acceptodds
Under review as a conference paper at ICLR 2027

Arbi3R: Arbitrary-Scale 3D Super-Resolution in Feed-Forward Novel View Synthesis

Abstract

Feed-forward novel view synthesis (NVS) is fast and generalizable, but renders at a fixed, low resolution. Recent feed-forward 3D super-resolution (SR) lifts this limit, yet predicts a fixed set of explicit 3D Gaussians, and trains a separate model for each integer scale. Arbitrary-scale SR otherwise exists only in 2D image space or in per-scene 3D NVS. None of the existing methods are at once generalizable and arbitrary-scale. We present Arbi3R, the first feed-forward (FF) framework to treat arbitrary-scale 3D-SR in FF NVS. For this, we replace the explicit Gaussians with an Implicit Function Decoder (IFD) that renders the novel views as a continuous function of pixel coordinates and scales. So, one trained model produces an output at any resolution from sparse, LR input views, including scales unseen at training. A Multi-View Geometry Encoder (MVGE) in our Arbi3R encodes the inputs into scene tokens, and the IFD decodes the target views through its scale-aware Coarse View Renderer (CVR) that produces an LR coarse view, an Epipolar Dual-Stream Detail Aggregator (EDDA) that recovers viewconsistent high-frequency details from the source views along epipolar lines, and a Coordinate-Modulated Neural Operator (CMNO) for continuous upsampling. On RealEstate10K, ACID and DL3DV, our Arbi3R outperforms the latest FF 3DSR specialists (SRSplat, SR3R) by 0.8 to 2.2 dB in PSNR, at their supported fixed scales, while uniquely spanning continuous and unseen resolutions. Arbi3R code will be released upon acceptance.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D reconstruction, 3DGS, Super Resolution

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_uLDwOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Arbi3R, a feed-forward framework for arbitrary-scale novel view synthesis from sparse low-resolution images. A VGGT-based encoder extracts multi-view features, a coarse renderer generates target-view features, and EDDA retrieves additional source-view information through epipolar-guided sampling. A coordinate-conditioned decoder then produces images at different resolutions. The model is trained across continuous scales and evaluated on reconstruction quality, cross-dataset generalization, and multi-view consistency.

Critical strengths
  1. The system achieves consistently strong reported reconstruction results. Improvements across several datasets and zero-shot evaluations make the method practically interesting. Supporting multiple output resolutions with one model also provides useful flexibility.

  2. Direct access to source-view information is effective. EDDA improves PSNR by approximately 1 dB in Table 6 and improves the reported consistency metrics. These results support the usefulness of supplementing coarse target-view features with source-view evidence.

Critical weaknesses
  1. The paper overstates its contribution to arbitrary-scale 3D reconstruction, and experimental comparisons do not adequately test the claimed advance.

In Eqs. (2)-(4), CVR and EDDA operate on a fixed low-resolution target grid. Higher-resolution outputs are produced by coordinate-based decoding of these target-view features, without refining the geometry-guided correspondence sampling for the additional output pixels. The features are multi-view conditioned, so this is more than ordinary single-image SR. Nevertheless, I view the main contribution as the addition of continuous-resolution decoding to a geometry-conditioned NVS pipeline, with limited innovation in the underlying 3D modeling.

The motivation further exaggerates the limitations of explicit representations. A fixed Gaussian representation can be rasterized at different resolutions. Changing the output grid does not inherently require changing the primitive count. Whether it produces good detail at those resolutions is an empirical question. SRSplat also reports cross-resolution generalization, including training at 2x and testing at 4x, contradicting the categorical "one trained model, only one scale" characterization.

The most relevant controls are missing. Table 3 enlarges fixed-scale baseline images using SRNO rather than directly rasterizing their Gaussian representations at the requested resolutions. It also omits comparison against Arbi3R's own x4 prediction resized to larger scales. These controls would directly test the value of continuous-scale decoding.

Other comparisons are less convincing for assessing the new architecture. Arbi3R has approximately four times as many parameters as SR3R, while the LargerNVS+SRNO baseline passes only rendered RGB to the SR model, unlike Arbi3R's richer feature and source-image connections. EDDA's overall gain is meaningful, but its dual streams, deformable offsets, top-m selection, and sparse gating receive little individual examination. Finally, all inputs are 64x64 under a single synthetic bilinear-resizing protocol, which substantially limits the practical scope of the evaluation.

  1. The scale encoding and sparse detail fusion have concrete technical problems that conflict with the paper's anti-aliasing and continuity explanations.

Appendix A derives a scalar variance from a pixel cell in normalized 2D image coordinates, but Eq. (5) applies it to six-dimensional Plucker-ray features. The corresponding footprint in ray coordinates depends on the camera projection and the transformation into ray directions and moments. Image-cell area alone does not describe that footprint. Moreover, the attenuation weakens as the requested output resolution increases, although CVR’s sampling grid remains fixed. I consider this heuristic scale conditioning. The claimed footprint integration and alias-free rendering do not follow from the derivation.

Eq. (10), as written, also introduces an avoidable discontinuity into the detail features. With a single active neighbor having feature (d≠0d\neq 0) and weight (ww), the normalized feature is (wd/w=dwd/w=d) for any (w>0w>0), but becomes zero when the valid weight reaches zero. Thus, an arbitrarily small contribution retains its full amplitude until its support disappears. View-dependent hard top-percentage selection introduces further discrete changes in the available features. The downstream decoder receives the valid weight and may compensate, so this does not prove visible flicker. However, the formulation introduces artificial feature jumps and relies on the decoder to suppress them, an awkward design for continuous-coordinate reconstruction.

Questions for authors
  1. Using existing checkpoints on a small DL3DV subset, please compare direct rasterization of the x4-trained Gaussian baselines at x3, x5.6, and x8 against the reported cascades. Please also compare Arbi3R's direct x5.6/x8 predictions with its own x4 output resized to those resolutions.

  2. How does the image-space variance in Eq. (16) map to a footprint in Plucker-ray coordinates, and why is the final output footprint appropriate for the fixed-resolution CVR? If this is heuristic scale conditioning, please revise the derivation and alias-free interpretation accordingly.

  3. Does the implementation of Eq. (10) use denominator stabilization or additional weighting not shown in the equation? Please clarify its behavior as the valid weight approaches zero, and provide a small existing-checkpoint test around mask-support boundaries to assess whether the feature discontinuity produces abrupt RGB changes.

Reviewer_1NBsOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Arbi3R, a feed-forward novel view synthesis framework that achieves arbitrary-scale 3D super-resolution within a single model. The key idea is to replace explic-it, fixed-budget 3D primitives with an Implicit Function Decoder that renders target pixels as a continuous function of coordinates and scale, using a scale-aware Coarse View Render-er, an Epipolar Dual-Stream Detail Aggregator that samples along deformable epipolar lines, and a Coordinate-Modulated Neural Operator for continuous upsampling. Across Re-alEstate10K, ACID, and DL3DV, Arbi3R achieves state-of-the-art performance versus recent feed-forward 3D-SR methods trained for fixed integer scales, while uniquely supporting continuous and unseen scale factors.

Critical strengths

To my knowledge, this is the first feed-forward framework that makes 3D super-resolution arbitrary-scale in a single model, replacing fixed-budget explicit primitives with an implicit decoder conditioned on coordinates and scale. The formulation is worked out rather than bolted on: the query cell is defined analytically from the target scale, the positional-encoding variance is derived from the cell footprint, and each decoding stage has a precise functional form. This capability is exactly what fixed-scale methods like SRSplat and SR3R lack.

The experimental validation is extensive. Beyond in-domain comparisons on RealEs-tate10K and ACID, the authors evaluate a 960P version of DL3DV at both seen and unseen scales, run zero-shot transfer to ACID, DL3DV, and DTU, measure multi-view consistency with Ewarp and MEt3R, and ablate the main proposed modules and the loss design. Both baselines were retrained under the same preprocessing protocol, so the 0.8–2.2 dB gains are measured under fair conditions.

Critical weaknesses

Although the title claims 3D super-resolution, the evidence for the '3D' is entirely indirect. The question on which the paper's framing rests is whether Arbi3R can truly super-resolve the 3D scene or can only super-resolve the rendered images. Yet the paper shows no direct 3D results. Although the paper explicitly states that it produces no explicit 3D Gaussians, this contradicts the title. The prior work that this paper is benchmarked against (SRSplat and SR3R) has established that 3D super-resolution involves obtaining a denser, higher-fidelity 3D representation from which higher-resolution views can be rendered. However, Arbi3R deliberately forgoes this. Instead, the evidence offered remains in view space. I would therefore ask the authors to reconsider the terminology, either substantiating the '3D' in the title with direct 3D-level evidence or softening the wording. A multi-view zoom panel of one surface would largely settle this.

Questions for authors
  1. To what extent are the results sensitive to the following EDDA hyperparameters? A brief sensitivity study would demonstrate that these are robust defaults rather than finely tuned ones.

  2. What is the runtime and memory breakdown for each stage (MVGE, CVR, EDDA, and CMNO) at representative scales, and how does this compare with SR3R and SRSplat on the same hardware? Table 15 reports totals only.

  3. Could you provide examples or statistics of failure cases involving wide-baseline targets, strong occlusions, reflective surfaces, or textureless surfaces?

  4. Training scale [1,4], so >×4 is extrapolation. App. K visualizes up to ×16. Is degradation onset observable beyond ×4? GT exists only up to ×5.6/×8, so ×12.3–×16 has no quantitative evaluation; ×16 output (1024²) exceeds source resolution, hence content is synthesized. To verify that beyond-range renders are scene-anchored rather than smooth extrapolations of the ×4 result, GT-free checks are needed.

  5. The formulation in Sec. 3.3, especially equations 7–9, introduces over a dozen symbols and module abbreviations in quick succession, making the derivation difficult to follow. Could you walk through one derivation with fewer abbreviations?

  6. The MVGE has a frozen VGGT-1B backbone, and the paper states that this backbone limits the quality (Sec. 5). How critical is this choice for performance?

  7. Section 3.4 justifies two-phase training due to the difficulty of learning continuous high-frequency details from scratch. However, no experiment substantiates this. Was single-phase end-to-end training attempted, and did it fail to converge or simply perform worse?

Reviewer_RtTcOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This work focuses on feed-forward 3D super-resolution, supporting arbitrary-scale super-resolution. It replaces the explicit Gaussians with an Implicit Function Decoder to render the novel view as a continuous/implicit function of pixel coordinates and scales. The decoder contains (1) a scale-aware Coarse View Renderer (CVR) to produce an LR coarse view, (2) an Epipolar Dual-Stream Detail Aggregator (EDDA) to recover view-consistent high-frequency details from the source views along epipolar lines, and (3) a Coordinate-Modulated Neural Operator (CMNO) for continuous upsampling. Finally, the proposed method outperforms other methods significantly.

Critical strengths

The proposed method is novel, introducing the first framework to support arbitrary-scale 3D super-resolution for feed-forward novel view synthesis. It achieves this by introducing an Implicit Function Decoder, including a scale-aware Coarse View Renderer (CVR), an Epipolar Dual-Stream Detail Aggregator (EDDA), and a Coordinate-Modulated Neural Operator (CMNO) to obtain high-resolution features from low-resolution features. The experimental results demonstrate that this work achieves significant improvement compared to other methods.

Critical weaknesses

None

Questions for authors
  • How is the hyperparameter ρ\rho in Line 220 selected, and how sensitive is the method to its value?
  • Is δ\delta in Eq. (7) a learnable parameter or predicted by an MLP?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.