Directly aligning trajectory-distilled video generators slashes rollout costs and optimizes the deployed few-step model without teacher fine-tuning or re-distillation. However, conventional flow-matching preference objectives assume single-head local velocity supervision, mismatching the finite-interval, block-average transition interface of trajectory-distilled generators such as Parallel Decoding Distillation (PDD) and Parametric Trajectory Distillation (PTD). Naively supervising individual unfused sub-interval velocity heads across all sub-steps rather than the fused block-average velocity at block starts (DirectAlign) induces a severe training--inference mismatch that degrades preference below the unaligned baseline. We introduce BlockAlign, an interface-consistent preference optimization framework that constructs states at valid block boundaries, queries the fused block-level velocity predictor, and supervises span-normalized transition residuals---seamlessly supporting DPO-, NFT-, and rollout-style objectives while preserving the native few-step sampler. We further build a unified post-training pipeline integrating offline pairs, reward-free semi-online continuations, and online rollouts with hierarchical preference selection and objective-aware block scheduling. Experiments on multi-reference video generation across MiniMax-H3 Ref2VA and an in-house model RefVid-V4 show that BlockAlign consistently outperforms unaligned distilled models and DirectAlign in structural plausibility, reference consistency, and visual quality. Notably, our approach unlocks PDD's 5x few-step rollout speedup for 3x faster end-to-end online training, achieving statistically significant human preference gains over PDD baselines (54.0% vs. 33.3% wins/losses, or 61.8% non-tie win rate on RefVid-V4; 51.3% vs. 34.7%, or 59.7% non-tie win rate on MiniMax-H3) and up to an 82.7% win rate against DirectAlign.
Figures & tables
Figure 1: Qualitative comparison on RefVid-V4. All methods use identical inputs, random seeds, inference settings, and matched timestamps. Red circles highlight representative baseline errors in the geometry and spatial relationships among the hands, apples, and basket handles. BlockAlign produces more plausible interactions while preserving the referenced identity, clothing, accessories, and basket. It also yields clearer, more saturated frames with fewer patch-like texture artifacts. Best viewed zoomed in.
Figure 2: Qualitative comparison on MiniMax-H3. Red circles highlight representative baseline errors in arm, sleeve, scarf, and garment structure. BlockAlign produces more coherent anatomy and garment geometry while preserving the referenced face, coat, trousers, boots, tote bag, scarf, and earrings across matched timestamps. Best viewed zoomed in.
Backbone
Method
SC
PF
VQ
DF
ST
Mean
PDD
3.35
3.79
3.93
3.55
3.75
3.67
RefVid-V4
DirectAlign
3.37
3.73
3.93
3.61
3.71
3.67
BlockAlign
3.41
3.81
3.94
3.64
3.75
3.71
PDD
3.41
3.62
3.96
3.60
3.65
3.65
MiniMax-H3
DirectAlign
3.40
3.61
3.95
3.55
3.59
3.62
BlockAlign
3.50
3.65
3.96
3.75
3.66
3.70
Table 1: Automated absolute assessment scores (1–5 scale; higher is better). Mean is the unweighted average over the five dimensions.
Backbone
Pair
SC
PF
VQ
DF
ST
Mean
Agr.
DirectAlign vs. PDD
45.7
42.8
42.0
50.8
43.3
44.9
55.9
RefVid-V4
BlockAlign vs. PDD
52.5
49.2
57.5
56.5
53.3
53.8
59.6
BlockAlign vs. DirectAlign
58.7
60.8
66.7
58.3
60.3
61.0
68.0
DirectAlign vs. PDD
38.0
41.7
37.3
38.2
43.3
39.7
57.7
MiniMax-H3
BlockAlign vs. PDD
52.5
55.2
52.8
55.3
53.8
53.9
57.2
BlockAlign vs. DirectAlign
62.5
61.7
67.3
65.7
62.0
63.8
58.9
Table 2: Adjusted pairwise win rates (%) for the first method against the second; 50% denotes parity. Agr. denotes bidirectional agreement averaged over dimensions.
Figure 3: Human preference evaluation on MiniMax-H3 (top) and RefVid-V4 (bottom). Bars show condition-level win, tie, and loss rates over 150 comparisons per method pair, aggregated from five independent annotators per pair. On both backbones, BlockAlign is consistently preferred over both DirectAlign and the original PDD model, while PDD is preferred over DirectAlign.
Appendix figures & tables23 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Qualitative comparison on MiniMax-H3 (PTD, Case 1). Both methods use identical inputs, random seeds, inference settings, and matched timestamps. The red circle highlights background blurring and spatial displacement of the wooden chair in the unaligned PTD baseline. BlockAlign preserves stable chair geometry and coherent scene structure alongside the referenced identity, knit top, trousers, flats, and ring. Best viewed zoomed in.
Figure 5: Qualitative comparison on MiniMax-H3 (PTD, Case 2). Red circles mark a severe anatomical deformity in the unaligned PTD baseline, which generates three feet/shoes. BlockAlign produces anatomically plausible lower-body structure and natural stepping motion while preserving the referenced face, jacket, skirt, handbag, socks, and Mary Jane shoes. Best viewed zoomed in.
Figure 6: Qualitative comparison on MiniMax-H3 (PTD, Case 3). Red circles highlight severe physical interpenetration between the khaki canvas tote and the blue cargo shorts in the unaligned PTD baseline. BlockAlign maintains physically valid object boundaries and natural hand–tote interactions while faithfully adhering to the referenced outfit and multi-shot prompt. Best viewed zoomed in.
Method
SC
PF
VQ
DF
ST
Mean
PTD
3.73
3.85
3.98
3.70
3.85
3.82
BlockAlign
3.73
3.88
3.98
3.74
3.88
3.84
Appendix
Table 3: Absolute assessment scores for the MiniMax-H3 (PTD) experiment (500 optimization steps; higher is better). Mean is the unweighted average over the five dimensions.
Dimension
Win Rate
Agr.
Subject Consistency
53.8
44.0
Prompt Following
50.5
56.0
Visual Quality
54.5
38.7
Deformation
55.0
42.7
Shot Transitions
50.2
73.3
Mean
52.8
50.9
Appendix
Table 4: Adjusted pairwise win rates (%) for BlockAlign (500 steps) against PTD on MiniMax-H3; 50% denotes parity. Agr. denotes bidirectional agreement between presentation orders.
Figure 7: Qualitative evaluation of BlockAlign-ART on MiniMax-H3 after 400 optimization steps under identical conditioning inputs, random seeds, inference settings, and matched timestamps. Top: The prompt specifies a single black shoulder bag; PDD produces an additional bag-like structure around the subject’s back and arm (red circles), whereas BlockAlign-ART consistently maintains a single referenced bag across viewpoints. Bottom: When the subject removes and holds her sunglasses, PDD generates a malformed handheld object conflated with the referenced bag, whereas BlockAlign-ART produces recognizable sunglasses and a plausible hand–object interaction. Best viewed zoomed in.
Method
SC
PF
VQ
DF
ST
Mean
PDD
3.41
3.62
3.96
3.60
3.65
3.65
BlockAlign-ART
3.44
3.69
3.98
3.66
3.68
3.69
Appendix
Table 5: Absolute assessment scores for the rollout-based BlockAlign-ART experiment on MiniMax-H3 (400 optimization steps; higher is better). Mean is the unweighted average over the five dimensions.
Dimension
Win Rate
Agr.
Subject Consistency
53.7
46.7
Prompt Following
51.3
60.0
Visual Quality
52.7
43.3
Deformation
51.0
41.3
Shot Transitions
51.5
74.7
Mean
52.0
53.2
Appendix
Table 6: Adjusted pairwise win rates (%) for BlockAlign-ART (400 steps) against PDD on MiniMax-H3; 50% denotes parity. Agr. denotes bidirectional agreement between presentation orders.
Figure 8: Additional RefVid-V4 comparison. Red circles highlight implausible hand–arm configurations and garment boundaries in the baselines. BlockAlign better preserves the referenced identity and outfit while producing clearer and less patchy local textures. Best viewed zoomed in.
Figure 9: Garment-reference consistency on RefVid-V4. Red circles mark deviations from the referenced sweatshirt in DirectAlign. BlockAlign better preserves its shape, closure, printed details, color, and associated accessories, with clearer and less patchy fabric textures. Best viewed zoomed in.
Figure 10: Scene-reference consistency on RefVid-V4. Red circles highlight corrupted storefront text in PDD and scene-layout deviations in DirectAlign. BlockAlign produces a more coherent conditioned scene and cleaner local details. Best viewed zoomed in.
Figure 11: Additional MiniMax-H3 comparison. Red circles highlight malformed hands and implausible interactions with the leaf or handbag. BlockAlign produces more coherent anatomy and interactions while preserving the referenced identity, garments, handbag, and shopfront. Best viewed zoomed in.
Figure 12: Additional MiniMax-H3 comparison with seated and walking poses. Red circles indicate malformed hands, implausible arm–torso configurations, and inconsistent lower-body structure. BlockAlign produces more coherent anatomy while preserving the referenced garments and accessories. Best viewed zoomed in.
Figure 13: Additional MiniMax-H3 comparison involving a wrapped object, kneeling motion, and a referenced cat. Red circles mark baseline errors in hand–object, lower-body, and subject–animal interactions. BlockAlign produces more plausible interactions while preserving the conditioned identity, clothing, accessories, and cat. Best viewed zoomed in.
Visual Quality
Motion (Quality & Amplitude)
Subject Consistency
Text Align.
Realism
Domain
Method
Aes ↑
Smooth. ↑
Motion-FB ↔
Motion-LK ↔
FaceSim-CUR ↑
FaceSim-ARC ↑
Nexus ↑
GME ↑
Natural ↑
Open-Domain (180 cases)
PDD
5.5259
0.9913
0.2232
4.8244
0.5406
0.4962
0.0282
0.6525
4.5574
DirectAlign
5.4663
0.9886
0.2329
5.1657
0.5488
0.5028
0.0245
0.6515
4.5833
BlockAlign
5.5591
0.9912
0.2003
4.4738
0.5455
0.5053
0.0237
0.6554
4.5111
Human-Domain (60 cases)
PDD
5.7233
0.9919
0.2728
5.4487
0.5581
0.5058
0.0053
0.6559
4.3444
DirectAlign
5.6798
0.9914
0.2812
5.6850
0.5692
0.5187
0.0048
0.6530
4.5167
Appendix
Table 7: Quantitative evaluation on the public OpenS2V-Eval benchmark [ 55 ] across all four official domains ( 360 videos per model; 1,080 videos total). We report raw metric means grouped into five functional categories. Note that Motion-FB and Motion-LK measure raw optical-flow displacement magnitude ( ↔ : amplitude indicator, not higher-is-better) and should be read alongside Smooth. ( ↑ ) and GME ( ↑ ). Symbols --- indicate subsets without applicable human-face references or official class_label metadata.
Eval. Protocol (RF/IR Judge)
Model
Content
Motion
Style
Structure
Narrative
Multi-content
Cross-aspect
Overall ↑
5-second Generation (Official Leaderboard & Ours, Gemini-3.1-Pro for RF / IR )
Seedance 2.5
78.88
66.10
67.13
73.35
73.97
77.77
71.53
72.68
MiniMax-H3
78.86
65.32
66.24
72.94
73.90
77.22
72.36
72.41
Seedance 2.0
79.00
63.05
63.73
68.36
73.76
76.63
71.83
70.91
Happy Horse 1.0
75.90
68.58
65.62
69.46
72.34
74.62
69.71
70.89
Gemini Omni
75.02
62.25
69.24
70.13
75.11
73.36
70.18
70.76
Kling 3.0 Omni
75.99
64.20
55.90
71.12
68.59
75.98
68.33
68.59
Appendix
Table 8: Task-family and overall quantitative evaluation on the public OmniVBench benchmark [ 19 ] ( 813 videos per model across seven task families). Each task family score averages Reference Fidelity ( RF ), Instruction Realization ( IR ), and Video Quality ( VQ ), and Overall denotes the macro-average across the seven families. Note that only RF and IR query a proprietary VLM judge, whereas VQ ( Technical , Aesthetic , and Physics ) is computed directly using fixed open-weight evaluation models and remains identical across VLM judges. Top block: 5-second generation results under the default Gemini-3.1-Pro judge for RF / IR , comparing our 10-step BlockAlign ( 24fps , 768p ) against official leaderboard entries. Middle block: controlled 8-second comparison ( 24fps , 768p ) under the Gemini-3.1-Pro judge for RF / IR . Bottom block: cross-judge verification using GPT-4o (the VLM judge used in OpenS2V-Eval) for RF / IR combined with the same fixed-model VQ scores. Best results within each 8-second comparison group (and top 5-second leaderboard entries) are highlighted in bold .
Metric / Sub-metric
Method
Content (112)
Motion (95)
Style (43)
Structure (108)
Narrative (120)
Multi-content (112)
Cross-aspect (223)
Macro ↑
Reference Fidelity (RF) ↑ ( Gemini-3.1-Pro Judge)
PDD (8-second)
91.15
46.54
61.05
78.40
87.90
93.75
79.88
76.95
DirectAlign (8-second)
92.41
41.84
65.70
75.89
87.57
92.95
79.79
76.59
BlockAlign (8-second)
91.52
43.64
59.30
79.90
87.29
92.22
79.82
76.24
BlockAlign (5-second)
94.01
47.81
63.37
78.78
81.93
92.06
79.19
76.74
Reference Fidelity (RF) ↑ ( GPT-4o Judge)
PDD (8-second)
99.11
71.05
76.16
98.26
98.96
99.65
94.80
91.14
DirectAlign (8-second)
99.00
72.89
79.07
97.86
98.54
99.47
94.50
91.62
Appendix
Table 9: Fine-grained breakdown on OmniVBench across Reference Fidelity ( RF ), Instruction Realization ( IR ), Video Quality ( VQ ), and the three VQ sub-dimensions ( Technical , Aesthetic , and Physics ) for all seven task families and the 7-family Macro average. For RF and IR , we report both the default Gemini-3.1-Pro judge and cross-judge verification with GPT-4o (the VLM judge used in OpenS2V-Eval). Within our controlled 8-second comparison, DirectAlign degrades Technical and Aesthetic quality below the unaligned PDD baseline across every single task family ( 7/7 ) as well as macro VQ ( 41.96 vs. 42.12 ), whereas BlockAlign (8-second) achieves the best Technical , Aesthetic , and VQ scores across all seven families ( 7/7 , bolded within the 8-second group). We also report BlockAlign (5-second) at the bottom of each block.
V-JEPA 2
VideoMAE v2
Backbone
Method
Cosine distance
ℓ2 distance
Cosine distance
ℓ2 distance
RefVid-V4
PDD
0.0238±0.0067
9.32±1.31
0.0438±0.0158
66.94±9.50
BlockAlign
0.0258±0.0072
9.74±1.37
0.0453±0.0159
68.90±8.89
MiniMax-H3
PDD
0.0308±0.0094
10.65±1.54
0.0432±0.0139
69.06±9.22
BlockAlign
0.0293±0.0087
10.37±1.48
0.0420±0.0140
67.74±9.51
Appendix
Table 10: Intra-condition generation diversity across five random seeds under identical reference images and text prompts. We report the mean and standard deviation of pairwise feature distances across test conditions (higher values indicate greater feature-space dispersion). Distances are computed independently within each encoder’s feature space.
Figure 14: Qualitative diversity comparison on RefVid-V4 (Case 1). Each row shows a video generated from a distinct random seed, and columns show matched timestamps. Corresponding rows for PDD and BlockAlign share identical seeds, reference images, and prompts. Both models produce diverse poses, camera framings, subject placements, and indoor–outdoor scene realizations while preserving the referenced identity and outfit, with no sign of mode collapse after BlockAlign post-training. Best viewed zoomed in.
Figure 15: Qualitative diversity comparison on RefVid-V4 (Case 2). Across five shared random seeds, BlockAlign retains rich variation in pose, viewpoint, framing, motion trajectory, and spatial layout while faithfully adhering to the reference and prompt conditions. Best viewed zoomed in.
Figure 16: Qualitative diversity comparison on MiniMax-H3 (Case 1). Across five shared random seeds and matched timestamps, BlockAlign preserves clear variation in seated and standing postures, walking trajectories, camera distances, subject placement, and handbag orientation while maintaining reference fidelity. Best viewed zoomed in.
Figure 17: Qualitative diversity comparison on MiniMax-H3 (Case 2). Different random seeds yield distinct realizations of the multi-shot prompt—varying in pose, facial expression, gesture, framing, and subject placement—confirming that BlockAlign improves structural plausibility without inducing mode collapse. Best viewed zoomed in.