Skip to content
Other

LeanScene - ultra-lightweight AI model for finding video cuts

2 min read

LeanScene by developer lucaspy is a compact model with only 0.39M parameters that automatically detects scene boundaries in videos: hard cuts, dissolves, and fades. It outputs exact frame ranges for each shot along with a list of transitions containing their types and probabilities. With a weight size of just 1.6 MB, the AI model outperforms the popular TransNet V2 model in quality across all three test datasets while requiring 115 times fewer computational operations (MACs) per frame. The tool is useful for video editors, video analytics system developers, and anyone preparing datasets from long videos.

Key features

  • Weighs only 1.6 MB (0.39M parameters) and requires 7.2M MACs (multiply-accumulate operations) per frame.
  • Outperforms TransNet V2 on the F1 metric across the BBC Planet Earth, ClipShots, and RAI benchmarks.
  • Runs locally without loading weights at runtime: models are already embedded in ready-to-use Python and Node.js packages.
  • Available in ONNX format (two graphs for feature extraction and temporal head), enabling it to run directly in the browser without sending videos to a server.

What you can do

  • Automatically split long videos into separate scenes when preparing datasets for training video models.
  • Integrate cut detection into a Python script or Node.js service for indexing and analyzing video archives.
  • Check your video's markup right in a browser tab via the author's web demo without uploading files to the cloud.

Example

import leanscene

result = leanscene.detect("video.mp4")
result.shots         # [(0, 343), (344, 2403), ...]
result.transitions   # [Transition(start=344, end=344, kind='cut'), ...]
result.probs         # (frames, 3) float32: same, cut, fade
Read moreDetails
Benchmarks (F1, average over 5 seeds): BBC Planet Earth - 97.4 (vs. 96.1 for TransNet V2), ClipShots - 79.1 (vs. 77.7), RAI - 93.7 (vs. 92.6).
Weights in the repository (one of the seeds): 97.45 / 78.82 / 93.03.
Limitations: The Python package decodes video via OpenCV, so timings may shift on variable frame rate (VFR) files, although frame indices remain accurate.
License: MIT.
Discuss Telegram 0

Will it run?

Runs in BF16

Fits without quantisation.

  • BF16weights 0.0 GB · ready files0.5 GB
  • FP8 / INT8recommendedweights 0.0 GB0.5 GB
  • 4-bitweights 0.0 GB0.5 GB

Tip: go 8-bit.FP8 / INT8 is within about a percent of full precision in quality, takes half the memory and runs faster on RTX 40/50 and H100. Full precision only pays off for fine-tuning.

What to get for this model

Massed ComputeOpen Massed Compute

Your own GPU: FP8 / INT8 — from RTX 2060 · BF16 — from RTX 2060

Massed Compute hourly prices, November 2026. The discounted price applies with the coupon at checkout.

Estimate: weights plus working memory. Long videos, high resolutions and long contexts need more.

Run it on Massed Compute

The script downloads the model to the rented machine. How to run it is in the model card.

  1. 1. Rent a machine with a fitting GPU

    Open Massed Compute
  2. 2. Paste the script into the machine's terminal

    Connect over SSH (the Massed Compute panel shows the command) or open a terminal on the machine's desktop.

    HF_TOKEN / CIVITAI_TOKENThe script starts with token lines: paste yours between the quotes. The Hugging Face token is for gated models, the Civitai one for Civitai models.

    # lucaspy/LeanScene
    mkdir -p ~/models
    curl -LsSf https://astral.sh/uv/install.sh | sh
    source "$HOME/.local/bin/env"
    uv tool install "huggingface_hub[cli]"
    hf download lucaspy/LeanScene --local-dir ~/models/LeanScene
    # The model is in ~/models. How to run it is in the model card (link above).
    Add more LoRAs to LeanScene
  3. 3. Use it

    The model is in ~/models. How to run it is in the model card (link above).

Stop the machine in the panel when you are done: it bills by the minute while it runs.

How to speed it up

Picked for this model type. Choose your hardware above to see what fits you exactly.

  • FlashAttention 44.0.0b40from November 18, 2026new release

    fastest exact attention on B200

    Kernels for Blackwell (beta). Exact attention without quantisation — fine for everything, LLMs included.

    How to enable: pip install flash-attn-4; vLLM and SGLang pick it up themselves.

    Earlier generations (2)
    • FlashAttention 3 — The Hopper (H100/H200) version with FP8.
    • FlashAttention 2 — The standard for RTX 30/40/50 and A100: exact attention without the memory cost of long contexts.
  • torch.compile2.15.0from October 28, 2026new release

    10–30% for free

    Compiles the model for your GPU. Slower first run, faster afterwards.

    How to enable: model = torch.compile(model); in ComfyUI the TorchCompileModel node.

Model family

New posts about versions of LeanScene (quants, finetunes, LoRAs) will show up under the bell.

No known relatives of this model yet. All derivatives on Hugging Face

About the model

lucaspy/LeanScene
Author
lucaspy
Category
Other
Released
Posted

On NeyroLearn

Hugging Face

00

Ask AI about this model

Answered by Gemini 4 · the chat is kept only in this tab

Comments