slopstudio

A high-performance, local-first AI video-generation NLE — a timeline editor where each row is one generation pipeline (TTS narration, an animated anime host, reaction images, code cards, diagrams, music, effects), composed in a GPU render graph with a latency-obsessed live preview and a deterministic record/export.

Why this exists

slopstudio is the ad-hoc video pipeline behind Gemma-san Explains (@GemmaExplains) — a channel of technical deep-dives narrated by a chuuni succubus-cosplaying cosmic architect named Gemma-san. Every video on that channel is authored, generated, and rendered with this tool. It is built for one format done well, not as a general-purpose editor.

It is open-sourced (MIT) for anyone who finds it useful — whether that's for making your own narrated explainers, as a reference for a latency-obsessed compositing NLE, or as an example of a pipeline a frontier LLM can drive end-to-end. There is no product, no account, no cloud: it runs on your own hardware against your own models.

Every frame above is composited by the editor from a single project file: a host sprite with viseme lip-sync, syntax-highlighted code that types itself in, labelled diagrams, sourced B-roll, and captions — each on its own timeline row, each a cached generation or a pure-GPU native clip.

Why it's fast

Two layers, deliberately separated:

• Compositing layer (instant). Timing, transforms, effects, transitions, and layout are pure GPU compositing over a pull-based render graph with a param-hashed GPU texture cache. Changing one of these and seeing the result is sub-frame — that is the whole point.
• Generation layer (async, cached). AI generation (image/video/voice/music) is slow, so it never sits in the interactive loop. Generations run as background jobs producing content-addressed cached assets; editing a generation parameter re-queues a job while the preview keeps showing the last result plus a "regenerating" badge.

Shape

  ┌──────────────────────────────┐         ┌──────────────────────────────────┐
  │  editor  (Windows-native)     │  HTTP/  │  providers  (the GPU host)        │
  │  C++ · Dear ImGui · D3D11     │   WS    │  Python services, one per model   │
  │                               │ ──────► │                                   │
  │  timeline of pipeline rows    │         │  tts   image  video  music  align │
  │  pull-based render graph      │ ◄────── │  (Qwen3-TTS, ComfyUI/Anima,       │
  │  live preview + record/export │ assets  │   Wan2.2, ACE-Step+Jamendo, …)    │
  │  HTTP control API (LLM-drive) │         │                                   │
  └──────────────────────────────┘         └──────────────────────────────────┘
       runs on the workstation                  defaults to `lame` (config-driven)

The editor is a Windows PE cross-compiled from the Nix flake (mingw-w64 + the Dear ImGui DX11 backend) and runs natively on Windows via WSLInterop — no WSLg performance tax. Providers are out-of-process so a crashing model can never take down the editor, and their endpoints live in config.toml so the whole thing is redeployable on different hardware by editing one file.

Build & run

Everything is provided by the flake; prefix host commands with nix develop --command.

nix develop                          # dev shell (python providers + mingw editor toolchain)
cp config.example.toml config.toml   # then edit endpoints/keys (config.toml is gitignored)

nix develop --command make -C editor            # build the editor → build/slopstudio.exe
nix develop --command python -m providers.tts   # run a provider locally (or deploy to a GPU host)

./build/slopstudio.exe examples/signature-opener.slop.json    # open a project (Windows / WSLInterop)

First run — stock assets. The editor ships no art: the stock host rig (Gemma-san's sprite sheet) and the shared backgrounds are a separate stock-assets.zip release asset, fetched and unpacked into presets/ on first launch. A fresh checkout plus that download is everything the bundled example projects need to render. Full video projects live in their own repo as portable project dirs (<name>.slop.json + assets/) and are opened the same way — asset uris resolve relative to the project file.

Prebuilt editor. Nightly cross-compiled slopstudio.exe binaries are published to the rolling nightly pre-release.

Render a video

# 1080p60, space-efficient x264 sized to ~300 MB (YouTube re-encodes anyway):
nix develop --command bash tools/export.sh path/to/project.slop.json \
    --scale 1080 --fps 60 --target-mb 300 --out exports/project.mp4

Or use File ▸ Render video… in the editor. The render is deterministic: the editor streams the full-res composite to ffmpeg, which encodes the video and muxes the audio assembled from the project's export plan (clip offsets, gains, sourced-music credits).

Author a video as an LLM

The project file (*.slop.json) is the human- and machine-editable source of truth. An agent composes a cut with tools/slop.py (a beats → skeleton → lint → render workflow) and generates voice/visemes headlessly — see docs/LLM_WORKFLOW.md and docs/PROJECT_FORMAT.md.

Models (all commercial-safe — this is MIT + monetized output)

Role	Model	License
TTS / voice design	Qwen3-TTS 1.7B VoiceDesign	Apache-2.0
Host sprite art	cloud GPT image gen → sprite rig (rembg matte + pose tools); local Anima 2B + character LoRA as backup	outputs owned / weights NC · outputs OK
Backdrops / anime stills	Anima 2B	weights NC · outputs commercial-OK
Diagrams / legible text	Qwen-Image	Apache-2.0
Image/video backend	ComfyUI	GPL-3.0 (separate process)
Image→video motion	Wan 2.2 (procedural 2.5D default)	Apache-2.0
Music gen / sourcing	ACE-Step 1.5 / Jamendo (CC-BY)	MIT / per-track
Avatar	pngtuber sprite state machine (Inochi2D planned, not yet supported)	—
Word timing / visemes	WhisperX / Rhubarb Lip Sync	BSD / MIT

See docs/RESEARCH.md for the full mid-2026 survey, the runner-ups, and the licensing rationale.

Documentation

docs/ARCHITECTURE.md	how it all fits together
docs/LLM_WORKFLOW.md	compose a video as an agent
docs/VIDEO_RUNBOOK.md	compose and finish a video by hand
docs/SCENE_COOKBOOK.md	reusable animated scenes for humans and agents
docs/PROJECT_FORMAT.md	the *.slop.json schema
docs/PROVIDER_PROTOCOL.md	how the editor talks to models
docs/RESEARCH.md	model survey + licensing rationale

License

MIT — see LICENSE. Generated content is yours; sourced music/art carry their own licenses, which slopstudio tracks and turns into automatic on-screen credits.