CtrlCache: Accelerating Interactive Video World Models with Control-Aware Caching

1 INTRODUCTION

• Existing caching policies primarily use fixed schedules or internal feature changes and do not explicitly account for action transitions in interactive generation.Chunk-wise autoregressive generation with few-step diffusion still requires multiple costly DiT forward passes per chunk.Source: §1 · p. 2
• CtrlCache assigns chunks to initial, transition, turning, or steady states and uses full computation for initial and transition chunks while allowing residual reuse for turning and steady chunks.The policy applies across chunk boundaries and within chunks; the history prior emphasizes low-frequency structure, attenuates high-frequency detail, and is enabled only for steady chunks.Source: §1 · p. 2
• 1.21× to 1.41× DiT-backbone speedups accompany the highest WBench Overall scores among evaluated methods across Matrix-Game 2.0 and LingBot-World v1/v2.CtrlCache also achieves the highest PSNR and SSIM and lowest LPIPS among accelerated methods on every backbone.Source: §1 · p. 2
• Clean-latent similarity drops around action changes, while low-frequency components persist across chunks more than high-frequency components during steady interaction.This supports more conservative reuse during transitions and carrying coarse scene layout more reliably than fine appearance details.Source: §1 · p. 2
• The framework combines control-aware scheduling and refresh with a frequency-mixed history prior to adapt computation reuse and history transfer without retraining.The prior uses the preceding chunk’s clean latent frame without an additional computational cost.Source: §1 · pp. 2–3

2 RELATED WORK

• Training-free caching includes fixed schedules and adaptive policies based on feature variation, token redundancy, or motion, but related methods generally ignore online action changes as an explicit refresh signal.The related work also notes that each newly generated chunk still requires multiple DiT forward passes despite few-step denoising.Source: §2 · p. 3
• Figure 3 depicts chunk-wise denoising, control-state-based residual reuse or full computation, and steady-interaction history guidance as CtrlCache’s overview.The diagram retains full computation at the first and final denoising steps.Source: Figure 3 · p. 4
• FlowCache makes chunk-wise reuse decisions without interpreting online transitions, while Light Interaction ties reuse to similar historical views and is less suited to interactions with few or no revisits.CtrlCache instead derives reuse and refresh decisions from controls, using full computation at transitions.Source: §2 · p. 4

3 CTRLCACHE

• CtrlCache combines control-based decisions about residual reuse with a history prior that transfers information from the preceding chunk.The framework uses incoming controls to decide when to reuse cached transformer residuals and when to restore full computation.Source: §3 · p. 4

3.1 PRELIMINARY

• Residual reuse approximates a denoising step by applying the most recent fully computed residual from the same chunk instead of recomputing transformer blocks.The current input projection, timestep embedding, and output head are still computed; the residual cache changes only after a full computation.Source: §3.1 · pp. 4–5

3.2 ACTION-AWARE SCHEDULING AND REFRESH POLICY

• The policy assigns chunks to initial, transition, turning, or steady states using control changes across chunk boundaries and within each chunk.The preceding chunk’s last control is included to detect boundary changes, and transition takes priority over turning.Source: §3.2 · p. 5

3.3 FREQUENCY-MIXED HISTORY PRIOR GUIDANCE

• The frequency-mixed prior uses the preceding chunk’s last clean latent frame, attenuating high-frequency content and applying temporal weights to later prior frames.The guidance blends this prior into the current chunk’s clean estimate at the first denoising step.Source: §3.3 · p. 6

3.2 ACTION-AWARE SCHEDULING AND REFRESH POLICY

• At one selected interior step, initial and transition chunks use full computation, while turning and steady chunks reuse the cached residual; the first and final steps are always fully computed.History prior guidance is enabled only for steady chunks and adds no DiT forward computation.Source: §3.2 · pp. 5–6, §3.3 · p. 6, §3.4 · p. 7

4.1 EXPERIMENTAL SETUP

• The evaluation compares three backbones on WBench navigation against Original, TeaCache, and EasyCache under shared prompts, controls, initial frames, and decoding settings.Source: §4.1 · p. 7

4.2 MAIN RESULTS

• CtrlCache achieves the best WBench Overall on all three backbones; on Matrix-Game 2.0, Overall rises from 0.6543 (Original) to 0.6786, above EasyCache (0.6625).On LingBot-World v1/v2, it also achieves best Overall, PSNR, SSIM, and LPIPS, with respective speedups of 1.26× and 1.21×.Source: §4.2 · pp. 8–9
• On LingBot-World v1/v2, CtrlCache achieves the best Overall, PSNR, SSIM, and LPIPS, with respective speedups of 1.26× and 1.21×.Source: §4.2 · p. 9
• With identical initial conditions and controls, CtrlCache better preserves local details and scene structure than TeaCache and EasyCache in qualitative comparisons.The figure’s blue dashed boxes mark layout, object-placement, and motion references; red and green boxes mark distortions and better-preserved details, respectively.Source: Figure 4 · p. 8, §4.2 · p. 9

4.3 ABLATION STUDY

• Naive residual reuse slightly improves Overall on both tested backbones but can reduce physics compliance; adding action-aware scheduling improves Overall and selected setting or physics dimensions.The policy restores full computation at control transitions while turning and steady states use residual reuse.Source: §4.3 · p. 9
• Applying frequency-mixed history guidance only to steady chunks yields the highest Overall among the tested variants; extending it to turning or transition chunks reduces performance.The ablation fixes residual reuse and the action-aware policy while varying the states receiving history guidance.Source: §4.3 · p. 9, Table 2 · p. 9

• CtrlCache improves WBench Overall over original inference on Matrix-Game 2.0 and LingBot-World v1/v2 while achieving comparable speedups.Source: §5 · p. 10
• Its action-aware policy controls residual reuse and restores full computation at control transitions, while frequency-mixed history guidance transfers persistent coarse structure during steady interaction.Source: §5 · p. 10
• The current method relies on backbone-specific action indicators and a fixed reuse step that varies with each backbone’s denoising schedule.The authors identify more general control representations and adaptive step selection as future work.Source: §5 · p. 10

• The appendix provides inference pseudocode that complements the main-paper method description.Source: APPENDIX · p. 13
• It includes implementation details and the evaluation protocol.Source: APPENDIX · p. 13
• It also provides hyperparameter analyses and additional qualitative examples.Source: APPENDIX · p. 13

• The procedure takes a pretrained world model, sampler, initial conditioning, online controls, chunk and denoising settings, a reuse step, and prior parameters.Source: A CTRLCACHE ALGORITHM · p. 13
• For each chunk, it reads current controls, derives control-state and reuse indicators, clears the within-chunk residual cache, and initializes the noisy latent.Source: A CTRLCACHE ALGORITHM · p. 13
• For steady chunks, it constructs a history prior from the preceding clean latent; after sampling, it retains the clean latent as the next chunk’s history reference and updates autoregressive context.Source: A CTRLCACHE ALGORITHM · p. 13
• On the selected reuse step, the algorithm uses the latest cached residual when reuse is permitted; otherwise, it fully computes the transformer blocks.Source: A CTRLCACHE ALGORITHM · p. 13

B IMPLEMENTATION DETAILS

• Original inference uses denoising schedules [1000, 908, 713] for Matrix-Game 2.0 and [999, 978, 947, 825] for LingBot-World v1.The supplied schedule passage also begins the LingBot-World v2 schedule but does not provide its values.Source: B IMPLEMENTATION DETAILS · p. 13
• The default residual-reuse timesteps are 908, 947, and 967 for Matrix-Game 2.0, LingBot-World v1, and LingBot-World v2, respectively.The history prior uses α = 0.5 and λ = 0.5; temporal decay is [1, 0.5, 0.25] for Matrix-Game 2.0 and v1, and [1, 0.5, 0.25, 0.125] for v2.Source: B IMPLEMENTATION DETAILS · p. 14
• Matrix-Game 2.0 classifies the first chunk as initial, later action changes as transition, mouse movement without a transition as turning, and otherwise as steady.The classifier checks changes across chunk boundaries and within chunks, using the unramped mouse command; turning requires a mouse-coordinate magnitude above 10^-6.Source: B IMPLEMENTATION DETAILS · p. 14
• LingBot-World v1/v2 encode controls as motion, yaw, and pitch; absent a transition, nonzero yaw or pitch marks a chunk as turning, otherwise it is steady.The passage maps W/S/A/D to local-frame motion and directional view controls to yaw or pitch.Source: B IMPLEMENTATION DETAILS · p. 14

C EVALUATION PROTOCOL

• 40 WBench navigation cases are sampled with generation seed 42, and every method uses the same cases and seed for each backbone.Original and accelerated runs also share prompts, controls, initial frames, and decoding settings, enabling paired fidelity comparisons.Source: C EVALUATION PROTOCOL · p. 14
• WBench Overall is the unweighted arithmetic mean of Quality, Consistency, Interaction, Setting, and Physics scores.Each component is averaged over the cases for which that measure is defined.Source: C EVALUATION PROTOCOL · p. 14
• PSNR, SSIM, and LPIPS are computed frame by frame against each paired Original video, averaged within video, then averaged equally across cases.Source: C EVALUATION PROTOCOL · p. 14
• Efficiency reports mean DiT denoising latency and speedup relative to Original, excluding KV-cache updates, model loading, condition preparation, VAE decoding, and video encoding.Timing uses CUDA synchronization and covers only DiT computation during denoising.Source: C EVALUATION PROTOCOL · p. 14

D HYPERPARAMETER STUDY

• History-prior hyperparameter studies cover frequency composition, blend, and temporal decay on Matrix-Game 2.0 and LingBot-World v1.The supplied figure text identifies blend guidance strength λ, but provides no result values.Source: D HYPERPARAMETER STUDY · pp. 14–15
• The supplied study description names the tested settings but does not report their comparative outcomes.Source: D HYPERPARAMETER STUDY · pp. 14–15
• The study varies the denoising step used for residual reuse on Matrix-Game 2.0 and LingBot-World v1.The supplied figure text identifies this as the effect of denoising step s; it provides no outcome values.Source: D HYPERPARAMETER STUDY · pp. 14–15

E ADDITIONAL QUALITATIVE EXAMPLES

• The additional comparisons cover Matrix-Game 2.0 and LingBot-World v1/v2, with six successive frames for each comparison.The shown controls use W/S/A/D for translation and directional commands for view changes.Source: E ADDITIONAL QUALITATIVE EXAMPLES · p. 15, Figure E.1 · p. 18
• The examples include a blocky game world, an anime-style subway, a rice field, a rainy highway, a warehouse, a library, a subway platform, and a war-torn street.The listed examples use varied control sequences, including steady movement and combinations of turns or direction changes.Source: E ADDITIONAL QUALITATIVE EXAMPLES · pp. 16–17
• Figure E.1 compares Original, TeaCache, EasyCache, and CtrlCache in rows using the same initial condition and WBench control sequence.Prompts and controls are shown with the frames.Source: Figure E.1 · p. 18