ClipFactory Request access

Local-first · Offline · Human publishes

Turn long footage into short clips, on your own machine.

ClipFactory is a local batch clipper for vertical video: transcript-first candidate selection, multi-beat FFmpeg rendering, deterministic QA gates, and Editing DNA style transfer from a reference video. No uploads, no auto-posting, no surprises.

  • 482tests in suite
  • 100%local render
  • 0auto-posts, ever
  • 9:16vertical-first

real clip record

clip_id   candidate-061
range     174.6s – 184.0s
render    8.5s · 720x1280
verdict   review_required
qa        PASS · gemmy captions

Safety promise: ClipFactory never joins a campaign, posts, submits, spends money, or mutates an account. Everything stops at local render — publishing is a human act.

What it does

Evidence in, clips out. Nothing else.

Transcript-first selection

Word-timed transcripts (local ASR with provenance) anchor every boundary to a real sentence. Proposals without transcript evidence are rejected, not guessed.

Multi-beat rendering

Clips render as structured beats — hook, body, payoff — reframed to 9:16 with captions overlaid exclusively by Gemmy Caption. Replaceable fills let you swap a beat without re-rendering the world.

Deterministic QA gates

Every render passes media checks (black frames, freezes, silence, A/V drift), caption QA, and visual QA. Outcomes are typed: publish-ready, review-required, or blocked — with reasons attached. The gates bite, including on our own showcase below.

Editing DNA transfer

Point at a reference video you admire. ClipFactory extracts its editorial behaviors, plans an edit of your raw footage in that style, validates the plan, then renders.

Gemmy Caption rails

One caption renderer, no fallbacks: HyperFrames rails (highlight, pill-karaoke, gradient-fill, editorial-emphasis, weight-shift…) overlaid by Gemmy. If Gemmy is missing, the render fails loudly instead of degrading silently.

Foundry plugin

Ships as a foundry-plugin.v1 capability pack: 8 commands (batch select/render/run, DNA extract/plan/render/assemble/variant) with schemas and typed results.

How it works

Two tracks, one principle: no evidence, no clip.

Track A · Batch clipping

  1. Inventory sources into a stable library manifest.
  2. Transcribe with word timings and provider provenance.
  3. Select a candidate bank — scored, reasoned, non-overlapping.
  4. Render multi-beat 9:16 clips, captions via Gemmy.
  5. Gate with media, caption, and visual QA.
  6. Review — you accept, hold, or reject. Nothing posts itself.

Track B · Editing DNA

  1. Extract behaviors from a reference video.
  2. Interpret them into a named, reusable style.
  3. Plan an edit of your raw footage in that style.
  4. Validate the plan against schemas and budgets.
  5. Render the styled edit, then assemble or variant the beats.

Ranking rule the engine enforces: QA status and evidence rank clips. Predicted views are never a ranking input.

Real outputs · rendered by the engine

Proof, not promises.

Everything below is genuine ClipFactory output, rendered on a local machine from a 27-minute 4K source (GTA VI Trailer 2, extended look — rights practice-only, excerpts shown to demonstrate the tool). Transcript: 2,623 words / 243 segments, provider gemmy-parakeet. Captions: Gemmy Caption exclusively. QA verdicts included unedited — including the failures.

Batch clips

Transcript-first selection picked 2 non-overlapping candidates from the full source. Both rendered 720×1280/30fps, both QA PASS, both review_required — the engine does not auto-approve.

Candidate 061 · score 0.64

174.6s – 184.0s → 8.5s render · highlight rail

“My partner's off the grid right now, you feel me?”

media QA pass caption QA pass review_required

Candidate 018 · score 0.517

36.2s – 51.8s → 5.4s render · highlight rail

“Yeah, go inside and ask for Raymond, bro”

media QA pass caption QA: 1 short cue review_required

Caption rails · one moment, five styles

Controlled variant: same 8.5s moment, only the caption rail changes (control highlight, challengers pill-karaoke / weight-shift / gradient-fill / editorial-emphasis). Two rails failed the A/V drift gate — Gemmy's karaoke and weight-shift overlays truncate ~3 frames, deterministically, twice in a row. The gate caught it; the clips stay review_required.

highlight (control)

QA pass

pill-karaoke

drift 0.090s > 0.08 gate

weight-shift

drift 0.090s > 0.08 gate

gradient-fill

QA pass

editorial-emphasis

QA pass

Editing DNA transfer

Reference: 60s slice → 12-shot event graph → interpretation (pacing, cut policy, caption policy, structure supported; zoom, b-roll unresolved). Planner grok-4.6 via local OAuth bridge: first draft rejected by the validator, repair pass accepted. The plan declines what the evidence can't support — zoom holds only, no b-roll, 4 rejected alternatives, 3 stated uncertainties. Target canvas preserved by design (DNA keeps source aspect; 9:16 is the batch track's job).

Styled edit · hook 0–18s, context 43.8–52.2s, payoff 57.0–59.8s

29.2s · captions auto/middle · validator: accept after 1 repair

plan accept uncertainties admitted: 3

How this page was made

  1. Transcribe 27 min via parakeet (chunked; engine validator merged).
  2. Select 2 candidates, transcript-first.
  3. Render 9:16 + 5 rails, Gemmy captions only.
  4. Gate media + caption + visual QA.
  5. Transfer extract → interpret → plan → render.
  6. Publish videos as release assets; page stays static.

All artifacts (candidate bank, edit plans, QA reports, planner I/O) live with the run; videos: showcase-v1 release.

Editing DNA

Steal like an editor, not like a thief.

Analyze a reference video into a reusable style — pacing, caption behavior, inserts, reframes — then apply it to your own raw footage. Three styles ship as examples (see one working above):

Hormozi

  • Short middle captions
  • Punch / reframe emphasis
  • Visual bridge retention
  • Faster pacing

Camilo

  • Short middle captions
  • Supporting inserts under VO
  • Faster pacing

Persona

  • Visual bridge retention
  • Faster pacing
python -m clipper_engine.cli style-analyze --source reference.mp4 --name MyStyle
python -m clipper_engine.cli style-list
python -m clipper_engine.cli edit-generate --style STYLE_ID --target raw.mp4 --out ./edit

Quickstart

On your machine in minutes.

Prerequisites

  • Python 3.12
  • ffmpeg + ffprobe on PATH
  • gemmy on PATH for ASR and caption overlays (typed error without it — never a silent fallback)
  • Optional: local LLM lane (Hermes proxy + CLIPFACTORY_LLM_ENDPOINT) for the DNA planner

Install & run

pip install -e .
python -m clipper_engine.cli batch-run --help
python -m clipper_engine.cli app        # desktop app, no quotas
python -m editing_dna --help            # style-transfer track

The engine source is in private beta — request access and the same commands work after checkout.

Platform

A Foundry plugin, not a black box.

ClipFactory ships as a foundry-plugin.v1 capability pack. Eight typed commands, JSON schemas for every input and output, structured events and results — spawned exactly like the host runs them and covered by a conformance suite.

CommandDoes
clip.batch-selectTranscript-first candidate bank across a batch
clip.batch-renderRender + QA the selected candidates
clip.batch-runSelect, render, and QA end to end
dna.extractReference video → validated event graph
dna.planPlanner input → validated edit plan
dna.renderAccepted plan + sources → styled video
dna.assembleReplaceable beat-list assembly
dna.variantSwap one beat, keep the body

Safety

The line it will not cross.

Never

  • Join a campaign
  • Post or submit anything
  • Spend money
  • Mutate an account

Always

  • Stop at local render
  • Attach reasons to every decision
  • Gate output: publish-ready, review-required, or blocked
  • Leave publishing to a human

FAQ

Asked, answered.

Does ClipFactory post to TikTok for me?

No. It renders video files and evidence on your machine and stops. Uploading, posting, and submitting are human acts it cannot perform.

Does it need the cloud?

No. Selection, rendering, and QA run locally with FFmpeg, Gemmy (local GPU), and deterministic validators. The DNA planner can point at any OpenAI-compatible local lane; the showcase transfer above was planned by Grok through a local OAuth bridge — still no provider keys in the repo.

Are the videos above real?

Yes. Real engine outputs from a 27-minute source, with real transcripts, real caption rails, and real QA verdicts — failures included. Videos stream from public release assets; this page itself is static and collects nothing.

How is ranking kept honest?

Clips rank by QA status and evidenced score. Predicted views are never a ranking input, and financial summaries never treat raw views as cash.

Where is the source code?

The engine is in private beta while the platform matures. Request access here — this page and its evidence are public and open.

Your footage. Your machine. Your call.

Watch the outputs, then request access to the engine when you're ready.