SceneForge Studio

A free, open-source desktop editor for narrated documentary videos. Turn a script, photos, clips and a voice into a finished film one scene at a time. You get film looks, captions that follow the narration, a multitrack audio timeline, stickers, titles and optional AI tools. Arabic and English are first-class.

Install · Highlights · Feature tour · AI providers and prices · Run from source · Tests · Version history · Licenses · Contributing

Release status. 0.8.0 is the latest release: the creator pack (silence and filler removal, auto-reframe to 9:16, GPU export, live playback, keyframes), on top of 0.7.1 (security and usability) and 0.7.0. 0.7.0 was a large update over 0.5.3: a multitrack audio timeline with pro edit tools, 71 caption styles, 400 stickers and emoji, per-effect settings, AI subject cutout and voice isolation, and optional text-to-video. Items marked (0.7.0) are new in this version. The full record is in desktop/RELEASE_NOTES.md and PROJECT_HISTORY_AND_HANDOFF.md.

The screenshots on this page come from the 0.7.0 build. They show a demo project made only from procedurally generated pictures, video and audio. scripts/capture_readme_screenshots.py rebuilds the project and the screenshots.

Highlights

• Scene-based editing with a real timeline. Pictures are edited per scene on V1, with text on T1, narration on A1 and clip sound on A2. Six independent audio tracks (A3–A8) take free audio clips.
• Pro edit tools for timeline audio: Select, Track select, Ripple, Roll, Slip, Slide and Blade, plus Razor All. There are also clip groups, colour and range markers, beat markers, volume envelopes and right-click menus. (0.7.0)
• 71 caption styles in six groups, with word-by-word timing and correct Arabic right-to-left layout. Captions can come from local Whisper at no cost, or from optional cloud services.
• Effects with per-effect settings: 36 presets have their own controls (0.7.0). There are film grades, halation, VHS, lens and camera effects, an effect stack with ordering and bypass, and drag-and-drop look preset packs.
• 400 stickers and emoji (0.7.0), textured titles, and subject cutout for photos and moving video, which puts text behind a person.
• Audio finishing: AI voice isolation, dialogue cleanup, beat-aware music fitting, beat sync, automatic ducking under narration and loudness levelling to −14 LUFS.
• Optional AI generation. Text-to-video runs locally through ComfyUI or in the cloud through Google Veo and Runway. Every cloud request shows a cost estimate and needs your confirmation. Image Studio supports several providers.
• Creator pack (0.8.0):
  • remove silences (jump cuts) and filler words;
  • auto-reframe to vertical 9:16;
  • GPU (NVIDIA NVENC) export;
  • live timeline playback;
  • keyframe animation for stickers, overlays and titles.
• Friendlier editing (0.9.0): a dedicated Typewriter box with sound preview, an Effects tab sorted into clear groups, waveforms on every audio lane, auto-ducking for A3–A8, automatic restore points, a sample project, caption translation to English (or bilingual) and title templates.
• Exports to YouTube, Shorts/Reels/TikTok, Instagram, ProRes, GIF, MP3/WAV and SRT/VTT. Rendering uses FFmpeg on your own computer.

Feature tour

Editor and timeline

• Tracks. T1 holds captions and titles. V1 holds the scene pictures, with transitions overlapping. A1 holds the accepted narration, A2 the embedded sound of each video clip, and A3–A8 independent audio clips. A whole-project music bed sits underneath. Each audio track has mute, solo and lock, and is dimmed when it will not be heard. (A4–A8: 0.7.0)

• Edit tools for timeline audio (0.7.0):

  Key	Tool
  V	Select
  A	Track select forward
  B	Ripple trim
  N	Roll
  Y	Slip
  U	Slide
  C	Blade
  Shift+C	Razor All

  Delete removes only the selected piece. Shift+Delete ripple-deletes it. Ctrl+C, Ctrl+X, Ctrl+V and Ctrl+D also work, and every edit can be undone.

• Markers are saved with the project. A marker can be a point or a range, in five colours, and can be attached to a clip so it moves with the clip. ♪ Beats… finds the beat of the music bed or of a selected clip and marks every beat, every 2nd beat or every bar. Cuts and clips snap to the beat markers. (0.7.0)

• Volume envelopes take up to 32 keyframes on an A3–A8 clip, with presets such as dip under a voice. The line is drawn on the clip and heard in scene renders and in the export. Clip groups select and move together: Ctrl+G groups, and Ctrl+Shift+G ungroups. (0.7.0)

• Right-click menus for audio clips, scenes, narration and clip sound. From these menus you can split, duplicate, move a clip to another track, add a marker to the clip, or detach narration or clip sound to A3 so you can cut it. (0.7.0)

• Everything that existed before: drag-and-drop of files and whole folders, 28 curated transitions with live previews, split and ripple delete, multi-select with batch apply, snapping, linked trimming of source video and audio, Ctrl+wheel zoom around the pointer, and undo/redo.

• The design notes are in docs/TIMELINE_ARCHITECTURE.md. V1 is still scene-based. Free multitrack video and nested clips are not implemented yet.

Captions and text

• 71 caption styles in the groups Social, Cinematic, Karaoke, Fun, Minimal and Arabic, with search (0.7.0: 40 new). You can also set the font, colours, outline, shadow, box, position, phrase or word-by-word splitting, karaoke highlighting and animation.
• Automatic captions from narration or video sound:
  • Local Faster-Whisper (the base model, about 145 MB) downloads once on first use. After that it runs offline on the CPU, with no key.
  • ElevenLabs Scribe and OpenAI Whisper are optional cloud services.
  • A Check local captions panel diagnoses and repairs the local engine.
• Title layers (Text, Text Box, Text+) take more than 25 animations. The reveal set includes from-right, up, down, split, clock wipe, counter-clockwise clock and iris (the last three: 0.7.0). There are also typewriter text with sound and per-letter and per-word animations.
• Arabic is rendered right-to-left, with Arabic caption styles and fonts (Noto Naskh, Amiri, Tajawal and Lalezar).

• Typewriter box (0.9.0): the caption typewriter, its speed and timing, and its keystroke sound (included or your own recording) in one place, with Preview sound. Also new: caption translation to English or bilingual captions, and title templates (lower third, intro, outro, chapter, quote).

Effects and looks

• Per-effect settings (0.7.0). 36 presets have 132 settings between them. Examples are warmth, tint and fade for colour looks; bleed, noise, scanlines and tracking for VHS; and threshold, radius and colour for halation. Settings are remembered per effect, and default settings render exactly as before.

• Looks: 35 scene looks and film treatments, 8 colour filters (Golden hour, Arctic, Portra film and others), and Chromatic split and Motion trail. Effects pack (0.7.0):

  • camera shake presets, lens flare and wiggle;
  • focus blur, tilt-shift and mosaic;
  • halation, 2383-style print, tungsten night and cross-process.

  All of these are rendered by FFmpeg. None of them are AI.

• Colour tools: sliders, colour wheels, split toning, .cube LUT import (3D, 1D and Resolve shaper LUTs), old film, spotlight, light leaks, and blur or pixelate to hide faces.

• Effect stack (0.7.0). Spotlight, light leak, lens flare, wiggle and camera shake can be reordered or bypassed. The order changes the look: shake before flare keeps the flare steady on the lens.

• Look preset packs (0.7.0):

  • Save the current look as a preset.
  • Drag a .json or .sflook pack onto the Effects tab to import it.
  • Export your own presets as a pack.
  • An 8-preset Starter pack is included.

  See docs/EFFECTS_GUIDE.md.

Stickers, titles and cutout

• 400 stickers and emoji in 14 categories, with search and a Recent tab (0.7.0). 322 are colour emoji from Twemoji (CC-BY 4.0). 78 are graphics drawn for SceneForge: subscribe and like buttons, badges, stamps, arrows, callouts, shapes and speech bubbles. The library is bundled, so it looks the same on every computer. Stickers can loop with float, bob or pendulum motion.
• Picture-in-picture, split-screen layouts, animated map routes, 3D photo (parallax), Ken Burns, restoration of old photos, and animated annotations. These were already in earlier releases.

• Textured titles (0.7.0) fill big letters with a texture. The texture can be one of 9 built-in patterns (lava, neon, gold, chrome, marble, ice, fire, pixel blocks, galaxy), any image in your Media Pool, or a texture generated from a prompt by your image engine. Cloud image engines may charge for that. Titles support glow, outline, 8 fonts and Arabic.

• Video inside text (0.9.1): a solid colour card with your title cut out of it, so the scene's video or picture plays inside the letters (the documentary "1942" / place-name look). Colour, opacity, font, size, position and an outline are adjustable.

• Subject cutout (AI, beta) (0.7.0) runs locally with onnxruntime. It has two uses:
  • Put text behind subject: captions and titles pass behind a person or object.
  • Remove background → Media Pool: saves a transparent PNG.
• Cutout options:
  • Edge refinement, a ±10 px edge grow/shrink, and a result preview on a checkerboard.
  • Models: U²-Net small (~4.6 MB, bundled with the installer), IS-Net (~170 MB, best general quality) and People (U²-Net human, ~176 MB). The two larger models download once, on first use, and are checked against a checksum.
  • Moving-video cutout works on one clip of up to 20 s, with a static camera and normal speed. It renders a VP9 alpha layer in the background, with progress and Cancel.

Audio

• AI voice isolation (0.7.0) uses a local MDX-Net model (Kim_Vocal_2, ~67 MB, downloaded once and checked against a checksum). It is available for narration takes, timeline clips and video clip sound. A strength slider controls it, and your original files are never changed. In a synthetic test it raised voice SDR from 0 to 7.75 dB.
• Dialogue cleanup (noise reduction, not AI), 1940s radio and telephone voice effects, and trim, volume, fades and waveform editing for narration.

• The music bed loops and ducks under narration automatically.
• Fit music to video length (0.7.0) is a beat-aware re-edit, not AI. It removes or repeats whole bars.
• Sync scene cuts to the beat, loudness levelling for YouTube (−14 LUFS) and a film countdown leader are also available.
• When you play a rendered scene, you also hear the timeline audio and the music under it. The rendered part itself stays clean, so the export mixes each clip only once.

AI generation (optional)

• Text-to-video. Run local ComfyUI workflows (LTX-2.5 Fast, Wan 2.1 T2V 1.3B, Wan 2.2 TI2V 5B, Wan 2.2 T2V A14B, or a custom API-format workflow), or use Google Veo 3.1 or Runway (Gen-4.5, WAN 3.0) with your own key. A step-by-step setup guide and an optional GPU/VRAM check help you choose a local model.
• Cost control. Cloud requests show the provider's published per-second price and an estimated total for the 1–3 takes you request. Generation starts only after an explicit cost confirmation, which both the UI and the backend check. You can compare the finished takes and add the one you choose to a scene.
• Image Studio generates 1–3 options with OpenAI, Google Gemini, Cloudflare Workers AI, Hugging Face, or local Stable Diffusion (AUTOMATIC1111).
• ComfyUI, model weights, custom nodes and workflow files are not bundled or downloaded by SceneForge. Prices and models are listed under AI providers and prices.

Local voices (Chatterbox and Kokoro)

• Chatterbox (Arabic and multilingual) and Kokoro give natural narration with no provider key. They run as local services in Docker Desktop in Linux-containers mode.
• The installer includes the service files. Start Chatterbox and Start Kokoro buttons detect a missing or stopped Docker, show a live log and connect when the engine answers.
• The first start downloads container images and models, several GB for Chatterbox. Chatterbox uses CPU-only PyTorch. Chatterbox has not yet been confirmed working on a Windows test PC. Kokoro and the cloud voices are the fallback.
• Cloud voices: ElevenLabs (with exact word timing) and Together AI. You can also upload your own recordings.

Export and share

• Platform presets: YouTube (1080p and 4K), TikTok/Reels/Shorts (needs a 9:16 project), Instagram, a small H.265 file, ProRes for editing, animated GIF, and audio only.
• Advanced settings: resolution up to 4K, frame rate, format (MP4/H.264, H.265, WebM, ProRes, GIF, MP3, WAV) and quality.
• SRT/VTT caption files are exported alongside the burned-in captions. The sound of each video clip is kept and can be adjusted per clip.
• The Share dialog is local-first. It offers a download, opens the file's folder in the desktop app, and gives upload steps for YouTube, TikTok, Instagram and Facebook. It does not upload to your accounts.

Preferences and themes

• Themes: Graphite Night (the default), Daylight, Midnight Blue and Warm Studio. Preferences are under File → Preferences and include an accent colour, panel spacing and reduced motion. The theme menu is also at the top right of the editor.
• You set the project format (16:9, 9:16, 1:1) and frame rate in Create project.
• The desktop close dialog offers Save, Don't save, Save and exit, and Cancel.
• Projects live in your user folder. The database is backed up automatically before a new version opens it (Help → Open project backups, last 5 kept).

Install

Installers are on the Releases page. The current version is 0.8.0.

Testing builds. Every push to claude/0.7.0-* (or a manual Run workflow) runs the Release workflow. That run produces test installers as temporary GitHub Actions artifacts. Public releases are created only from v* tags.

Updates. On Windows and Linux (AppImage), SceneForge downloads new versions in the background and asks to restart. On macOS it tells you about a new version and opens the download page. Help → Receive beta updates opts you in to test versions.

What is bundled and what downloads on first use

Downloaded models are kept in %USERPROFILE%\.sceneforge\models\{whisper,cutout,voice} (or ~/.sceneforge/models/…). Provider keys are stored in the operating system's credential store (Windows Credential Manager, macOS Keychain, Linux Secret Service), never in project files.

AI providers and prices

AI is optional. You can build a complete video from your own photos, clips and recordings without any account.

Text-to-video prices

This table shows the provider's published USD rate per generated second, as listed in the app's catalog. The prices were checked on 2026-09-29 and can change. Provider billing, taxes, credits, retries and regional terms can change the final charge. Local generation has no API fee, but it uses your hardware and electricity. Check each model's license before you publish or use its output commercially.

Price references: Google Veo pricing, Veo API options, Runway API pricing, Runway API changelog. Key setup: Google AI Studio API keys (instructions), Runway API setup (Dev console).

Local setup notes

• Install and start ComfyUI yourself, then connect its loopback address under Settings → Providers.
• In ComfyUI, load the model's workflow and export it with File → Export Workflow (API). Older versions call this Save (API Format). Import the JSON into Generate video.
• Custom ComfyUI nodes run inside ComfyUI with your user permissions. Only import workflows and nodes you trust.
• Custom local dimensions must be multiples of 16.
• On an 8 GB GPU, start at 480p with Wan 2.2 5B (ComfyUI's native offloading) or the lighter Wan 2.1 1.3B. See the Wan 2.2 setup and the LTX ComfyUI setup.
• Takes are generated one after another. Each finished take is saved in the Media Pool, and completed takes are kept if a later one fails.
• Vast.ai remote GPUs are only a proposal and are not connected to the app. See docs/VAST_GPU_RELAY_PROPOSAL.md.

Run from source

Requirements:

• Python 3.12. The backend pins numpy==2.5.3, which needs 3.12.
• Node.js 20 or later. CI uses 22.
• FFmpeg on your PATH.

The desktop installer bundles its own Python runtime and FFmpeg. You only need these tools to run from source.

Windows

git clone https://github.com/EzioDEVio/SceneForge.git
cd SceneForge
scripts\setup.bat
scripts\start.bat          :: serves the built app at http://127.0.0.1:8000
scripts\dev.bat            :: or: backend with --reload on :8000 + Vite dev server on :5173

macOS / Linux

git clone https://github.com/EzioDEVio/SceneForge.git && cd SceneForge
python3.12 -m venv .venv && .venv/bin/pip install -r backend/requirements.txt
npm --prefix frontend ci && npm --prefix frontend run build
cd backend && ../.venv/bin/python -m uvicorn app.main:app --host 127.0.0.1 --port 8000
# open http://127.0.0.1:8000 (FastAPI serves frontend/dist on the same origin)

For frontend work, run npm --prefix frontend run dev. Vite then serves on port 5173 and proxies /api to the backend on port 8000.

Useful environment variables:

Build the desktop installers yourself

pip install -r desktop/requirements-build.txt
npm --prefix frontend ci && npm --prefix frontend run build
npm --prefix desktop ci
python desktop/scripts/fetch_ffmpeg.py        # downloads and verifies FFmpeg for your OS
python desktop/scripts/fetch_models.py        # the bundled cutout model (u2netp)
python desktop/scripts/build_backend.py       # bundles the Python backend (PyInstaller)
npm --prefix desktop run dist:win             # or dist:mac / dist:linux

Tests

npm --prefix frontend run build && npm --prefix frontend test   # editor, units, timeline, video-generation and share-dialog checks
ELECTRON_SKIP_BINARY_DOWNLOAD=1 npm --prefix desktop ci && npm --prefix desktop test
python tests/integration/test_combined.py                       # each integration test is a standalone script

• Build and render checks (.github/workflows/checks.yml) run on every push. They run the frontend build and tests, then 32 Python integration scripts that render real media with FFmpeg.

• On Linux, the credential tests need a Secret Service keyring. Run them under dbus-run-session with gnome-keyring-daemon unlocked, as CI does.

• Tests never call paid providers. Those calls are mocked.

• Last full local run (RC9):

  Suite	Checks
  Frontend component checks	264
  Unit checks	19
  Timeline checks	40
  Video-generation flows	2
  Share-dialog checks	6
  Desktop checks	18
  Integration scripts	32

  All passed. This does not replace testing on Windows.

Version history

The table summarizes the documented history. Versions marked WIP or RC were development snapshots, not public releases. Details are in desktop/RELEASE_NOTES.md and handoff.md.

Releasing (maintainers)

1. Set the version in desktop/package.json, for example 0.7.0, or 0.7.0-beta.1 for a beta.
2. Commit, then tag and push: git tag v0.7.0 && git push origin v0.7.0.
3. The Release workflow tests everything and builds the Windows, Linux and macOS installers. It attaches them to a draft release.
4. Review the draft and click Publish release. Users' apps offer the update only after that.

Licenses and credits

SceneForge Studio is free software. You can redistribute it and modify it under the terms of the GNU General Public License v3.0 or later; see LICENSE. It is provided WITHOUT ANY WARRANTY.

• Bundled components keep their own licenses: FFmpeg (GPL builds), Electron, Python libraries, and fonts under the SIL OFL. See desktop/THIRD_PARTY.md.
• Emoji: Twemoji v17.0.3 graphics, Copyright Twitter, Inc. and other contributors, licensed CC-BY 4.0 (assets/stickers/LICENSE-TWEMOJI-GRAPHICS.txt). The 78 graphic stickers were drawn for SceneForge (GPL-3.0-or-later).
• Cutout models are ONNX files published with rembg (MIT). According to backend/app/render/cutout.py, IS-Net general use and U²-Net small are Apache-2.0. The People model (u2net_human_seg) comes from the same rembg release; its licence is not stated in the code.
• Voice isolation model: Kim_Vocal_2 (MDX-Net, by Kimberley Jensen) comes from the Ultimate Vocal Remover model repository. UVR and python-audio-separator are MIT-licensed. The weights have no explicit licence from their author, and this has not been confirmed. Check it before you rely on the feature commercially. It must be resolved before a public release.
• Other AI models and services (Whisper models, Chatterbox, Kokoro, Stable Diffusion, ComfyUI models, cloud providers) are not part of SceneForge. They have their own terms, which may restrict commercial use.

Copyright © 2026 EzioDEVio and SceneForge contributors.

Contributing and security

• Read CONTRIBUTING.md before opening a pull request. Keep database changes additive, add a real-render test for rendering changes, and never commit keys or personal media.
• Report vulnerabilities privately, as described in SECURITY.md.
• PROJECT_HISTORY_AND_HANDOFF.md and HANDOFF_TO_CHATGPT_RC9.md contain the implementation inventory, decisions and open items.