Editing quality89

Illustrative rubric from the evaluation spec. Real scores are produced per render once the evaluator runs against a project.

A workspace built around one button

Create AI Edit is always the primary action. Everything else — projects, storage, exports, the optional manual timeline — sits behind it.

Dashboard

New AI Edit, recent projects, processing, completed, drafts, exports and storage in one view.

Project cards

Thumbnail, duration, live status, created and last-edited dates, export and quick actions.

Optional timeline

Video, audio, caption, B-roll, music, SFX and effect tracks — with “Re-edit with AI” always present.

Storage & exports

Object storage for originals, proxies, thumbnails and renders. Databases hold references, never media.

How the brain learns to edit

Training pairs raw footage with a professional final edit, its timeline and the decisions behind it — not tutorials.

Level 1
Fundamentals

Cuts, shots, scenes, continuity, rhythm, transitions, audio, composition.

Level 2
Professional editing

Best-take selection, pacing, reaction shots, B-roll, music timing, colour.

Level 3
Content types

YouTube, Shorts, vlogs, podcasts, gaming, education, documentary, ads, interviews.

Level 4
Editorial judgment

What matters, what's boring, what's emotional, what to emphasise or cut.

Level 5
Self-evaluation

The edit is scored by a separate evaluator, then revised and re-scored.

No silent training

Your footage is never used to improve models without explicit, revocable consent.

Private by default

Signed URLs, scoped access, validated uploads, rate-limited APIs and server-held secrets.

Retention control

Choose how long media is kept, delete a project and its derivatives at any time.

Honest states

Anything not yet working is labelled Research or In build. No fake buttons, no fake progress.

Research center

Architecture is decided from documented, currently available technology — speech recognition, vision and scene models, FFmpeg-based server rendering, GPU queues — not from invented APIs. Capabilities that need training data or infrastructure we do not yet have are recorded as open work.

AI models

Speech, vision, embeddings and a reasoning layer behind swappable providers.

Video tech

FFmpeg graphs, WebCodecs proxies, GPU queues and cloud rendering.

Editing theory

Film grammar, shot selection, pacing, highlight detection.

Datasets

Raw + professional-edit pairs with decision-level annotation.

Competitors

CapCut, Premiere, Descript, VEED, OpusClip, Runway, Kapwing — what they automate vs. leave manual.

Evaluation

Story, pacing, audio, visual, caption and overall professional scores, 0–100.

Questions worth asking

Do I need to know how to edit?

No. The whole point is that you upload footage, state a goal, and receive a finished edit. The manual timeline exists only for people who want to nudge the result.

Is the AI actually editing, or applying a template?

It writes a per-project timeline from your content — takes, beats, pacing curve, captions and music are decided from what's in the footage, not from a preset.

What runs today versus later?

Transcription, scene analysis, smart cutting, captions, audio repair, planning, rendering and the review loop form the MVP. Reframe, scoring and colour are in build; B-roll reasoning and sound design are research.

Where does rendering happen?

Server-side, on an FFmpeg-based render graph with background jobs. The browser only ever plays proxies and previews.

Hand it the footage. Get the film back.

Zero editing knowledge required. The AI is the editor — you're the director.

Create AI Edit