acceptodds
Withdrawn from ICLR 2027

R3D:Residual Reliability Routing for Training-Free Image-to-3D Acceleration

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Recent diffusion- and flow-based image-to-3D generators synthesize high-quality 3D assets, but remain computationally expensive due to repeated Transformer computation throughout sampling. Existing training-free acceleration methods reduce this cost by exploiting temporal, stage-wise, or token-wise redundancy, yet they do not characterize whether the residual updates from Transformer blocks can be reliably predicted for faster inference. In this paper, we first characterize the reliability of predicting these residual updates within the texture stage. We find that prediction reliability varies substantially across sampling time and network depth, while the relative ordering of reliable and unreliable locations remains consistent across inputs. Moreover, fine-grained residual units within a Transformer block can remain predictable even when whole-block residual prediction is unreliable. Motivated by these findings, we propose Residual Reliability Routing (R3D), a training-free framework that uses the reliability of residual prediction to construct a reusable routing policy. R3D first constructs reusable reliability profiles from a small calibration set through Residual Reliability Profiling (RRP). Then, we propose Granularity-Aware Routing (GAR), which selects the residual unit within a Transformer block that reduces the most computation. Finally, Exact-Anchor Refresh (EAR) recomputes selected residual units in the Transformer block to refresh prediction histories. The resulting routing policy is fixed before inference and generalized to unseen inputs without modifying the pretrained generator. Experimental results show that R3D consistently reduces texture stage inference cost while largely preserving 3D generation quality.

Primary area:
generative models
Keywords:
3D Asset Generation, Training-Free Acceleration, Temporal Redundancy

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Y4JnOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates feature-caching-based acceleration of the denoising process in diffusion-based 3D generative models. Exploiting the consistency of residual prediction reliability across inputs, it uses offline analysis to identify computations that can be replaced by prediction. Periodic exact refreshes provide historical residuals for extrapolation, reducing computation and achieving substantial target-stage speedups while largely preserving generation fidelity.

Critical strengths

The paper establishes a reasonably coherent link between its analysis, method, and experimental validation. For example, the depth-permutation ablation preserves the prediction budget while changing prediction locations, providing targeted evidence for the usefulness of reliability-aware placement.

Critical weaknesses

The incremental contribution over closely related methods (e.g., ERTACache) remains unclear. Further clarification or experimental evidence is needed to assess the novelty; please see the questions below for details.

Questions for authors
  1. Could the authors clarify the standalone contribution of RRP relative to prior offline error-profiling methods, such as SmoothCache and ERTACache, independently of GAR and EAR?
  2. Could the authors explain Table 3 in more detail and clarify the advantage of GAR over simpler submodule-level caching or forecasting methods, such as FORA, SmoothCache, and TaylorSeer? In particular, what do “Cells” and “Units” represent, how do the Submodule and Mixed settings differ in execution, and why does Mixed achieve a higher speedup despite having the same number of prediction cells?
  3. The experiments compare R3D with TeaCache and SmoothCache, but do not include a forecasting-based baseline such as TaylorSeer. Could the authors provide such a comparison under comparable speed–quality settings, or explain any adaptation barriers? This would help distinguish the benefits of residual forecasting itself from those of the proposed reliability routing.
  4. Could the authors clarify how much of the speedup in Table 1 is attributable to R3D itself versus the additional exact caching optimizations described in Appendix C.4? Were equivalent optimizations enabled for the baselines? A breakdown on one representative backbone would help connect the controlled ablations to the main deployment results.
Reviewer_PXWNOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes residual reliability routing to accelerate image-to-3D models with various architectures. The performaces are encouraging and the training-free nature makes the method easy to be adapted to 3D object foundation models.

Critical strengths
  • The paper incorporates structured predictability of Transformer residuals to enhance the efficiency of diffusion sampling. The main comparison and ablation studies provides sufficient results.
Critical weaknesses
  • The overview figure (Fig. 4) may be confusing as different mentioned methods adopt different inference strategy in image-to-3D pipeline. Aslo, there might be more than one Transformer in mentioned methods (for instance, LumiTex).
  • The implementation details, especially how the method operates on different architecture methods are not enough.
  • I think the most critical concern on this paper is the lack of 3D native designs like the relations of 3D tokens, the 3D consistency, the geometry and texture nature. Its foundamental designs seem to orthorgonal to 3D foundation models but can also apply to image and video models. In this case, why are 3D models selected as the base models?
Questions for authors
  • How is R3D adapted to different backbone architectures, particularly TRELLIS and LumiTex, which employ different latent representations and generation pipelines?
Reviewer_bi7uOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes R3D, a training-free method for accelerating the texture stage of image-to-3D generators built on diffusion. It has three components: Residual Reliability Profiling (RRP) estimates a Q90 normalized-error profile from 20 calibration objects; Granularity-Aware Routing (GAR) selects, at each location, the residual unit that saves the most compute among those whose error is below a threshold ϵ\epsilon; Exact-Anchor Refresh (EAR) refreshes the extrapolation anchors with scheduled exact evaluations and never feeds predicted residuals back into the history. Experiments demonstrate promising results.

Critical strengths
  1. Useful change of perspective. Most caching methods ask whether computation is redundant. This paper asks whether the residual can be reliably predicted, which is a different question.
  2. The article's experiments are thorough and effectively validate the method.
Critical weaknesses
  1. Limited novelty relative to prior work. The method described in the article appears to be a combination of previous methods(SmoothCache, TaylorSeer), but there is no discussion of its similarities and differences compared to those methods.
  2. Missing key baselines, especially 3D-specific accelerators. Training-free accelerators like Fast3Dcache, Fast-SAM3D and TaylorSeer are necessary.
  3. The headline speedups do not measure R3D. The controlled studies evaluate R3D as proposed (Mixed-H2, ε = 0.05), which gives 1.22× on TRELLIS. The 1.842× in Table 1 comes from a different configuration (App. C.4): a whole-block-only H2 route with exact negative-branch CA caching. In Table 15, the same Whole-H2 route at ε = 0.05 gives only 1.07×. The extra speedup therefore comes from somewhere other than RRP or GAR, and GAR is not used at all in this setting. The other backbones also use their own configurations: K/V caching on TRELLIS.2, an H1 route on LumiTex, and a damped predictor on MatLat. None of these differences is mentioned in the main text or in Table 1.
Questions for authors

See weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.