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

Salt++. Context-Aligned Post-Training for Few-Step Streaming Multimodal Generation. By Xingtong Ge and nine others, from the Hong Kong University of Science and Technology, Vivix Group Limited, and two other institutions. Published on the twenty-ninth of September. Its code repository is a placeholder: a readme, and nothing to run yet.

Streaming audio and video generation in a few steps needs two things at once: causal modeling, and step distillation. Standard training recipes run into two challenges, and both are about context. Teacher forcing pairs a clean history with a noisy target, so the predictive contextual representations are supervised only indirectly, through velocity prediction. And in causal Distribution Matching Distillation, or DMD, directly reusing bidirectional score models creates a mismatch between the context content is generated in and the context it is scored in. One challenge concerns what the model learns about its history. The other concerns how its outputs are judged.

Salt++ is a two-stage post-training framework that answers each challenge with a stage of its own. The first is Causal Self-Flow, or CSF. The second is context-aligned autoregressive DMD.

CSF exploits an asymmetry in contextual information. It varies the history while keeping the noisy target fixed. Since the target stays put and only the history gets noisier, the thing being varied is how much context there is to rely on. A student that sees a noise-mixed history aligns its intermediate representations with those of a teacher that sees a clean history, the teacher being an exponential moving average. That self-supervised signal encourages the student to extract semantic information, and it improves cross-modal alignment.

The second stage shares the causal mask and the prefix across all three places they appear: generator sampling, fake-score training, and real-score evaluation. Generated and reference distributions are then matched under a block-conditional KL objective. With calibrated teacher guidance, this stage performs few-step distillation from a clean prefix and then adapts to generated histories, without switching objectives and without a separate consistency distillation.

At four hundred and eighty p, under the same four-step causal setting, Salt++ improves visual quality by fifty-seven percent and motion quality by forty-five percent over OmniForcing on JavisBench. A separate scale-wise post-training stage extends it to four-step generation at one thousand six hundred and sixty-four by nine hundred and sixty. At that resolution it outperforms the bidirectional LTX-2 on six of seven reported metrics. The comparison there is with a bidirectional model, the kind of model whose scores the causal setting could not simply reuse.

Both stages rest on one observation: in streaming generation, context is what goes wrong. The first stage supervises the use of context directly rather than through velocity alone. The second makes sure the model is scored under the same context it generates in. Each fix targets one of the two mismatches, and the pair is what makes few-step streaming work.