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

Add some Salt 🧂 to your video generation distillation.

Xingtong Ge^1,2, Yi Zhang², Yushi Huang¹, Dailan He², Xiahong Wang², Bingqi Ma², Guanglu Song², Yu Liu², Jun Zhang¹

¹ Hong Kong University of Science and Technology
² Vivix Group Limited

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.

Code Release

Salt improves few-step video generation with two complementary components:

1. SC-DMD regularizes the consistency of direct and composed denoising updates.
2. Cache-conditioned alignment uses high-quality cache trajectories as references. This release provides the spatial-only TRD alignment used by the selected autoregressive recipes.

The implementation supports Self Forcing, Causal Forcing, and LongLive initializations. The canonical mixed-step distribution samples 8-, 4-, and 2-step trajectories with probabilities 0.4 / 0.4 / 0.2.

Selected Results

Text-to-video generation on VBench

Diffusion models

Model	NFE	Total	Quality	Semantic
rCM	4	82.73	83.65	79.04
DMD	4	82.78	84.39	76.36
Salt (SC-DMD)	4	83.19	84.42	78.30

Autoregressive models

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

Image-to-video generation on VBench-I2V

Method	NFE	I2V Score	Quality	Background Consistency	Motion Smoothness	Dynamic Degree	Imaging Quality	Temporal Flicker
PCM	8	93.63	78.52	97.34	98.24	30.98	70.42	97.67
DMD	4	93.09	78.89	92.79	97.99	58.46	70.35	95.21
LightX2V	4	93.50	80.92	95.87	97.89	60.33	71.67	96.30
Salt (SC-DMD)	4	93.90	80.86	95.97	98.37	52.85	72.16	97.41
Salt-alpha	4	93.88	81.71	95.46	98.30	68.13	72.08	96.48

Long-horizon autoregressive generation on VBench-Long

Backbone	Total	Quality	Semantic
Causal Forcing	78.11	82.57	60.25
Salt + Causal Forcing	78.28	82.15	62.77
LongLive	79.03	82.82	63.88
Salt + LongLive	79.27	82.90	64.74

Qualitative Results

The figure below shows additional qualitative comparisons with the Causal Forcing baseline. Salt better preserves subject identity, object geometry, scene composition, and motion smoothness across challenging examples.

Figure: Qualitative comparisons with the Causal Forcing baseline. Salt better preserves subject identity, object geometry, scene composition, and temporal coherence across challenging examples, including the umbrella, trombone, reading-girl, and grape cases.

Repository layout

configs/
  baseline/             original Self/Causal/LongLive training recipes
  mixed_sc_dmd/         mixed-step SC-DMD recipes
  mixed_sc_dmd_trd/     mixed-step SC-DMD + TRD recipes
  inference/            released-checkpoint inference recipes
model/                  DMD, SC-DMD, TRD, and streaming model losses
pipeline/               Self Forcing, Causal Forcing, and LongLive rollouts
trainer/                distributed distillation training loop
utils/                  data, FSDP, scheduler, and Wan wrappers
wan/                    modified Wan2.1 and causal/LongLive backbones

See configs/README.md for the recipe matrix and the paper-to-code map.

Requirements

The code targets Linux, Python 3.10, PyTorch 2.4+, and NVIDIA GPUs. Training uses FSDP and was developed for multi-GPU H100/A100 environments. Inference can run on fewer GPUs, subject to model size and available memory.

conda create -n salt python=3.10 -y
conda activate salt

pip install -r requirements.txt
# Recommended on supported NVIDIA GPUs:
pip install flash-attn --no-build-isolation
pip install -e .

Optional Wan prompt-extension and USP dependencies are listed in requirements-optional.txt.

Model and prompt preparation

Download the Wan2.1 backbones into the paths used by the public configs:

huggingface-cli download Wan-AI/Wan2.1-T2V-1.3B \
  --local-dir wan_models/Wan2.1-T2V-1.3B
huggingface-cli download Wan-AI/Wan2.1-T2V-14B \
  --local-dir wan_models/Wan2.1-T2V-14B

Download the released Salt checkpoints and the training prompts together. Run this command from the repository root so all relative paths match the configs:

hf download domiso/Salt --local-dir .

This creates:

checkpoints/
  salt_cf.pt               Salt + Causal Forcing inference checkpoint
  salt_ll.pt               Salt + LongLive inference checkpoint
prompts/
  vidprom_filtered_extended.txt

Checkpoint files are intentionally not stored in Git. Training is data-free with respect to videos. The included text file contains one training prompt per line; replace data_path if using another prompt collection.

Training from the three initialization families additionally requires checkpoints/ode_init.pt, checkpoints/causal_ode.pt, or the optional original checkpoints/self_forcing_dmd.pt, depending on the selected recipe. Obtain these initialization weights from their Self Forcing/Causal Forcing sources or update generator_ckpt to your local equivalents.

Training

The examples below use a single 8-GPU node. Adjust --nproc_per_node, FSDP settings, and total_batch_size for your environment.

Mixed-step SC-DMD

# Self Forcing initialization
torchrun --nproc_per_node=8 train.py \
  --config_path configs/mixed_sc_dmd/self_forcing.yaml \
  --logdir logs/sc_dmd_self_forcing --disable-wandb

# Causal Forcing initialization
torchrun --nproc_per_node=8 train.py \
  --config_path configs/mixed_sc_dmd/causal_forcing.yaml \
  --logdir logs/sc_dmd_causal_forcing --disable-wandb

# LongLive initialization
torchrun --nproc_per_node=8 train.py \
  --config_path configs/mixed_sc_dmd/longlive.yaml \
  --logdir logs/sc_dmd_longlive --disable-wandb

Mixed-step SC-DMD + TRD alignment

CONFIG=configs/mixed_sc_dmd_trd/self_forcing.yaml \
  bash scripts/train.sh

CONFIG=configs/mixed_sc_dmd_trd/causal_forcing.yaml \
  bash scripts/train.sh

CONFIG=configs/mixed_sc_dmd_trd/longlive.yaml \
  bash scripts/train.sh

All six mixed-step recipes use:

scfm_step_lists_8: [1000, 875, 750, 625, 500, 375, 250, 125]
scfm_step_lists_4: [1000, 750, 500, 250]
scfm_step_lists_2: [1000, 500]
scfm_step_probs: [0.4, 0.4, 0.2]  # 8-step / 4-step / 2-step

Inference

Salt + Causal Forcing uses the EMA generator, while the selected Salt + LongLive checkpoint uses the regular generator:

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

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

Acknowledgements

This codebase builds on Self Forcing, CausVid, and Wan2.1. Please also cite the original projects when using their components.

License

The repository contains components under more than one license. Most code is provided under Apache-2.0, while wan/modules/causal_model_longlive.py is marked CC-BY-NC-SA-4.0. See THIRD_PARTY_LICENSES.md before redistribution or commercial use. Model weights are governed by their respective upstream licenses.

Citation

@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}
}