---
license: apache-2.0
tags:
  - video-understanding
  - fps-prediction
  - visual-chronometer
  - pulse-of-motion
pipeline_tag: video-classification
---

# Visual Chronometer — Physical FPS Prediction

Safetensors conversion of the [Visual Chronometer](https://github.com/taco-group/Pulse-of-Motion) model for predicting the **Physical Frame Rate (PhyFPS)** from video motion patterns.

## Model Details

- **Architecture:** 2+1D VAE encoder + attention-pooled probe + MLP regression head
- **Training range:** 10–60 FPS
- **Output:** `log(FPS)` — take `exp()` to get predicted PhyFPS
- **Input:** 30-frame clips at 216×216 resolution, normalized to `[-1, 1]`
- **Parameters:** ~170M (683 MB safetensors)

## Format

This repo provides the model in **safetensors** format, converted from the original PyTorch Lightning checkpoint.

| File | Size | Format |
|---|---|---|
| `vc_common_10_60fps.safetensors` | 683 MB | safetensors |

## Original

- **Paper:** [Pulse of Motion](https://github.com/taco-group/Pulse-of-Motion)
- **Original weights:** [xiangbog/Visual_Chronometer](https://huggingface.co/xiangbog/Visual_Chronometer)
- **License:** Apache 2.0

## Usage

Used by [ComfyUI-FFMPEGA](https://github.com/AEmotionStudio/ComfyUI-FFMPEGA) as the `phyfps` no-LLM mode for automated video temporal quality assessment.