autoshorts：AI 辅助选段的长视频转短视频工具

Turn long-form videos into short clips with AI-assisted moment selection, GPU rendering, styled captions, and optional voiceovers.

Showcase

Four clips generated from gameplay footage:

Clip 1	Clip 2	Clip 3	Clip 4

What it does

Content-aware clipping: 15 presets with their own moment categories, analysis prompts, caption styles, and voice presets. Gaming/esports use action heuristics plus AI ranking; other types use AI-first selection.
Flexible analysis: Gemini, including full-video deep analysis; OpenAI vision; or local heuristic selection without API analysis.
Captions and narration: Whisper transcription, AI-written captions, PyCaps styling with FFmpeg fallback, and cross-clip story narration.
Optional voiceovers: Local Qwen3-TTS with multilingual voice design, adjustable source/voice volumes, and subtitle timing matched to speech.
GPU rendering: CUDA decoding and analysis, cropping, blurred backgrounds, configurable aspect ratios, and NVENC encoding.
Dashboard: Choose content and caption styles, manage the video queue, monitor jobs, and browse generated clips.

Supported content: gaming, esports, podcasts, interviews, news/commentary, entertainment, TV shows, comedy, music, vlogs, cooking/DIY, fitness, education, documentaries, and sports.

Examples: English roast / Japanese narration / Chinese news style.

Get started

You need an NVIDIA GPU and driver, a CUDA toolkit with nvcc, FFmpeg with NVENC support, and Python 3.10+. For voiceovers, 6 GB+ VRAM and 16 GB system RAM are recommended.

git clone https://github.com/divyaprakash0426/autoshorts.git
cd autoshorts
make install
cp .env.example .env

The Makefile creates a single .venv with the Python dependencies and CUDA-enabled Decord. Decord's build requires FFmpeg 4.x libraries; see environment.yml. PyTorch 2.10 pairs with TorchCodec 0.10.0. An interactive installer and Docker setup are also provided; Docker requires the NVIDIA Container Toolkit.

Edit .env before running: choose a content type and AI provider, add the corresponding API key, and set DEBUG_SKIP_ANALYSIS=false and DEBUG_SKIP_RENDER=false for normal processing.

Dashboard or CLI

# Dashboard: configure settings, add videos, and start processing
.venv/bin/python -m streamlit run src/dashboard/About.py

# CLI: read videos from gameplay/ and write clips to generated/
.venv/bin/python run.py

Use the dashboard's file picker or place source videos in gameplay/. Output clips are saved in generated/.

Configuration

Settings live in .env and can also be changed through the dashboard.

Setting	Purpose
`AI_PROVIDER`	`gemini`, `openai`, or `local`
`VIDEO_TYPE`	Content preset, such as `gaming`, `podcasts`, or `educational`
`ENABLE_SUBTITLES` / `SUBTITLE_MODE`	Toggle captions; choose `speech`, `ai_captions`, or `none`
`CAPTION_STYLE`	`auto` matches the detected moment; explicit styles are content-specific
`ENABLE_TTS` / `TTS_LANGUAGE`	Toggle voiceovers and select their language
`SCENE_LIMIT`	Maximum clips per source video
`MIN_SHORT_LENGTH` / `MAX_SHORT_LENGTH`	Clip duration bounds; `0` selects automatic lengths
`TARGET_RATIO_W` / `TARGET_RATIO_H`	Output aspect ratio, typically `9` / `16`

See .env.example for all settings and voice design for advanced TTS options.

Privacy and cost: Cloud providers receive video content. Gemini deep analysis uploads the full source video; API usage may incur charges. local uses heuristic selection instead, but GPU rendering still requires NVIDIA hardware.

Development

# Lightweight suite; real GPU/model checks are skipped
.venv/bin/python -m pytest tests/ -q

# Include actual CUDA decoding, rendering, and local model inference
.venv/bin/python -m pytest tests/ -q --run-gpu --run-models

See Contributing for test prerequisites and bounded footage checks, pipeline flow for architecture, and Roadmap for planned features.

License and credits

Apache 2.0; see NOTICE. Dependencies and model weights retain their own licenses.

Built on work from shorts-maker-gpu and Auto-YouTube-Shorts-Maker.

Support the project -- donations go to charity.