CobbleGen turns Reddit-style story text into vertical short-form reels. It ships with a FastAPI web studio for pasting stories, choosing narration voices, uploading gameplay footage and music, watching live generation progress, and browsing finished reels with copy-ready metadata.
The repository intentionally does not include private stories, rendered output videos, temporary files, voice previews, footage, music, .env, or local pipeline state. Drop your own assets into the folders after cloning.
- Browser studio for story input, asset uploads, voice preview, settings, job progress, and reel library
- CLI pipeline for one story or batch processing a
stories/folder - Ollama story analysis with host failover and per-host model selection
- NVIDIA Magpie/Riva narration with emotional voice variants and chunk stitching
- Edge TTS fallback for simpler local runs
- Karaoke-style ASS subtitles burned into the final video
- Optional Unsplash scene images and captions
- Drag-and-drop background footage and music
- FFmpeg rendering with NVENC preference when available
- Python 3.11 or newer
- FFmpeg and FFprobe available on
PATH - Ollama, either on your PC or reachable on another host
- At least one vertical-compatible background video in
footage/ - Optional music files in
music/ - For NVIDIA narration: an NVIDIA API key with Riva/Magpie access
- Optional Unsplash API key for contextual scene images
Install Ollama from https://ollama.com/ and pull or expose the model you want to use. Cloud-backed Ollama models work too as long as your configured Ollama host serves them.
git clone https://github.com/Dr4cule/CobbleGen.git
cd CobbleGen
python -m venv .venv
# Windows PowerShell:
.venv\Scripts\Activate.ps1
# macOS/Linux:
# source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .envThen edit .env.
For a single local Ollama instance:
OLLAMA_BASE_URL=http://localhost:11434
OLLAMA_MODEL=nemotron-3-super:cloudFor a master/slave setup or multiple remote Ollama hosts:
OLLAMA_HOSTS=http://master:11434|nemotron-3-super:cloud,http://slave:11434|qwen3.6:27bEach OLLAMA_HOSTS entry uses url|model. CobbleGen tries them in order and falls back automatically if one host is unavailable. If OLLAMA_HOSTS is not set, it uses OLLAMA_BASE_URL plus a default slave fallback.
Pick a narration engine per reel from the TTS engine dropdown in the Create tab, or set the default with TTS_BACKEND:
| Engine | TTS_BACKEND |
Notes |
|---|---|---|
| NVIDIA Magpie | nvidia_magpie |
Fast, 400+ voices, per-script emotion. Unlimited. |
| ElevenLabs | elevenlabs |
Most natural & consistent; character-accurate karaoke timing. Free tier ~10k chars/month. |
| Edge TTS | edge_tts |
Free Microsoft voices, offline-friendly fallback. |
NVIDIA Magpie:
TTS_BACKEND=nvidia_magpie
NVIDIA_API_KEY=your_key_here
NVIDIA_TTS_VOICE=Magpie-Multilingual.EN-US.Mia
# Emotion mode: dominant (consistent voice for the whole reel, default) |
# dynamic (switch per chunk) | off
NVIDIA_EMOTION_MODE=dominant
NVIDIA_TTS_NORMALIZE_CHUNKS=true # match loudness across stitched chunksdominant mode picks one emotion for the entire reel so the voice timbre never jumps between chunks (the usual cause of "inconsistent" Magpie audio), while still matching the story's mood. Chunks are loudness-normalised before stitching so volume stays even.
ElevenLabs (key needs only the Text to Speech permission; voice listing/quota are not required):
TTS_BACKEND=elevenlabs
ELEVENLABS_API_KEY=your_key_here
ELEVENLABS_VOICE_ID=JBFqnCBsd6RMkjVDRZzb # George (free-tier premade voice)
ELEVENLABS_MODEL=eleven_multilingual_v2Reels are sent to ElevenLabs in a single request whenever they fit (almost always), which gives the most consistent voice and exact word timing with no glitches. Free-tier premade voices include George, Brian, Bill, Sarah, Charlotte, and more — all selectable in the UI.
Edge TTS (lighter fallback):
TTS_BACKEND=edge_tts
TTS_VOICE=en-US-GuyNeuralPlace your own files here:
footage/ gameplay or background videos, such as .mp4 or .mov
music/ optional background tracks, such as .mp3 or .wav
stories/ optional .txt story files for CLI or existing-file web runs
These folders are ignored by Git except for .gitkeep, so your private assets and generated media stay local.
python -m webapp.runOpen http://127.0.0.1:8000.
Useful variants:
python -m webapp.run --port 9000
python -m webapp.run --host 0.0.0.0 --port 8000Process one story:
python main.py --story stories/my_story.txtProcess every unprocessed .txt file in stories/:
python main.py --allForce reprocessing:
python main.py --story stories/my_story.txt --reprocessList voices for the configured TTS backend:
python main.py --list-voicesReset local processing history:
python main.py --reset-footage
python main.py --reset-storiesFinished reels and metadata are written to output/:
output/
story_stem_safe_title.mp4
story_stem_safe_title_meta.txt
The metadata file includes the title, description, hashtags, hook, intro/outro text, source story, selected footage, music, photo credits, and final video path.
CobbleGen/
main.py CLI entrypoint and progress events
config.py .env loading and typed settings
modules/ story, TTS, subtitles, images, video, state
webapp/ FastAPI studio
run.py local web server launcher
server.py API routes
jobs.py background job manager
static/ UI, logo, favicon assets
docs/screenshots/ README preview screenshots
footage/ local user videos, ignored
music/ local user music, ignored
stories/ local story text, ignored
output/ generated reels, ignored
temp/ render workspace, ignored
If Ollama fails, confirm the configured host is reachable and has the model:
curl http://localhost:11434/api/tags
curl http://master:11434/api/tagsIf NVIDIA voice listing or generation fails, check NVIDIA_API_KEY, network access to grpc.nvcf.nvidia.com:443, and:
python main.py --list-voicesIf no video is produced, confirm FFmpeg/FFprobe are installed, footage/ contains at least one supported video file, the story is not empty, and your TTS backend can generate audio.
If GPU rendering fails, set:
VIDEO_RENDER_PREFER_GPU=false
VIDEO_CODEC=libx264