README.md

---
license: other
pipeline_tag: text-to-video
tags:
- video-generation
- text-to-video
- diffusion
- distribution-matching
- distillation
- wan
- arxiv:2604.03118
---

# 🧂 Salt: Self-Consistent Distribution Matching with Cache-Aware Training for Fast Video Generation

[![arXiv](https://img.shields.io/badge/Arxiv-2604.03118-b31b1b)](https://arxiv.org/abs/2604.03118)
[![Project Page](https://img.shields.io/badge/Project-Page-green)](https://xingtongge.github.io/Salt/)
[![GitHub Repo stars](https://img.shields.io/github/stars/XingtongGe/Salt.svg?style=social&label=Star&maxAge=60)](https://github.com/XingtongGe/Salt)
[![Hugging Face](https://img.shields.io/badge/🤗%20Hugging%20Face-Model-yellow)](https://huggingface.co/domiso/Salt)

[Xingtong Ge](https://xingtongge.github.io/)<sup>1,2</sup>, [Yi Zhang](https://zhangyi-3.github.io/)<sup>2</sup>, Yushi Huang<sup>1</sup>, Dailan He<sup>2</sup>, Xiahong Wang<sup>2</sup>, Bingqi Ma<sup>2</sup>, Guanglu Song<sup>2</sup>, Yu Liu<sup>2</sup>, Jun Zhang<sup>1</sup>

<sup>1</sup>The Hong Kong University of Science and Technology, <sup>2</sup>Vivix Group Limited

European Conference on Computer Vision (**ECCV**), 2026

## Abstract

Distilling video generation models to extremely low inference budgets (e.g., 2-4 NFEs) is crucial for real-time deployment, yet remains challenging. Trajectory-style consistency distillation often becomes conservative under complex video dynamics, yielding over-smoothed appearance and weak motion. Distribution matching distillation (DMD) can recover sharp, mode-seeking samples, but its local training signals do not explicitly regularize how denoising updates compose across timesteps, making composed rollouts prone to drift. To overcome this challenge, we propose Self-Consistent Distribution Matching Distillation (SC-DMD), which explicitly regularizes the endpoint-consistent composition of consecutive denoising updates. For real-time autoregressive video generation, we further treat the KV cache as a quality-parameterized condition and propose cache-distribution-aware training. This training scheme applies SC-DMD over multi-step rollouts and introduces a cache-conditioned feature alignment objective that steers low-quality outputs toward high-quality references. Across extensive experiments on both non-autoregressive backbones (e.g., Wan 2.1) and autoregressive real-time paradigms (e.g., Self Forcing, Causal Forcing, and LongLive), Salt consistently improves low-NFE video generation quality while remaining compatible with diverse KV-cache memory mechanisms.

## Released Models

### Salt + Causal Forcing

- `checkpoints/salt_cf.pt`
- Supports 2-step and 4-step autoregressive generation.
- Use the EMA generator for inference.

### Salt + LongLive

- `checkpoints/salt_ll.pt`
- Supports 4-step autoregressive generation with LongLive KV-cache memory.
- Use the regular generator for inference.

The training prompt collection used by the released recipes is also included
at `prompts/vidprom_filtered_extended.txt`.

## Selected Results

### Text-to-video generation on VBench

| Model | NFE | Total | Quality | Semantic |
| --- | ---: | ---: | ---: | ---: |
| Self Forcing | 4 | 84.20 | 84.74 | **82.05** |
| Salt + Self Forcing | 4 | **84.47** | **85.27** | 81.28 |
| LongLive | 4 | 84.40 | 85.12 | 81.53 |
| Salt + LongLive | 4 | **84.93** | **85.41** | **83.00** |
| Causal Forcing | 4 | 84.62 | 85.41 | 81.47 |
| Salt + Causal Forcing | 4 | **85.08** | **85.96** | **81.59** |
| Salt + Causal Forcing | 2 | **84.80** | **85.63** | **81.49** |

![Qualitative comparisons with Causal Forcing](https://raw.githubusercontent.com/XingtongGe/Salt/main/assets/vis_supp.png)

## Usage

### 1. Prepare the code and artifacts

```bash
git clone https://github.com/XingtongGe/Salt.git
cd Salt

# Downloads checkpoints/ and prompts/ into the paths expected by the configs.
hf download domiso/Salt --local-dir .
```

Follow the installation and Wan2.1 preparation instructions in the
[GitHub repository](https://github.com/XingtongGe/Salt).

### 2. Salt + Causal Forcing

```bash
python inference.py \
  --config_path configs/inference/salt_causal_forcing.yaml \
  --checkpoint_path checkpoints/salt_cf.pt \
  --data_path prompts/example_prompts.txt \
  --output_folder outputs/salt_cf \
  --use_ema
```

### 3. Salt + LongLive

```bash
python inference.py \
  --config_path configs/inference/salt_longlive.yaml \
  --checkpoint_path checkpoints/salt_ll.pt \
  --data_path prompts/example_prompts.txt \
  --output_folder outputs/salt_ll
```

## Training

The released prompt collection contains one prompt per line and matches the
default public recipes. Salt training does not require a video dataset.

The code repository provides recipes for:

- Self Forcing, Causal Forcing, and LongLive baselines;
- mixed-step SC-DMD; and
- mixed-step SC-DMD with TRD alignment.

All canonical mixed-step recipes sample 8-, 4-, and 2-step trajectories with
probabilities **0.4 / 0.4 / 0.2**. See the
[configuration guide](https://github.com/XingtongGe/Salt/tree/main/configs) for
the complete recipe matrix.

## License

Please review the code repository's license and third-party notices before
redistribution or commercial use. In particular, the LongLive backbone carries
a file-level CC-BY-NC-SA-4.0 notice. Model weights may also be subject to their
upstream backbone licenses.

## Citation

If you find this work useful, please cite:

```bibtex
@article{ge2026salt,
  title={Salt: Self-consistent distribution matching with cache-aware training for fast video generation},
  author={Ge, Xingtong and Zhang, Yi and Huang, Yushi and He, Dailan and Wang, Xiahong and Ma, Bingqi and Song, Guanglu and Liu, Yu and Zhang, Jun},
  journal={arXiv preprint arXiv:2604.03118},
  year={2026}
}
```