Glance Forcing: Making Bidirectional Diffusion Autoregressive at a Glance

1 INTRODUCTION

• More than 92.5% of trajectory curvature lies in the high-noise regime, and differing regime losses make one-sample training of a monolithic student prone to representation interference.The high-noise regime is associated with semantic layout and motion, while low noise concerns detail refinement.Source: §1 · p. 2
• Glance Forcing uses a dual-LoRA architecture with a Slow expert for high-noise semantic synthesis and a Fast expert for low-noise detail refinement.It trains only LoRA parameters, keeps the pretrained teacher frozen, and uses temporal-frequency and spatial high-pass objectives without adding a student network.Source: §1 · p. 2
• VBench Total rises from 84.04 to 84.22 while trajectory-collection cost falls from thousands of GPU-hours to about one A100-hour.With a matched training budget, Glance Forcing reaches a peak VBench Total score of 84.33.Source: §1 · p. 2
• The paper argues that PF-ODE distillation primarily straightens few-step trajectories, while one trajectory can suffice when exploited across noise regimes.Its contributions also report that the method matches or surpasses other methods while reducing training cost by orders of magnitude.Source: §1 · p. 2

2 RELATED WORKS

• Bidirectional video diffusion denoises clips jointly, with quadratic sequence-length complexity that makes it impractical for interactive low-latency scenarios.Autoregressive generation instead performs causal rollout across frames or chunks while retaining diffusion generation within each segment.Source: §2 · p. 3
• Autoregressive methods use Teacher Forcing or Diffusion Forcing, while recent systems address train–test discrepancy and architectural gaps but still rely on costly dense PF-ODE trajectories.Autoregressive factorization supports streaming by using prior blocks as context through a key-value cache.Source: §2 · p. 3
• Trajectory-based distillation compresses transport paths into few-step generators, while score distillation matches generated and target distributions; both lines have stated shortcomings.Trajectory methods assume sufficiently smooth paths for compression, whereas the cited video score-distillation methods are described as neglecting intrinsic spatiotemporal dynamics.Source: §2 · p. 3

3.1 PRELIMINARY

• Slow-LoRA handles coarse structure in early denoising stages, while Fast-LoRA refines details later in the chunk-wise autoregressive framework.Consistency distillation trains a student to map noisy trajectory states to the ODE endpoint.Source: Figure 3 · p. 4, §3.1 · p. 4

3.2 WHY CAN ONE SAMPLE SUFFICE?

• ODE initialization improves four-step trajectory straightness and directional alignment, while increasing supervision from one trajectory to 3K changes geometry only marginally.The subsequent DMD stage improves generation quality while leaving trajectory geometry largely unchanged.Source: §3.2 · p. 5

3.3 SUBOPTIMAL TRAJECTORY FITTING IN CONSISTENCY DISTILLATION

• More than 92.5% of trajectory curvature lies in the high-noise region, making uniform-timestep regression to clean endpoints an inconsistent target.The method separates high-noise layout and motion from low-noise refinement and fine-detail generation, fitting each sub-trajectory with a specialized student.Source: §3.3 · p. 5
• The Slow and Fast experts outperform vanilla VCM in modeling structural semantics and fine details, respectively.The authors report that decoupling trajectory optimization alleviates the capacity bottleneck of a unified model fitting conflicting objectives.Source: §3.3 · p. 6

3.4 EXPERT-SPECIFIC OPTIMIZATION OBJECTIVE

• A temporal-frequency loss guides Slow toward coherence, while a spatial high-pass loss guides Fast toward sharp textures and structural details.The temporal loss suppresses high-frequency flickering while preserving motion dynamics; the Laplacian-based loss encourages crisp boundaries and fine textures without compromising semantic structure.Source: §3.4 · p. 6

4.1 SETUP

• The evaluation uses Wan2.1-1.3B to generate 480×832 videos with 81 frames, and reports VBench, VisionReward, Dynamic Degree, Instruction Following, latency, and throughput.Training is conducted on 8× A100 GPUs; latency and throughput are measured on a single H100 GPU, excluding VAE-related time.Source: §4.1 · pp. 6–7

4.2 MAIN RESULTS

• Glance Forcing records Total scores of 82.09 with one sample, 82.69 at matched budget, and 84.58 after autoregressive distillation, with further Dynamic, VisionReward, and Instruction Following gains.The 82.09 one-sample setting also has higher Dynamic, VisionReward, and Instruction Following scores than Causal Forcing; 82.69 is the best overall Total among initialized autoregressive models.Source: Table 1 · p. 7, §4.2 · p. 7
• At matched training budget, Glance Forcing produces more temporally consistent frames, sharper textures, and clearer scene details than its one-sample setting.With one training sample, it also produces visually competitive videos with more stable appearance, motion, and fine-grained details than existing distilled autoregressive models in several cases.Source: §4.2 · p. 7, Figure 6 · p. 8
• Human annotators prefer Glance Forcing over Self Forcing and Causal Forcing, with win rates of 91.3% and 63.3%, respectively.The study uses three annotators and 50 prompts per comparison.Source: §4.2 · p. 7, Table 2 · p. 8

4.3 ABLATION STUDIES

• Slow–Fast optimization raises the overall score from 83.98 to 84.33; TFS adds further quantitative gains, and HP further improves overall quality and spatial fidelity.The reported qualitative effects are improved temporal consistency from TFS and sharper fine-grained spatial details from HP.Source: §4.3 · p. 8, Table 3 · p. 8, Figure 7 · p. 8
• Within each four-step chunk, S2–F2 gives the best overall performance, while S4 favors visual quality and F4 better preserves motion dynamics and semantic consistency.The imbalanced S1–F3 and S3–F1 schedules degrade performance.Source: §4.3 · p. 8, Table 6 · p. 9

4.4 DISCUSSION

• With one training sample, an 8-step bidirectional model distilled from 50-step Wan2.1 achieves comparable performance at over 5× lower sampling cost, with modest degradation at 4 NFEs.This tests Glance Forcing beyond autoregressive generation.Source: Table 5 · p. 9, §4.4 · p. 9
• One-sample initialization is more sensitive to high motion and specialized content, while DMD improves quality across all tested samples; semantic scores remain relatively stable.DMD gains are particularly pronounced for High Motion and Simple Geometry, and effectiveness depends on the selected sample’s characteristics.Source: §4.4 · p. 9, Table 4 · p. 9

• A single PF-ODE trajectory suffices when denoising is decoupled by noise regime, with trajectory correction and distributional refinement playing separable roles.The proposed dual-LoRA design assigns layout and motion to Slow and detail to Fast, each with a tailored objective.Source: §5 · p. 9
• Using one trajectory, Glance Forcing matches or surpasses the baseline while reducing trajectory-collection cost from thousands of GPU-hours to about one A100-hour.The conclusion reports consistent gains as the training budget scales.Source: §5 · p. 9

• Generative AI tools assisted with translation during drafting.Source: AI USE STATEMENT · p. 10
• The authors state they did not use generative AI for dataset generation, model development, mathematical claims or proofs, hypotheses, methodology, implementation, data analysis, or results interpretation.Source: AI USE STATEMENT · p. 10
• The authors used AI tools to improve readability, reviewed and revised the assisted text, and take responsibility for the final content.Source: AI USE STATEMENT · p. 10

• The flow-matching scheduler uses a shifted noise level with shift factor k = 5 for generator rollouts and DMD/GAN critic timestep sampling.The model predicts vθ(xt, t, c) and converts it to a clean latent estimate during generation.Source: A MORE IMPLEMENTATION DETAILS · p. 13
• Training constructs a 3K-sample synthetic dataset and a 3K-sample ODE Slow-Fast dataset, then performs 1K teacher-forced causal ODE-distillation steps before asymmetric DMD.The datasets use about the same prompts; the ODE student is initialized from the autoregressive diffusion teacher.Source: A MORE IMPLEMENTATION DETAILS · p. 13
• Inference uses four shared timesteps, assigning the first two to Slow-LoRA and the next two to Fast-LoRA; both adapters have rank 32.Training uses batch size 64 and Adam with learning rate 2 × 10^-6; the single data sample is trained for 60 steps, or 60 effective epochs.Source: A MORE IMPLEMENTATION DETAILS · p. 13

Trajectory Curvature

• The curvature analysis uses Wan2.1-T2V-1.3B with 100 diverse prompts, 50 denoising steps, noise schedule shift 8, guidance scale 6, and one deterministic seed per prompt.The deterministic seed is used to eliminate randomness from different initial noise realizations.Source: B ANALYSIS DETAILS · p. 13
• 91.82% ± 0.15% of total curvature mass falls at t ≥0.9 across 100 trajectories, with a high/mid-noise curvature ratio of 32.4 ± 0.6.The 95% bootstrap confidence intervals are [91.53, 92.11]% and [31.2, 33.6], respectively; the authors report the uncertainty as SEM.Source: B ANALYSIS DETAILS · p. 13
• 87.9% of curvature mass remains at t ≥0.9 under a temporal-difference metric that removes static appearance and emphasizes motion structure.The high/mid-noise ratio under this metric is 18.6.Source: B ANALYSIS DETAILS · p. 13

One-Sample Selection Criteria

• The six representative samples contrast motion dynamics, visual complexity, and domain specificity while holding video settings, camera trajectories, subject scales, and prompt structures consistent.The pairs are high/low motion, complex textures/simple geometry, and open/specific domain.Source: B ANALYSIS DETAILS · p. 14
• Sample videos use Wan2.1-T2V-1.3B with consistent sampling settings and normalized prompt structure, subject scale, and camera trajectory descriptions.These controls are intended to make the selected dimensions capture differences in motion, visual complexity, and domain specificity rather than generation settings.Source: B ANALYSIS DETAILS · p. 14

Study platform

• The web-based evaluation platform randomizes model-output order and spatial placement to reduce potential presentation bias.Figure 9 shows the interface.Source: C USER STUDY · p. 14, Figure 9 · p. 14

Study design

• Each comparison uses 50 randomly sampled prompts, with each generated video pair independently evaluated by three human annotators.Prompts are fixed before evaluation.Source: C USER STUDY · p. 14

Rating procedure

• Annotators choose a preferred video or a tie after viewing two videos side-by-side in a continuous loop.They judge overall visual quality, temporal consistency, motion realism, and prompt adherence.Source: C USER STUDY · p. 14

Rating totals

• Each baseline comparison yields 150 individual ratings, for 300 ratings across both comparisons.The aggregated votes are reported in Table 2 in the main paper.Source: C USER STUDY · p. 14

Impact of κ

• Figure 10 is identified as showing the impact of κ, but the supplied caption does not state the direction or magnitude of that impact.Source: Figure 10 · p. 15

Additional qualitative results

• Five new prompts provide additional qualitative results covering all six models in Table 1, with five videos per model generated using the same random seed.Figure 11 compares approaches during ODE initialization, while Figure 12 presents comparisons during DMD distillation.Source: E ADDITIONAL QUALITATIVE RESULTS · p. 15

Table 4 prompts

• The Table 4 prompt describes a red sports car drifting on a mountain road, while other supplied prompts specify a nearly still woman and a crowded Tokyo night market.The scenes specify differing amounts of subject and environmental motion.Source: F PROMPTS · pp. 15–16

Figure 6 prompts

• The Figure 6 prompts include a biologist on a rainforest suspension bridge and a violinist on a Seoul rooftop at night.Source: F PROMPTS · p. 18

Figure 7 prompts

• The Figure 7 prompts describe a steam train crossing Glenfinnan Viaduct and a reindeer in a snowy forest at sunset.Source: F PROMPTS · p. 18

Figure 11 prompts

• Figure 11 prompts include a tea-valley scene and an origami dragon riding a boat through waves.Source: F PROMPTS · p. 18

Figure 12 prompts

• Figure 12 prompts range from a tea valley and snowy-forest reindeer to a coding llama and a woman picking wildflowers by a lake.Other supplied prompts specify a rotating chrome cube, a green valley, and a bioluminescent jellyfish.Source: F PROMPTS · pp. 16, 18, 19