🧂 Salt++: Context-Aligned Post-Training for Few-Step Streaming Multimodal Generation

Xingtong Ge^1,2, Yutong Wang³, Lunjie Zhu¹, Haitao Lin⁴, Fangyu Lin¹, Yushi Huang¹, Xin Zhang², Yi Zhang², Yu Liu², Jun Zhang¹

¹The Hong Kong University of Science and Technology, ²Vivix Group Limited, ³The University of Sydney, ⁴Westlake University

Preprint, 2026

A successor to Salt, extending few-step distribution matching from video distillation to causal joint audio–video generation.

📝 Abstract

Few-step streaming audio–video generation requires both causal modeling and step distillation, yet standard training recipes face two context-related challenges. Teacher forcing pairs clean history with a noisy target, but supervises predictive contextual representations only indirectly through velocity prediction. Meanwhile, directly reusing bidirectional score models in causal Distribution Matching Distillation (DMD) creates a mismatch between generation and scoring contexts. We address these challenges with Salt++, a two-stage post-training framework comprising Causal Self-Flow (CSF) and context-aligned autoregressive DMD. CSF exploits contextual information asymmetry by varying the history while keeping the noisy target fixed: a noise-mixed-history student aligns its intermediate representations with those of a clean-history exponential-moving-average teacher. This self-supervised signal encourages the student to extract semantic information and improves cross-modal alignment. Context-aligned AR DMD shares the causal mask and prefix across generator sampling, fake-score training, and real-score evaluation to match generated and reference distributions under a block-conditional KL objective. With calibrated teacher guidance, it performs clean-prefix few-step distillation and then adapts to generated histories without switching objectives or requiring separate consistency distillation. At 480p, Salt++ improves visual and motion quality by 57% and 45% over OmniForcing on JavisBench under the same 4-step causal setting. A separate scale-wise post-training stage extends Salt++ to 4-step 1664×960 generation, outperforming bidirectional LTX-2 on six of seven reported metrics.

Four-step streaming audio–video generation at 480p. Salt++ retains scene and facial detail that the causal baseline loses at the same step budget.

🔊 Watch the generated clips with audio on the project page. Audio and video are generated jointly, so audio–visual synchrony can be judged directly.

✅ TODO List

• Paper on arXiv
• Project page with audio–video samples
• Inference scripts
• Open-source model weights
• Training scripts

Stay tuned 🚀

🔍 Method

Prior recipes switch objectives between stages and score the causal generator with bidirectional models. Salt++ keeps a single AR DMD objective throughout post-training and only shifts its conditioning from clean context to generated rollout:

1. Causal Self-Flow (CSF) varies the history while holding the noisy target fixed, turning contextual information asymmetry into a representation-learning signal for the autoregressive teacher.
2. Context-aligned AR DMD shares one causal mask and prefix across generator sampling, fake-score training, and real-score evaluation, so distribution matching alone suffices — no separate consistency-distillation stage.
3. Scale-wise post-training extends the recipe to 1664×960 with two low-resolution and two high-resolution generator calls per block, still four generator calls in total.

Two roles of causal context in post-training. (a) CSF exploits information asymmetry in causal histories for self-supervised representation learning. (b) Mismatched (top) and aligned (bottom) contexts across the generator, real score, and fake score.

📈 Results

Table 1: JavisBench-mini, 480p audio–video generation

Bold = best, underline = second best among the 4-step rows.

Model	Causal	Steps	VQ ↑	MQ ↑	AQ ↑	CLIP ↑	IB-AV ↑	Javis ↑	DeSync ↓
LTX-2 Base	✗	40	1.884	0.566	4.986	0.311	0.239	0.200	0.608
AR Teacher (CSF)	✓	40	2.304	0.857	4.565	0.316	0.202	0.162	0.746
OmniForcing	✓	4	1.807	0.699	4.718	0.303	0.163	0.124	0.745
Salt++ (TF-dCM route)	✓	4	2.013	0.826	4.976	0.313	0.229	0.185	0.710
Salt++	✓	4	2.838	1.010	4.991	0.316	0.184	0.146	0.759

Table 2: VBench video fidelity and temporal consistency, 480p

Model	Causal	Steps	Aesthetic ↑	Imaging ↑	Subject Consistency ↑	Background Consistency ↑
LTX-2 Base	✗	40	53.89	67.06	95.94	95.63
OmniForcing	✓	4	56.74	68.41	96.62	95.32
Salt++ (TF-dCM route)	✓	4	53.65	62.87	96.58	96.05
Salt++	✓	4	55.43	69.98	97.01	96.24

Table 3: JavisBench-mini, 1664×960 audio–video generation

Bold = best across all methods.

Model	Causal	Steps	VQ ↑	MQ ↑	AQ ↑	CLIP ↑	IB-AV ↑	Javis ↑	DeSync ↓
LTX-2 Base	✗	40 + 3	2.231	0.607	4.867	0.311	0.169	0.145	0.658
OmniForcing	✓	4	2.298	0.832	4.721	0.293	0.165	0.128	0.755
Salt++	✓	4	2.730	0.957	5.113	0.318	0.193	0.157	0.768

Scale-wise post-training at 1664×960, using two low-resolution and two high-resolution generator calls per block.

📚 Citation

If you find this work useful, please cite:

@misc{ge2026saltpp,
  title={Salt++: Context-Aligned Post-Training for Few-Step Streaming Multimodal Generation},
  author={Ge, Xingtong and Wang, Yutong and Zhu, Lunjie and Lin, Haitao and Lin, Fangyu and Huang, Yushi and Zhang, Xin and Zhang, Yi and Liu, Yu and Zhang, Jun},
  year={2026},
  eprint={2609.36995},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2609.36995}
}

🙏 Acknowledgements

This work builds on LTX-2 as the bidirectional audio–video foundation model, follows the causal streaming setup of OmniForcing, and evaluates with JavisBench and VBench. Please also cite the original projects when using their components.