Few-step sampling reduces the cost of streaming video generation. Distribution Matching Distillation (DMD) supervises few-step generators with a separate fake-score network, which adds training memory and computation and must track the evolving generator. We extend APEX to streaming video by using condition shifting to estimate the velocity field of the generated distribution within the generator. The original text condition selects generation, while a shifted condition selects fake estimation using the same parameters. This removes the independent fake network while retaining S2-Avatar's frozen reference teacher, Self-Replay Forcing, and perceptual regularization. We evaluate APEX on 180-second videos using three backbone evaluations per block and a one-step spatial Refiner, with a 70-case human study and a 320-case StreamAV-Bench comparison. Compared with S2-Avatar, APEX achieves higher video instruction fulfillment on the Progressive track (3.856 versus 3.831) and higher visual aesthetics across both tracks (0.641898 versus 0.632606). Human raters favor APEX most strongly on instruction following, with +30.0% net preference, followed by consistency and video quality at +4.3% each. Motion quality favors S2-Avatar, audiovisual synchronization is predominantly tied, and usability is unchanged. Separate measurements under matched training settings show 4.4% lower peak allocated GPU memory per GPU and 9.2% shorter mean iteration time than DMD.
Figures & tables
Figure 1: APEX streaming generation and interaction. Top: Progressive (left) and Interactive (right) StreamAV-Bench samples (3 NFE, 2× Refiner). The lower four rows show object interaction, turning, dancing, and storytelling, with each sequence read from left to right. Quoted user prompts are translated from Chinese, with rounded onset times. All timestamps are in seconds. Demo times are measured relative to the start of each recording.
Figure 2: DMD and APEX in Self-Replay Forcing. DMD uses independent fake parameters η ; APEX shares generator and fake parameters θ and, by default, AdamW states. Teacher parameters ϕ are frozen. Left: xθ is the rollout/replay output, t′ the scoring time, dt′ the normalized correction, and \mathop{{\color[rgb]{1,0,0}\operatorname{sg}}}\nolimits stop-gradient. The correction contributes ∇θLscore ; L is the full generator objective. The fake-velocity subscript θ identifies the generated distribution. Right: H and Hout denote input and output hidden states; HC contains projected text tokens c . Within each row, Q,K,V are queries, keys, and values, and Aattn=softmax(QK⊤/dh) , with head dimension dh . Superscript shift marks states under cshift=−c . The single-head example uses fixed weights and omits the causal mask and other transformer operations; values are schematic.
Figure 3: Paired examples at 3 backbone NFE. (a) APEX retains both bags late in the sequence, while S2-Avatar shows one. (b) APEX brings the tube to the other hand for the applying instruction; S2-Avatar continues holding it at its side. Each pair uses the same reference and timestamps, with a common one-step 2× Refiner.
Model
VA ↑
VQ ↑
PQ ↑
AQ ↑
AVAlign ↑
AVSync ↓
VIF ↑
AIF ↑
SC ↑
BC ↑
SoulX-FlashTalk
0.281
2.515
7.091
3.006
0.121
0.648
–
2.499
0.946
0.964
AptAvatar
0.458
2.040
7.075
2.582
0.099
1.284
–
2.448
0.860
0.923
OmniForcing
0.554
2.374
6.351
2.528
0.119
1.423
2.423
2.343
0.962
0.953
Hallo-Live
0.574
–
5.396
–
0.102
1.353
–
–
0.991
0.977
Self-Forcing
0.585
2.753
6.453
2.623
0.256
0.919
3.019
2.810
0.981
0.969
IAMFlow
0.602
2.840
6.415
2.625
0.268
1.016
3.192
2.789
0.986
0.973
Table 1: Comparison on ten StreamAV-Bench metrics. VIF, AIF, SC, and BC use Progressive; the other metrics use both tracks. APEX and S2-Avatar use our evaluation results; baseline VIF is from StreamAV-Bench ( Liu et al., 2026f ) and other baseline scores from Vidu S2 ( Zhang et al., 2026c ) . Values are rounded to three decimals. Bold and underlined values mark the best and second-best available results in each column; a dash marks an unreported value. AVSync uses seconds.
Metric
Overall ( n=320 )
Progressive ( n=160 )
Interactive ( n=160 )
APEX
S2
Δ
APEX
S2
Δ
APEX
S2
Δ
VQ ↑
3.256
3.267
−0.011
3.180
3.131
+0.049
3.333
3.403
−0.070
AQ ↑
2.725
2.788
−0.063
2.706
2.809
−0.103
2.744
2.766
−0.022
VIF ↑
3.891
3.834
+0.057
3.856
3.831
+0.025
3.925
3.836
+0.089
AIF ↑
2.610
2.570
+0.040
2.835
2.768
+0.067
2.385
2.372
+0.013
VA ↑
0.641898
0.632606
+0.009292
0.645374
0.636402
+0.008972
0.638422
0.628810
+0.009612
Table 2: Per-track and pooled StreamAV-Bench results. Progressive and Interactive each contain 160 cases; Overall pools all 320 cases for every metric. Values retain source precision, and Δ is APEX minus S2-Avatar. AVSync is measured in seconds and is lower-is-better; all other reported metrics are higher-is-better.
Figure 4: Paired human preferences at 3 backbone NFE. APEX versus S2-Avatar on 70 I2V sequences of 180 seconds, with a common one-step 2× Refiner. Good, Same, and Bad denote APEX wins, ties, and losses.
Model
Instruction MOS
Synchronization MOS
Usable sequences
Usability
APEX
4.17
4.23
62/70
88.6%
S2-Avatar
3.99
4.26
62/70
88.6%
Table 3: Absolute human ratings and usability. Usable sequences have both MOS ratings above 2.
Figure 5: Training cost of DMD and APEX. Peak allocated memory per GPU (GiB) and mean iteration time (seconds) under matched training settings. Percentage reductions are measured relative to DMD.
Figure 6: Human preferences at 2 backbone NFE. Each model is evaluated against S2-Avatar at 3 NFE in a separate 70-case paired comparison. All configurations use the same one-step 2× Refiner.
Figure 7: APEX at 2 backbone NFE versus S2-Avatar at 3. Selected face-contact, forward-leaning, and side-bending frames have similar overall appearance, with differences in jewelry, hand placement, and body pose. Reference images, frame timestamps, and the one-step 2× spatial Refiner are matched across methods.
Configuration
NFE
Consistency
Instruction
Sync.
Video
Motion
Default
2
−2.9
+27.1
−7.1
−2.9
−7.1
DF10
2
−18.6
+21.4
−5.7
−8.6
−4.3
DF4
2
−8.6
+17.1
−8.6
−11.4
−8.6
Default
3
−5.7
+21.4
−7.1
−5.7
+2.9
DF10
3
−11.4
+11.4
−7.1
−11.4
+1.4
DF4
3
−14.3
+10.0
−10.0
−7.1
+1.4
Table 4: Sampling and optimizer configurations. Net preference (%) over S2-Avatar at 3 NFE; each block is a separate 70-case round. Sampling variants share optimizer states; optimizer rows use AdamW and also differ in NFE.
Understanding the design space of streaming video diffusion is essential to exploring its potential for generation quality and computational efficiency. We develop a unified analytical framework that relates model and sampler choices, historical conditioning, execution scheduling, and training strategies. The framework accommodates a broad family of causal context-selection policies and makes their computational dependencies and training-inference alignment explicit. Within this design space, we study three representative policies: clean, same-level, and progressive history. On the full VBench prompt set, same-level and progressive history achieve aggregate scores of 85.24 and 85.60, respectively, compared with 84.45 for the clean-history reference. Long-video comparisons further show improved subject consistency and more coherent motion with progressive history. By allowing multiple denoising nodes to be processed together, progressive-history pipelining achieves 1.57-2.83× steady-state DiT speedups under our evaluated conditions. We additionally find that LoRA adaptation of the DMD fake-score network improves generation quality using only 2.15% as many trainable fake-score parameters as full-parameter adaptation. Together, these findings show that fully denoised history is not a prerequisite for high-quality streaming generation and motivate the joint design of historical conditioning, execution, and training.
The push for efficient text to image synthesis has moved the field toward one step sampling, yet existing methods still face a three way tradeoff among fidelity, inference speed, and training efficiency. Approaches that rely on external discriminators can sharpen one step performance, but they often introduce training instability, high GPU memory overhead, and slow convergence, which complicates scaling and parameter efficient tuning. In contrast, regression based distillation and consistency objectives are easier to optimize, but they typically lose fine details when constrained to a single step. We present APEX, built on a key theoretical insight: adversarial correction signals can be extracted endogenously from a flow model through condition shifting. Using a transformation creates a shifted condition branch whose velocity field serves as an independent estimator of the model's current generation distribution, yielding a gradient that is provably GAN aligned, replacing the sample dependent discriminator terms that cause gradient vanishing. This discriminator free design is architecture preserving, making APEX a plug and play framework compatible with both full parameter and LoRA based tuning. Empirically, our 0.6B model surpasses FLUX-Schnell 12B (20× more parameters) in one step quality. With LoRA tuning on Qwen-Image 20B, APEX reaches a GenEval score of 0.89 at NFE=1 in 6 hours, surpassing the original 50-step teacher (0.87) and providing a 15.33× inference speedup. Code is available https://github.com/LINs-lab/APEX.
Deyuan Liu, Peng Sun, Yansen Han +3
Westlake University · Zhejiang University · Shanghai Innovation Institute +1
Few-step autoregressive video generation enables low-latency streaming, yet prevailing training pipelines separate causal adaptation from distribution matching. Recent representation-based text-to-video methods retain this separation by starting with an ODE initialization stage, even when the target distribution is defined directly by reference videos. We introduce \method, the first single-stage framework that uses reference videos both as paired examples of how a history continues and as a target population for self-generated rollouts. Specifically, Direct Forcing is trained with two complementary signals: (1) Causal Continuation Alignment learns paired spatiotemporal correspondence under reference histories; and (2) Rollout Distribution Matching employs maximum mean discrepancy to align the representation distributions of self-generated and reference videos. To coordinate these signals, we introduce Dynamic Gradient Balancing, which adapts their relative gradient contributions during training. Experiments with Wan2.1-1.3B show that Direct Forcing learns a four-step autoregressive generator in 250 updates without a separate ODE initialization stage or online diffusion-score supervision, achieving 84.63 VBench Total and 86.01 Quality.