01
Drop it in
A film, an episode, or a pile of episodes. Pick Live action or Anime and one of three quality levels. Splitcast tells you roughly how long it will take before you commit.
free · open source · runs offline
One film in.
One cast per character out.
Drop in a film, an episode or a whole season. Splitcast finds every person who appears, groups their shots, and exports each character’s scenes as clean clips — ready for your edit. The AI runs on your own graphics card. Nothing is uploaded.
Windows 10/11 · NVIDIA GPU · 16 languages
Splitcast — Tears of Steel
click a character
Real screenshot. Footage: Tears of Steel (CC-BY, Blender Foundation).
offlineyour GPUno accountfilmsserieswhole seasonsanime modeframe-accurateH.264 · HEVC · ProRes16 languagesopen sourcefree
offlineyour GPUno accountfilmsserieswhole seasonsanime modeframe-accurateH.264 · HEVC · ProRes16 languagesopen sourcefree
/how-it-works
Three steps. The middle one is the computer’s.
02
It finds the cast
Shots are detected, faces are tracked through each shot, and tracks are grouped into characters. Two faces in the same shot are never merged. You get a list, ranked by screen time.
03
Pick, trim, export
Click through scenes, set In and Out to the frame, then export as separate clips or one file. Copy the original stream, or re-encode with H.264, HEVC or ProRes.
/review
Review fast. Trim to the frame.
- Space plays, I and O set the cut, ← → change scene.
- The strip under the grid shows where the character appears in the film.
- Filter by shot size (close, medium, wide), solo shots only, minimum length.
- Wrong match? Move scenes to another character, or merge two characters.
/features
Made for people who cut things together.
Seasons, not just films
Add every episode to one project. A character keeps the same identity from episode to episode, so you get one scenepack per character for the whole season.
Anime mode
Drawn faces are a different problem. A dedicated model looks at hair, eyes and outfit instead of facial geometry.
Pick your speed
Three quality levels. Slower means fewer mix-ups between similar faces.
× real time, 1080p, RTX 5060. Decoding uses your CPU too.
16 languages
Including right-to-left layout for Arabic.
Azərbaycanca
Cuts that land on the frame
Copy the original stream for instant, lossless cuts, or re-encode when you need frame accuracy. NVENC is used when your card has it.
Nothing leaves your PC
No account, no cloud, no telemetry. The installer carries everything, so analysis runs offline from the first minute.
0real-time speed on an RTX 5060, 1080p (Balanced)
0interface languages
0characters found in the 5-minute demo film
0bytes uploaded while it works
/under-the-hood
What actually runs, and what to expect.
- shotsTransNetV2 finds the cuts (or a content-adaptive detector on Fast)
- facesSCRFD detects faces on your GPU through ONNX Runtime + CUDA
- identityArcFace embeddings; tracks grouped with cannot-link inside a shot
- animeYOLOv8 anime face detector + CCIP character embeddings
- seasonscharacters matched across episodes by several “looks” each
Things worth knowing
- Windows + NVIDIA for now. It uses CUDA. Other GPUs are on the list, not in the box.
- Balanced models are inside the installer. Fast, Maximum and the anime models download once, the first time you pick them.
- Characters are numbered, not named. It does not know who an actor is. You rename them.
- Anime background extras can mix. Leads are reliable; crowds are not.
- Face models have their own licences. InsightFace weights are for non-commercial research use; see the repository for details.
/install
Download, double-click, done.
- Download Splitcast-Setup.exe from the latest release. It is about 2 GB because everything is inside: the app, the AI runtime, ffmpeg and the models.
- Run it. No admin rights, no Python, no internet needed while installing.
- Open Splitcast from the Start menu and drop in a film.
Windows may show a “SmartScreen” notice because the installer is not code-signed yet. Choose “More info → Run anyway”.
Prefer building from source?
# needs Python 3.12, Node 20+ and an NVIDIA driver git clone https://github.com/elcann7/splitcast cd splitcast .\setup.ps1 .\ScenepackStudio.bat
/faq
Questions people ask.
Is it really free?
Yes. Open source, no account, no paid tier. If it saves you time, a star on GitHub is the nicest thanks.
Does my footage leave my computer?
No. Analysis and export run locally. The only network traffic is a one-time model download, when you pick a quality level whose models are not installed yet.
Which graphics cards work?
NVIDIA cards with CUDA support. It was developed on an RTX 5060. Without a supported GPU it will fall back to the CPU, which is much slower.
Can it tell me the character’s name?
No, and it does not try. Characters appear as “Character 01, 02…” ranked by screen time. Double-click to rename.
What about animated films and anime?
Choose Anime when you create the project. It uses a model trained for drawn characters. Leads work well; background crowds may get mixed.
Can I cut a whole season?
Yes. Add all the episodes to one project; each character is matched across episodes, so you export one scenepack per character for the season.
Is it legal to use on films I don’t own?
Splitcast is a tool; what you cut and what you do with it is your responsibility. Check the rights to the footage you use.