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.
Figures & tables
Figure 1: Representative historical context policies. Columns correspond to video chunks and rows to denoising levels. Highlighted historical states condition the current denoising update shown in orange: (a) clean history, (b) same-level history, and (c) progressive history.
Figure 2: Fully serial and pipelined denoising. (a) Serial execution completes all K updates of one chunk before starting the next, yielding BK sequential rounds. (b) Pipelined execution advances multiple chunks at staggered denoising levels, yielding B+K−1 rounds under sufficient parallel resources while preserving the same BK chunk-level updates.
Context policy
Total
Quality
Semantic
Clean: ϕ(j,r)=0
84.446
84.862
82.783
Same-level: ϕ(j,r)=j
85.240
85.541
84.038
Progressive: ϕ(j,r)=max(j−r,0)
85.599
85.961
84.149
Table 1: Generation quality under three context policies. Higher is better.
Figure 3: Three-policy comparisons under matched conditions and policy-specific weights. (a) Short-video example for “a bus accelerating to gain speed” (seed 0). (b) Long-video bee example (seed 1). Columns show the indicated times.
Fake-score adaptation
Trainable params.
Total ↑
Quality ↑
Semantic ↑
Full-parameter
14.288B
84.126
84.477
82.720
LoRA
0.307B
85.599
85.961
84.149
Table 2: Fake-score adaptation under progressive history.
Devices (CP)
Clean serial (ms/chunk)
Progressive pipeline (ms/chunk)
Speedup
4
527.7
336.0
1.57×
8
441.7
222.4
1.99×
12
626.3
221.0
2.83×
Table 3: Steady-state DiT latency (ms/chunk) and speedup.
Appendix figures & tables2 assets
Supplementary material from the paper’s appendix.
Appendix
Item
Setting
Student initialization
Causal-rCM; Wan2.1-T2V-1.3B; TF-sCM; c3-3 .
Teacher / fake-score backbone
Wan2.1-T2V-14B; frozen teacher.
Training samples
Student-generated videos; 81 frames at 16 FPS, approximately five seconds.
VAE temporal compression
Wan2.1 VAE; 4× temporal compression; 81 video frames correspond to 21 latent frames.
Video chunking
Three latent frames per chunk; seven chunks.
Spatial resolution
832×480 .
Appendix
Table 4: Shared training configurations for clean, same-level, and progressive context adaptation with fake-score LoRA. The additional full-parameter fake-score variant is evaluated under progressive history only. The student is fully trainable in all configurations.
Dimension
Clean
Same-level
Progressive
Quality dimensions
Subject consistency
95.196
95.801
96.286
Background consistency
95.489
95.212
95.406
Temporal flickering
98.511
98.424
98.453
Motion smoothness
98.070
97.886
98.089
Aesthetic quality
65.285
66.352
66.082
Appendix
Table 5: Complete VBench dimension-level results on the full prompt set. Values are raw scores multiplied by 100, before normalization for aggregation. Higher is better in each dimension; the best result in each row is bold, including ties.