smol-tldw

Most tl;dw apps usually use caption context for generating summaries. smol-tldw uses pure video context to generate a description using SmolVLM2. It is experimental but quality of summaries are great for such smol models! (256M/500M) Obvious caveat is this is limited to visual context.

Setup

Install dependencies using uv

uv sync

Usage

python -m smol_tldw "https://www.youtube.com/watch?v=..."

Native frame decoding using torchcodec. Pipeline first downloads the video using yt-dlp. Loads HuggingFaceTB/SmolVLM2-256M-Video-Instruct model by default. You can change it by passing --model.

Run with --help to see all options.

Examples

python -m smol_tldw.cli \
    --video-input frames \
    --max-frames 64 \
    --repl \
    "https://www.youtube.com/watch?v=..."