Salt++: Context-Aligned Post-Training for Few-Step Streaming Multimodal Generation
Salt++ post-trains few-step streaming audio-video models, beating OmniForcing on quality and LTX-2 on most metrics.
Salt++ is a two-stage post-training method for few-step streaming audio-video generation, combining Causal Self-Flow with context-aligned autoregressive distribution matching distillation. Causal Self-Flow trains a noise-mixed-history student against a clean-history teacher, while AR DMD shares the causal mask and prefix across sampling and scoring. At 480p under a 4-step causal setting it improves visual and motion quality by 57% and 45% over OmniForcing on JavisBench, and a later stage reaches 4-step 1664x960 generation, outperforming bidirectional LTX-2 on six of seven reported metrics.
- Salt++ combines Causal Self-Flow with context-aligned autoregressive DMD.
- Shared causal masks align generator sampling with fake and real score models.
- At 480p and 4 steps it beats OmniForcing by 57% visual and 45% motion quality on JavisBench.
- A scale stage reaches 4-step 1664x960 and beats LTX-2 on six of seven metrics.
- Products
- JavisBench
- AI models
- Salt++OmniForcingLTX-2
From the original article
1 min read at huggingface.co
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.
By Xingtong Ge, Yutong Wang, Lunjie Zhu · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.