README.md

---
license: mit
tags:
- Serial Scaling Hypothesis
- Diffusion
pipeline_tag: video-to-video
---

<h1 align="center">The Seriality Gap in Video Diffusion Models</h1>

<p align="center">
  <strong><a href="https://jdiazchao.com">Jorge Diaz Chao</a><sup>*</sup></strong>&nbsp;&nbsp;&nbsp;&nbsp;
  <strong><a href="https://konpat.notion.site">Konpat Preechakul</a><sup>*</sup></strong>&nbsp;&nbsp;&nbsp;&nbsp;
  <strong><a href="https://yuxi.ml">Yuxi Liu</a></strong>&nbsp;&nbsp;&nbsp;&nbsp;
  <strong><a href="https://yutongbai.com">Yutong Bai</a></strong>
</p>

<p align="center">
  UC Berkeley<br>
  <code>{jdiazchao,konpat,yuxi_liu,yutongbai}@berkeley.edu</code>
</p>

<p align="center">
  <a href="#">Paper</a> |
  <a href="#">Project Site</a> |
  <a href="https://github.com/jdiazchao/seriality-gap">Code</a>
</p>

<figure align="center">
  <div>&nbsp;</div>
  <img src="figures/teaser.jpg" alt="The seriality gap in video diffusion models">
  <div>&nbsp;</div>
  <figcaption>
    <p align="justify">
    Figure 1. <strong>Dependent-event prediction exposes the seriality gap.</strong>
    (a) Hard-sphere dynamics separate non-serial from serial video prediction.
    (i) In the single-ball control, any future state can be computed directly from the initial state, without resolving intermediate states.
    (ii) With multiple balls, each ball-ball collision changes the state governing later collisions, creating dependent-event chains that must be resolved in temporal order.
    (b) Given initial frames, can a diffusion model remain accurate as longer prediction horizons demand more serial computation?
    </p>
  </figcaption>
</figure>

## Download a Model

Clone the [seriality-gap GitHub repository](https://github.com/jdiazchao/seriality-gap), then follow its [setup instructions](https://github.com/jdiazchao/seriality-gap#setup).

```bash
git clone https://github.com/jdiazchao/seriality-gap
cd seriality-gap
```

Once setup is complete, download a checkpoint by passing its training configuration to `download.sh`:

```bash
bash download.sh configs/49f-5n/Train-20k-49f-5n-0c-B-768d-30l-64b-2e4lr.yaml
```

You may equivalently pass the model name directly:

```bash
bash download.sh 20k-49f-5n-0c-B-768d-30l-64b-2e4lr
```

Downloaded weights are saved to:

```text
ckpt/huggingface/<model-name>/model-avg.safetensors
```

## BibTeX

```bibtex
@misc{chao2026serialitygapvideodiffusion,
      title={The Seriality Gap in Video Diffusion Models}, 
      author={Jorge Diaz Chao and Konpat Preechakul and Yuxi Liu and Yutong Bai},
      year={2026},
      eprint={2607.13031},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2607.13031}, 
}
```