PaperlayerView exact revision on arXiv

Source-linked AI summary

Fast3Dcache: Training-free 3D Geometry Synthesis Acceleration

Mengyu Yang, Yanming Yang, Chenyi Xu, Chenxi Song, Yufan Zuo, Tong Zhao, Ruibo Li, Chi Zhang

arXiv:2511.22533v2cs.CV

TL;DR

3D caching is challenging because errors in reused features can disrupt geometry, unlike the small texture inaccuracies often tolerated in 2D or video generation. Fast3Dcache uses voxel stabilization to schedule and select cached features; when combined with adapted TeaCache, it reports 3.41× throughput versus 2.84× for TeaCache alone, with improved geometric fidelity.

  • Problem

    Caching methods that reuse redundant computations accelerate 2D and video generation, but applying them to 3D can let small errors accumulate into structural and topological inconsistencies.

    Source: §1 · p. 1

  • Method

    Fast3Dcache uses voxel stabilization to set cache quotas with PCSC and selects stable tokens for reuse with SSC, based on velocity magnitude and acceleration.

    Source: §1 · pp. 1–2

  • Results

    3.41× throughput with Fast3Dcache combined with adapted TeaCache exceeds 2.84× for adapted TeaCache alone, and the combination also improves geometric fidelity.

    Source: §5.2 · p. 7

  • Takeaways & Limitations

    The comparison underscores that caching strategies developed for 2D image synthesis do not directly generalize to 3D geometry because they overlook volumetric stabilization patterns and topology-sensitive dynamics.

    Source: §5.2 · p. 6

    Fast3Dcache is optimized for TRELLIS's sparse voxel grid; applying its voxel-based metrics to continuous or implicit representations requires tailoring the stability criteria.

    Source: §13 · p. 17

Abstract

from arXiv · show

Diffusion models have achieved impressive generative quality across modalities like 2D images, videos, and 3D shapes, but their inference remains computationally expensive due to the iterative denoising process. While recent caching-based methods effectively reuse redundant computations to speed up 2D and video generation, directly applying these techniques to 3D diffusion models can severely disrupt geometric consistency. In 3D synthesis, even minor numerical errors in cached latent features accumulate, causing structural artifacts and topological inconsistencies. To overcome this limitation, we propose Fast3Dcache, a training-free geometry-aware caching framework that accelerates 3D diffusion inference while preserving geometric fidelity. Our method introduces a Predictive Caching Scheduler Constraint (PCSC) to dynamically determine cache quotas according to voxel stabilization patterns and a Spatiotemporal Stability Criterion (SSC) to select stable features for reuse based on velocity magnitude and acceleration criterion. Comprehensive experiments show that Fast3Dcache accelerates inference significantly, achieving up to a 27.12% speed-up and a 54.83% reduction in FLOPs, with minimal degradation in geometric quality as measured by Chamfer Distance (2.48%) and F-Score (1.95%).

1. Introduction

Fast3Dcache targets the geometric errors that can arise when 2D caching methods are applied to 3D diffusion, using voxel stabilization to guide computation reuse. Its scheduler sets cache quotas over timesteps, while a stability criterion selects tokens for reuse.

1. Introduction

  • Fast3Dcache exploits sparse-voxel stabilization, with PCSC adjusting cache ratios over timesteps and SSC selecting stable tokens using velocity magnitude and acceleration.The paper reports progressively stabilizing voxel occupancy and a roughly logarithmic decline in active updates.Source: §1 · pp. 1–2
  • Fast3Dcache is reported to achieve state-of-the-art acceleration-performance trade-offs on 3D generation tasks.The framework leverages voxel occupancy stabilization during denoising and combines PCSC with SSC.Source: §1 · p. 2

2. Related Work

  • 2D caching methods reuse redundant computations, but their direct transfer to 3D can overlook geometry-specific dynamics and cause topological errors.Prior feature-caching methods reuse features based on spatial or temporal similarity; 3D-adapted use can cause topological errors [5] [10] [11] [26] [29] [33] [34] [42] [59] [73] [76] [77].Source: §1 · p. 1, §2 · p. 2

3.1. Flow Matching (FM)

  • In Flow Matching, inference evolves a sample from noise by numerically solving an ODE, and the predicted velocity field at each step informs caching.The field is learned to approximate the ideal velocity along a path interpolating data and noise.Source: §3.1 · pp. 2–3

3.2. Sparse Structure Generation

  • TRELLIS separates structure generation from SLAT generation; Fast3Dcache accelerates the structure stage, where voxel changes diminish in a three-phase pattern.The structure is represented by active voxel coordinates, and its latent grid is evolved by a flow transformer conditioned on an image.Source: §3.2 · p. 3

4. Methodology

Fast3Dcache exploits predictable voxel and latent-feature stabilization to decide when and which tokens to cache, then integrates those choices into a three-stage inference workflow.

4.1. 3D Geometry Synthesis Observation

  • Voxel changes follow three phases: rapid early evolution, log-linear intermediate stabilization, and sharp final decline, motivating increasingly aggressive caching.The proposed schedule disables caching early, grows the budget during the intermediate phase, and applies it most aggressively during final refinement.Source: §4.1 · p. 3
  • Persistently small velocity magnitudes and low acceleration identify stable tokens, while large or changing updates indicate tokens that need recomputation.Velocity measures update intensity; acceleration measures how much the current update direction deviates from the previous one.Source: §4.1 · pp. 3–4

4.2. Fast3Dcache Core Components

  • PCSC predicts the changing cache budget from log-linear voxel decay, calibrating at an anchor step and converting predicted geometric change into a latent-token quota.The quota constrains how many tokens may be cached during Phase 2.Source: §4.2 · pp. 4–5
  • SSC ranks tokens by velocity and acceleration stability, recomputing the less stable active subset while reusing cached features for stable tokens.Attention is performed only on the active subset selected under the PCSC budget.Source: §4.2 · pp. 5–6

4.3. Fast3Dcache Integration

  • Fast3Dcache uses full sampling and anchor calibration first, PCSC-budgeted SSC caching with periodic refreshes second, then fixed-ratio CFG-free caching with periodic full corrections.The refresh interval limits error accumulation during dynamic caching; the final phase uses a separate correction cycle.Source: Figure 3 · p. 5, §4.3 · p. 6

5. Experiments

Experiments evaluate Fast3Dcache on TRELLIS and DSO, measuring inference efficiency and geometric fidelity. The results report better efficiency-quality trade-offs than the baselines, while ablations examine the scheduler, token selection, and correction step.

5.1. Implementation Details

  • Experiments use TRELLIS and DSO on one RTX 4090, and evaluate 71 Toys4K objects through 852 image prompts using throughput, FLOPs, CD, and F-Score.Meshes are normalized and aligned to ground truth with ICP before geometric metrics are computed.Source: §5.1 · p. 6

5.2. Results Analysis

  • Fast3Dcache delivers higher throughput and lower FLOPs while preserving CD and F-Score across TRELLIS and DSO; 3D-adapted RAS drops TRELLIS F-Score by 26.53%.Table 1 compares Fast3Dcache with TRELLIS, DSO, and RAS.Source: §5.2 · p. 6, Table 1 · p. 7
  • Combining Fast3Dcache with 3D-adapted TeaCache raises throughput speedup from 2.84× to 3.41× and improves CD and F-Score over TeaCache alone.The combined method is presented as complementary to TeaCache.Source: §5.2 · p. 7, Table 2 · p. 7
  • PCSC outperforms fixed-rate sampling by adapting cache budgets to prompt-specific geometric complexity; its slope sensitivity is tested over an order-of-magnitude range.Small slope fluctuations of about ±10% have negligible effect because cache quotas are discretized.Source: §5.2 · p. 7, Table 3 · p. 7
  • SSC combines velocity and acceleration because neither component alone is sufficient to assess geometric stability.The joint criterion yields better geometric fidelity than relying on a single metric.Source: §5.2 · pp. 7–8, Table 4 · p. 8
  • Figure 4 contrasts input images and generated meshes; RAS shows artifacts and surface noise, while Fast3Dcache retains structural fidelity comparable to TRELLIS.The comparison includes RAS at varying sampling ratios.Source: §5.2 · p. 7, Figure 4 · p. 8
  • Disabling the elimination correction step degrades geometry, with CD reaching 0.0724 and F-Score falling to 51.8157.The authors report periodic full-sampling updates as essential for correcting accumulated approximation errors.Source: §5.2 · p. 8

6. Conclusion

Fast3Dcache is a training-free accelerator tailored to the TRELLIS series and spatially explicit sparse-voxel representations. Its PCSC and SSC modules use voxel stabilization patterns to reduce computation while preserving geometric fidelity.

6. Conclusion

  • Fast3Dcache combines PCSC cache-budget allocation based on voxel decay trends with SSC selection of the active tokens requiring updates.The framework is designed for sparse-voxel geometry synthesis and aims to reduce FLOPs while preserving geometric fidelity.Source: §6 · p. 8

Supplementary Material

The supplementary material extends the experiments across accelerator combinations, tasks, datasets, and hyperparameter settings. It reports data-driven parameter selection and examines how refresh intervals affect the speed-quality trade-off.

7.1. More Experiment Details

  • Fast3Dcache is applied to DSO without modification, while RAS is re-implemented on 3D grids using its original design and default sampling ratios.The RAS ratios are 25% and 12.5%.Source: §7.1 · p. 13

7.2. Full Results of Complementarity with Modality-Agnostic Accelerators

  • Combining Fast3Dcache with EasyCache reaches 10.33× acceleration and F-Score 54.77; with TeaCache, the combined method reaches 3.41× speedup while surpassing the standalone baseline's geometric quality.The authors characterize the combinations as compounding efficiency gains without compromising generation quality.Source: §7.2 · p. 13, Table 5 · p. 13

7.3. More Results of Other Dataset and Task

  • On TRELLIS-text-xlarge and OmniObject3D, Fast3Dcache maintains geometry close to Vanilla on OmniObject3D, with F-Score 73.9062 versus 74.7461.Across tasks, the reported throughputs are 0.4781 and 0.6775 Iter/s; on OmniObject3D, CLIP is 19.3325.Source: §7.3 · p. 13, Table 6 · p. 13, Table 7 · p. 13

7.4. Data-driven Parameter Selection

  • Table 10 reports that f_corr = 0 does not improve visual quality despite better CD and F-Score metrics, and uses more FLOPs than f_corr = 3.The authors select f_corr = 3 after considering the balance between speed and quality.Source: Table 10 · p. 14, §7.4 · p. 14
  • Data-driven slope selection uses median RANSAC fits from 100 sampled instances, and the resulting parameters yield comparable metrics and superior performance on two datasets.Figure 5 shows the estimated slope distributions and log-decay trends for Toys4K and OmniObject3D.Source: §7.4 · pp. 13–14, Figure 5 · p. 13, Table 8 · p. 14

7.5. More Ablation Study

  • The reported default settings include μ = −0.07, ω = 0.7, τ = 3, ξ = 0.7, f_corr = 3, ρ_a = 0.2, and ρ_CFG-OFF = 0.75.The authors state that hyperparameters can be controlled flexibly to suit users' speed-quality requirements.Source: §7.5 · p. 14, §7.4 · p. 14

7.6. More Results of Hyperparameter Sensitivity

  • Small variations in μ leave performance robust, while a larger refresh interval τ consistently yields higher acceleration and serves as a speed-quality control.Table 11 analyzes μ and τ.Source: §7.6 · p. 14, Table 11 · p. 14

8. More Visualizations

The additional visualizations show voxel dynamics across prompts, velocity and acceleration fields, and generated shapes under different settings. The reported patterns support the use of log-linear voxel decay for cache-budget prediction.

8.1. More Visualizations of Voxel Dynamics

  • Across diverse prompts, Phase 2 shows stable decay in dynamic voxels, supporting a log-linear approximation for predicting the caching budget.These visualizations are presented as further validation of PCSC's universality.Source: §8.1 · p. 14

8.2. More Visualizations of Velocity Field

  • Additional velocity and acceleration field maps show similar patterns to the main-text examples, which the authors use to select active voxels during inference.Source: §8.2 · p. 14

8.3. More Visualizations of Generation

  • Across examples including humans, buildings, objects, animals, and creative objects, visualizations show geometry quality at different τ settings and continued quality in combined methods.Source: §8.3 · p. 14

9. Impact of Sampling Parameters on Voxel Dynamics

Voxel stabilization depends on the CFG interval and shifting factor η: ending CFG causes a sharp drop in active voxels, while continuous CFG removes it. Fast3Dcache’s phase division uses TRELLIS’s default schedule but can adapt to user-defined schedules.

9. Impact of Sampling Parameters on Voxel Dynamics

  • Phase 1 forms the outline, Phase 2’s dynamic-voxel count trends downward despite fluctuations, and Phase 3 is the CFG-off phase.The experiments confirm that a log-linear approximation is acceptable for Phase 2’s downward trend.Source: Figure 6 · p. 15
  • For η=2, cutoff t=0.5 maps to t_l=1/3, moving the transition to step 17 of 25; t_l denotes the original uniform-schedule timestep.Source: §9 · pp. 14–15
  • Under the default CFG interval [0.5, 1], η controls when CFG ends and the transition triggers a sharp drop in active voxels; continuous CFG over [0, 1] removes that drop.The timing of the drop also depends on the shifted sampling schedule.Source: §9 · pp. 14–15, Figure 7 · p. 15
  • Fast3Dcache’s phase division is calibrated to TRELLIS’s default η=3 and CFG interval [0.5, 1], but the framework can adapt to user-defined schedules.η=3 further delays the refinement stage.Source: §9 · p. 15

10. Impact of Computational Cost on Geometry Quality

The reported results indicate that Fast3Dcache can accelerate inference while maintaining or even enhancing geometric fidelity, and that greater computational volume does not always yield better geometry.

10. Impact of Computational Cost on Geometry Quality

  • Lower sampling ratios can yield better geometric metrics, including lower CD and higher F-Score, in some regimes.Source: §10 · p. 15
  • A sampling ratio of 0.3 yields lower CD than 0.4, illustrating that higher computational volume does not consistently produce better quality.Source: Figure 8 · p. 16
  • Fast3Dcache achieves significant acceleration while maintaining or even enhancing geometric fidelity.Source: §10 · p. 16

11. Algorithm of Fast3Dcache

The paper points readers to Algorithm 1 for the Fast3Dcache algorithm details.

11. Algorithm of Fast3Dcache

  • Algorithm 1 presents the details of the Fast3Dcache algorithm.Source: §11 · p. 16

12. FLOPs Calculation

The FLOPs calculation focuses on flow transformer blocks, while the inference algorithm schedules caching across denoising phases and selects tokens using spatiotemporal stability.

12. FLOPs Calculation

  • FLOPs calculations count operations inside flow transformer blocks because other modules have significantly lower computational workloads.Source: §12 · p. 16
  • SSC scores tokens using velocity and acceleration, caches the quota’s lowest-score tokens, and disables caching after τ consecutive cache steps to eliminate error accumulation.Source: §12 · p. 17
  • The algorithm precomputes a PCSC cache budget, uses full sampling in Phase 1 and dynamic caching in Phase 2, then applies CFG-free refinement with periodic full corrections and fixed-ratio caching.Source: §12 · p. 17
  • The pipeline updates active tokens while reusing others and returns the final denoised feature grid S0.Source: §12 · p. 17

13. Limitation and Future Work

Fast3Dcache is currently optimized for TRELLIS’s sparse voxel grids; extending its stability metrics to continuous or implicit representations requires tailoring. Texture-specific constraints for SLAT are left for future work.

13. Limitation and Future Work

  • The current implementation is optimized for the sparse voxel grid representation used by TRELLIS.The authors identify continuous and implicit representations, such as Signed Distance Fields, as requiring tailored stability criteria.Source: §13 · p. 17
  • Applying the voxel-based metrics to continuous or implicit representations requires tailoring their stability criteria.The authors plan to extend the caching paradigm to a broader range of 3D representations.Source: §13 · p. 17
  • The second stage, SLAT, can use established 2D acceleration methods, while texture-specific constraints remain future work.Source: §13 · p. 17
Loading 2511.22533v2…