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

LTX-2 · bidirectional · 40 steps

OmniForcing · causal · 4 steps

Salt++ (TF-dCM route) · 4 steps

Salt++ (AR–AR DMD route) · 4 steps

LTX-2 · bidirectional · 40 steps

OmniForcing · causal · 4 steps

Salt++ (TF-dCM route) · 4 steps

Salt++ (AR–AR DMD route) · 4 steps

01 / 02

Fast motion at the same four-step budget

Same prompt and seed. LTX-2 is bidirectional at 40 steps; OmniForcing and both Salt++ routes are causal at four steps. Tap a panel's ♫ chip to switch which soundtrack you hear.

01 / Overview

Abstract

Few-step streaming audio–video generation requires both causal modeling and step distillation. Teacher forcing pairs clean history with a noisy target, but shapes 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 predicts intermediate representations from a clean-history exponential-moving-average teacher. Together with flow matching, this self-supervised signal encourages the student to extract semantic information useful for next-block prediction 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 scale-wise post-training stage further extends Salt++ to 4-step 1664×960 generation, outperforming bidirectional LTX-2 on six of seven reported metrics.

Joint audio and video

A single causal generator produces temporally aligned video and audio blocks, so every clip below carries its own generated soundtrack.

Causal Self-Flow

Varying only the history while holding the noisy target fixed turns information asymmetry into a representation-learning signal for the autoregressive teacher.

Context-aligned AR DMD

Generator sampling and both score models share one causal mask and prefix, so distribution matching alone suffices — no separate consistency-distillation stage.

02 / Method

Method

Salt++ keeps one autoregressive DMD objective throughout post-training and only shifts its conditioning from clean context to generated rollout.

Comparison with recent causal post-training recipes

Two roles of causal context in post-training

Score-context configurations in causal DMD

Cross-modal alignment during AR teacher training

03 / Videos

Generated results

Audio and video are generated jointly. Every clip carries its own generated soundtrack, so audio–visual synchrony can be judged directly. Press Unmute, or Play together to start a whole row in sync.

480p qualitative comparison between LTX-2 Base, OmniForcing and both Salt++ routes

480p streaming generation

Four-step causal generation against the bidirectional foundation model and the causal baseline, under matched prompts and seeds.

Wellness host · paper Fig. 8

[Shot 1][0.0s-5.0s]: A calm wellness-channel video in a sunlit living room. A Caucasian man in his late 60s with short white hair and a neatly trimmed white beard sits upright in a simple armchair. A static chest-up composition and soft frontal daylight keep his face natural, unobstructed, and sharply focused. He wears a pale blue shirt. Both hands rest separately on the armrests and remain fully visible. He looks directly into the lens, takes a small breath, and says in a warm, steady voice, "Rest is not a reward for finishing everything. Nothing is ever finished." He relaxes his shoulders slightly and gives one gentle nod. The audio is his voice, a distant bird outside the window, and quiet room ambience; the camera never moves.

LTX-2 Base

40 steps

OmniForcing

4 steps

Salt++ (TF-dCM route)

4 steps

Salt++ (AR–AR DMD route)

4 steps

1664×960 scale-wise generation

Two low-resolution and two high-resolution generator calls per block, still four calls in total.

Fitness coach · paper Fig. 6

[Shot 1][0.0s-5.0s]: In a minimalist fitness studio with a neutral grey wall, a mature Black man with a shaved head and a short grey beard stands on a black mat. The camera is locked in an eye-level medium shot, and broad soft lighting keeps his face, shoulders, and hands crisp without motion blur. He wears a maroon athletic shirt. Both hands begin open at waist height, palms angled inward and clearly separated. He brings them slowly upward by a few inches while maintaining direct eye contact and says in a calm, supportive voice, "You don't need more motivation; you need one smaller promise." His hands stop and remain still as he gives a gentle nod. The studio is quiet except for his resonant voice and faint ventilation hum.

LTX-2 Base

40 + 3 steps

OmniForcing

4 steps, direct 960p

Salt++ 480p

4 steps, bilinear 2×

Salt++ 960p

2 LR + 2 HR

Ablations

Score–context configuration for few-step initialization, and teacher-guidance calibration during distillation.

Cherry-blossom garden · paper Fig. 7

A serene garden with cherry blossom trees featuring pink and white flowers is shown. In the background, a traditional Japanese-style building with a sloped roof stands. The garden includes a pond with clear water surrounded by rocks and greenery. The sky is clear and bright, indicating a sunny day. Soft, gentle piano music plays in the background throughout.

TF-dCM

consistency distillation

BI–BI DMD

both scores bidirectional

BI–AR DMD

fake score causal

AR–AR DMD

context aligned

04 / Results

Quantitative results

JavisBench-mini and VBench, all causal models at four steps with CFG = 1.

+57%

Visual quality over OmniForcing, 480p, 4 steps

+45%

Motion quality over OmniForcing, 480p, 4 steps

1664×960

Four generator calls per block, scale-wise

Table 1(a). JavisBench-mini, official prompts, 480p

Model	Causal	Steps	VQ↑	MQ↑	AQ↑	CLIP↑	IB-AV↑	Javis↑	DeSync↓
LTX-2 Base	No	40	1.884	0.566	4.986	0.311	0.239	0.200	0.608
AR teacher (CSF)	Yes	40	2.304	0.857	4.565	0.316	0.202	0.162	0.746
OmniForcing	Yes	4	1.807	0.699	4.718	0.303	0.163	0.124	0.745
Salt++ (TF-dCM route)	Yes	4	2.013	0.826	4.976	0.313	0.229	0.185	0.710
Salt++ (AR–AR DMD route)	Yes	4	2.838	1.010	4.991	0.316	0.184	0.146	0.759

Bold is best and underline is second best among the 4-step rows.

Table 1(b). JavisBench-mini, official prompts, 1664×960

Model	Causal	Steps	VQ↑	MQ↑	AQ↑	CLIP↑	IB-AV↑	Javis↑	DeSync↓
LTX-2 Base	No	40+3	2.231	0.607	4.867	0.311	0.169	0.145	0.658
OmniForcing	Yes	4	2.298	0.832	4.721	0.293	0.165	0.128	0.755
Salt++	Yes	4	2.730	0.957	5.113	0.318	0.193	0.157	0.768

Bold marks the best result across all methods.

Table 1(c). VBench video quality and within-clip consistency, 480p

Model	Causal	Steps	Aesthetic↑	Imaging↑	Subject↑	Background↑
LTX-2 Base	No	40	53.89	67.06	95.94	95.63
OmniForcing	Yes	4	56.74	68.41	96.62	95.32
Salt++ (TF-dCM route)	Yes	4	53.65	62.87	96.58	96.05
Salt++ (AR–AR DMD route)	Yes	4	55.43	69.98	97.01	96.24

Table 2. 1,000 LTX-2-enhanced JavisBench prompts, 480p

Model	Causal	Steps	VQ↑	MQ↑	AQ↑	IB-AV↑	AVH↑	Javis↑	DeSync↓
LTX-2 Base	No	40	1.981	0.640	4.959	0.256	0.245	0.211	0.573
OmniForcing	Yes	4	1.774	0.607	4.613	0.158	0.152	0.121	0.734
Salt++ (TF-dCM route)	Yes	4	2.074	0.760	4.838	0.245	0.233	0.193	0.704
Salt++ (AR–AR DMD route)	Yes	4	2.820	0.999	4.997	0.192	0.186	0.151	0.696

05 / Ablations

Ablations

Which conditional distributions are compared, and which reference is distilled.

Table 3(a). Few-step initialization: score contexts and objectives

Objective	Scores	VQ↑	MQ↑	AQ↑	CLIP↑	IB-TV↑	IB-TA↑	DeSync↓
TF-dCM	–	1.916	0.611	4.686	0.307	0.264	0.145	0.766
TF-DMD	BI/BI	0.837	0.140	4.796	0.301	0.266	0.144	0.811
TF-DMD	BI/AR	1.131	0.157	4.571	0.287	0.255	0.149	0.531
TF-DMD	AR/AR	3.147	1.351	4.750	0.317	0.271	0.153	0.726

BI/AR denotes the visibility of the real and fake score models respectively.

Table 3(b). Teacher guidance under aligned AR–AR scores

Teacher guidance	VQ↑ (off.)	MQ↑ (off.)	VQ↑ (rew.)	MQ↑ (rew.)	AQ↑	CLIP↑	DeSync↓
Fixed 4.0	1.643	0.572	1.712	0.598	4.934	0.315	0.843
Randomized U(1.0, 3.5)	3.147	1.351	3.132	1.261	4.750	0.317	0.726

Guidance inherited from consistency distillation versus guidance calibrated for DMD.

06 / BibTeX

BibTeX

@misc{ge2026saltcontextalignedposttrainingfewstep,
  title     = {Salt++: Context-Aligned Post-Training for Few-Step Streaming
               Multimodal Generation},
  author    = {Xingtong Ge and Yutong Wang and Lunjie Zhu and Haitao Lin and
               Fangyu Lin and Yushi Huang and Xin Zhang and Yi Zhang and
               Yu Liu and Jun Zhang},
  year      = {2026},
  eprint    = {2609.36995},
  archivePrefix = {arXiv},
  primaryClass = {cs.CV},
  url       = {https://arxiv.org/abs/2609.36995}
}

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

Fast motion at the same four-step budget

Abstract

Method

Generated results

480p streaming generation

1664×960 scale-wise generation

Ablations

Quantitative results

Ablations

BibTeX