Steady-Forcing: Balancing Spatial Persistence and Motion Continuity in Long-Horizon Nature Video Diffusion

Challenges

• Autoregressive rollouts accumulate prediction errors and temporal-position extrapolation failures, causing background drift, identity collapse, and reduced long-horizon stability.Attention-sink methods can mitigate spatial drift while suppressing motion, producing stagnation in natural flows.Source: §1 · p. 2

Motivation

• Fixed-camera generation separates geometric background stability from continuous motion in dynamic regions, making the setting a controlled testbed for the stability–motion trade-off.The setting also supports extended coherent streams for ambient media, procedural game environments, and dynamic background synthesis.Source: §1 · p. 2

Approach

• Steady-Forcing targets the interaction between spatial drift and motion decay, which prior works generally address in isolation.The framework specializes autoregressive video diffusion for static-scene nature streams through a task-specific dual-memory policy.Source: §1 · pp. 2–3
• The Unified Dual-Sink Mechanism separates persistent scene identity in V-Sink from compressed kinetic history in EMA-Sink during multi-minute rollouts.This bounded-memory design assigns distinct memory roles to scene identity and dynamic motion.Source: §1 · p. 3
• Periodic KV Flush resets the autoregressive cache at regular intervals, suppressing accumulated errors before they stabilize into repeated texture artifacts.The strategy is presented as cache purification for long rollouts.Source: §1 · p. 3

Training

• A ground-truth-video-free distillation pipeline combines a 21,000-prompt synthetic corpus, motion-rewarded prior initialization, and a Wan2.1-14B teacher.The pipeline specializes a general-purpose autoregressive model for fixed-camera nature streams without ground-truth video supervision.Source: §1 · p. 3

Results

• Across seven forcing-based baselines, Steady-Forcing improves long-horizon background consistency and visual quality while maintaining plausible flow and higher perceived motion continuity.The result is supported by experiments and a blind user study.Source: §1 · p. 3

Video Diffusion and DiTs

• Diffusion Transformers provide strong quality baselines, but bidirectional models require future temporal context and therefore preclude low-latency streaming inference.Wan2.2 increases capacity through a Mixture-of-Experts formulation without increasing inference cost.Source: §2 · pp. 3–4

Autoregressive Long Video Generation

• Causal distillation and forcing methods enable streaming generation by addressing train–test gaps, error accumulation, or motion amplitude.Causal Forcing is reported to surpass Self-Forcing by 19.3% in Dynamic Degree, while LongLive extends autoregressive generation.Source: §2 · p. 4

Unbounded Horizon Extension

• Infinity-RoPE and related operators extend generation length through moving temporal reference frames, KV Flush, and RoPE Cut, but do not address fixed-camera natural-flow stability–motion trade-offs.These methods focus on horizon extension rather than the joint interaction of drift and motion decay.Source: §2 · p. 4

Motion-Enhanced Distillation

• Reward Forcing combines EMA-Sink and Re-DMD to improve dynamic amplitude, and Steady-Forcing specializes their interaction for fixed-camera nature streams.Reward Forcing is reported at 23.1 FPS with an 88.38% improvement in dynamic amplitude.Source: §2 · p. 4

Video Generation Evaluation

• VBench reports Background Consistency and Dynamic Degree, but these dimensions do not expose long-horizon flow decay or stability achieved by suppressing motion.Dynamic Degree measures aggregate motion amplitude rather than persistence or late-stage decay.Source: §2 · p. 4

Base Model

• Wan2.1-T2V is a flow-matching DiT operating in a compressed latent space with a causal 3D VAE.The VAE compresses video by 4× temporally and 8× spatially, reducing transformer-attention token count.Source: §3 · p. 4

3D-RoPE

• Wan2.1 uses 3D-RoPE for temporal, height, and width coordinates, but temporal indices beyond the 1024-index training horizon become out-of-distribution.This limits attention reliability during extended autoregressive rollouts.Source: §3 · p. 5

Self-Forcing DMD Distillation

• Self-Forcing trains each frame on the model’s previously generated outputs, exposing it to compounding prediction errors encountered during inference.This bridges the train–test distribution gap and mitigates exposure bias.Source: §3 · p. 5

Unified Dual-Sink Mechanism

• Steady-Forcing decouples drift and stagnation through separate memory pathways for long-range scene identity, compressed kinetic history, and local temporal continuity.The resulting context has constant memory footprint O(w + s) as generation length grows.Source: §4 · p. 5, §4.1 · pp. 5, 7
• The V-Sink permanently retains Frame 0 keys and values as an immutable spatial reference against background displacement during long self-rollouts.It preserves the scene’s initial layout, color distribution, and structural identity.Source: §4.1 · pp. 5–6, Figure 3 · p. 6
• The EMA-Sink fuses evicted key-value pairs into an exponentially weighted motion summary, retaining long-range kinetic context without growing memory size.The EMA-Sink is refreshed at every generation step, while the V-Sink remains fixed; α is set to 0.99 for training and inference.Source: §4.1 · p. 6

Block-Relativistic RoPE

• Block-Relativistic RoPE assigns newly generated blocks relative temporal positions while keeping sink positions fixed, extending temporal attention beyond the 1024-index training horizon.The V-Sink retains temporal index 0, and the EMA-Sink receives a fixed intermediate index.Source: §4.1 · pp. 6–7

Periodic KV Flush

• Periodic KV Flush resets the cache every Npurify = 21 blocks to suppress accumulated errors before they stabilize into repeated texture artifacts.The reset retains the V-Sink and m = 5 recent latent frames, preserving scene identity and local temporal continuity; the accumulated EMA-Sink is discarded.Source: §4.2 · pp. 7–8, Figure 4 · p. 7

Training

• Training combines motion-biased initialization from Reward-Forcing with Self-Forcing DMD distillation from a frozen Wan2.1-T2V-14B teacher under the same memory configuration used at inference.The student generates chunk-by-chunk from its own previous outputs, while motion-rewarded initialization supplies higher motion amplitude than standard DMD initialization.Source: §4.3 · p. 8

Experimental Setup

• Evaluation uses seven forcing-based baselines, six VBench dimensions, and four horizon tiers spanning 5s to 240s.The evaluation excludes Subject Consistency because prompts contain no foreground subject.Source: §5.2 · p. 9, §5.1 · p. 9, §5.3 · p. 10
• The fixed-camera protocol evaluates public baseline checkpoints as task-adapted variants rather than native release scores.Daggered baselines use the same fixed-camera steady-motion inference wrapper as Steady-Forcing.Source: §5.2 · p. 9

Qualitative and Mechanistic Results

• The V-Sink reduces background displacement, EMA-Sink preserves downstream directional flow, and Periodic KV Flush suppresses hardened texture artifacts.Each component addresses a distinct long-rollout failure mode described in the experiments.Source: §5.2 · p. 9

Metric Interpretation

• VBench Dynamic Degree rewards drift-induced and genuine fluid optical flow equally, so the paper interprets it alongside Background Consistency and human preferences.The method is optimized for the fixed-camera nature-flow objective rather than the generic VBench average.Source: §5.3 · pp. 10–11

Quantitative Results

• Steady-Forcing achieves the highest Background Consistency and Imaging Quality at every reported horizon: 98.06/69.87 at 5s, 95.60/71.59 at 60s, 95.07/66.21 at 120s, and 92.57/71.27 at 240s.Background Consistency generally degrades more slowly for Steady-Forcing than for the baselines.Source: §5.3 · p. 10
• Steady-Forcing is best or second-best on Motion Smoothness, Temporal Flickering, and Aesthetic Quality across the reported horizons.These metrics complement the strongest Background Consistency and Imaging Quality results.Source: §5.3 · p. 10

6.1 User Study

• Steady-Forcing receives the highest preference rate across all five user-study criteria.The study compares anonymized videos from Steady-Forcing and three representative baselines.Source: §6.1 · p. 11
• Static-View Stability preference is 0.746 versus 0.101 for the next-best method, while Motion Continuity is 0.710 versus 0.116.The simultaneous lead covers the two criteria most directly tied to the stability–motion trade-off.Source: §6.1 · p. 11

6.2 Ablation Study

• The V-Sink alone preserves coarse layout but allows stagnation, whereas EMA-Sink alone retains motion history but allows background drift.Combining both sinks improves the stability–motion balance, and Periodic KV Flush further reduces accumulated cache artifacts.Source: §6.2 · p. 12, Figure 6 · p. 12
• At 240s and other extreme horizons, residual limitations include fine-texture inconsistency, mild color drift, and partial motion stagnation in large flow regions.The trade-off is mitigated but not fully eliminated.Source: §6.2 · p. 12, Figure 7 · p. 13

Efficiency Analysis

• Steady-Forcing runs at approximately 17 FPS on a single A100 GPU with KV memory O(w+s) relative to rollout length.The fixed-size V-Sink, EMA-Sink, and local window keep memory bounded as rollouts lengthen.Source: §6.2 · p. 13

Inherited Model Constraints

• Underrepresented fluid interactions in the Wan2.1 training distribution can produce implausible dynamics during long rollouts.This limitation is inherited from the backbone’s physics priors.Source: §7 · p. 13

Mid-Sequence Forgetting

• The fixed-size memory cannot recall content generated between the V-Sink and sliding-window contexts, causing mid-sequence details to be progressively lost at each KV Flush.The design provides global and local context but lacks an intermediate recall mechanism.Source: §7 · p. 13

Residual Stagnation

• Motion stagnation remains partially unresolved in large-area flow regions because a single compressed EMA-Sink state cannot represent broad-scale coherent dynamics.The flush period is also not adaptive to local motion complexity.Source: §7 · p. 13

Future Work

• Future directions include hierarchical memory, stronger teachers via VLM-integrated Re-DMD, and adaptive flush scheduling.These directions are proposed to address the identified memory and long-rollout limitations.Source: §7 · p. 13

• Steady-Forcing combines V-Sink, EMA-Sink, Block-Relativistic RoPE, Periodic KV Flush, motion-rewarded initialization, and Self-Forcing distillation for fixed-camera natural-flow generation.The framework uses a Wan2.1-14B teacher and domain-specific negative prompting without ground-truth video supervision.Source: §8 · p. 13
• Evaluations across seven baselines improve long-horizon background consistency, imaging quality, and perceived motion continuity.Source: §8 · p. 14
• Aggregate VBench scores under-penalize fixed-camera artifacts and reward drift-induced optical flow as Dynamic Degree, motivating task-specific static-camera nature-stream benchmarks.Source: §8 · p. 14

• The supplement provides figures, tables, equations, and public code, data, models, and generated videos for Steady-Forcing.Source: Supplementary Material:… · p. 19
• Figure S1 shows stable background identity and category-specific motion across six fixed-camera nature-flow categories at 0, 20, 40, and 60 seconds.The EMA-Sink preserves river directionality, smoke’s upward drift, ocean turbulence, and particle coherence in rain and snow.Source: §A.1 · p. 19
• Figure S2 extends the main baseline comparison to additional scene types using the same task-adapted protocol.Source: §A.2 · p. 19

• At horizons of at least 120 seconds, Steady-Forcing’s primary failure cases include motion flattening, texture hardening, temporal repetition, and local geometric drift.Source: §A.3 · pp. 19–20
• Open-ocean wave amplitude progressively decays because the EMA-Sink’s single compressed state cannot represent broad-scale coherent dynamics.Source: §A.3 · p. 19
• Rare cache contamination surviving a flush can reinforce repeated patterns in textured static regions over time.Source: §A.3 · p. 19
• Low-optical-flow scenes can enter short-period loops, while dense rain or smoke can cause minor background displacement at scene boundaries.Source: §A.3 · pp. 19–20

Dynamic Degree and quantitative evaluation

• VBench aggregate averages do not directly measure the qualities most critical for fixed-camera static-scene generation.Dynamic Degree measures aggregate motion amplitude, while Background Consistency does not penalize stability achieved by suppressing motion.Source: §B.3 · p. 23
• Dynamic Degree combines genuine foreground fluid motion with background drift caused by exposure bias, so suppressing drift necessarily lowers total frame-level optical flow.Source: §B.3 · p. 23
• Causal-Forcing reports the highest Dynamic Degree and VBench average but the lowest Background Consistency, including scores of 95.85 at 5s, 87.73 at 60s, 88.09 at 120s, and 86.87 at 240s.Source: §B.3 · p. 23
• Drift-induced optical flow can inflate Dynamic Degree without preserving static background identity, so the metric cannot reliably distinguish desired fluid motion from camera drift.Source: §B.3 · p. 24
• Across horizons, Steady-Forcing ranks first in Background Consistency and Imaging Quality despite ranking last or second-last in Dynamic Degree.Source: §B.3 · p. 24, Table S5 · p. 25

Statistical significance

• In a blind four-way study, 23 participants evaluated six prompts using anonymized, randomized videos across overall quality, stability, motion continuity, temporal consistency, and artifact-free quality.Source: §C.1 · pp. 24–25
• Steady-Forcing led all five preference criteria significantly, with p < 0.001 under a one-sided binomial test against the 0.25 chance level.Source: §C.3 · p. 27

Teacher model scale

• Past 100 seconds, the Wan2.1-1.3B-teacher variant suffers structural collapse and loses fluid dynamics, while the Wan2.1-14B-teacher model preserves rich texture boundaries.Both variants are evaluated on an identical river-scene prompt out to 200 seconds.Source: §D.1 · p. 27

Memory trade-off

• The bounded-memory design retains the initial frame and recent context but cannot preserve fine-grained scene details generated in the middle of a long sequence.Intermediate frames are compressed into EMA-Sink and eventually discarded during KV Flush, yielding O(w + s) constant memory.Source: §E.1 · p. 27

Teacher model scale

• At a 200-second horizon, distillation from Wan2.1-14B improves preservation of downstream flow details and perspective geometry over the smaller-teacher variant.The comparison is shown in the teacher-training ablation across the full rollout horizon.Source: Figure S5 · p. 29