acceptodds
Under review as a conference paper at ICLR 2027

Puffin-World: Scaling a Unified Multimodal Model with Native 3D World States

Abstract

We propose Puffin-World, a unified multimodal architecture that integrates physical understanding, spatial simulation, and 3D world generation and reconstruction without relying on external offline modules. To reliably construct and interact with 3D worlds, our framework jointly models three native world states: physics (gravity field and latitude), geometry (depth), and appearance (image), together with a unified Omni-Camera representation that supports diverse tasks and flexible motions. Beyond modeling these states, we introduce a strategy for propagating physical dynamics across future frames. By grounding absolute camera properties in the real world, Puffin-World enables physically consistent and visually stable world generation. We further couple appearance and geometry within a single generative process, jointly synthesizing each future view and reconstructing its underlying geometry. This unified paradigm enables interleaved closed-loop applications requiring synergy across multiple tasks, including mimic and self-calibrated world exploration. To scale Puffin-World to complex scenarios, we construct Puffin-16M, comprising 15 million vision–language–camera triplets and 1 million trajectories featuring various and challenging motions. To foster further research in this area, we will release the code, models, and datasets.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Unified Multimodal Model, World Model, Spatial Intellience

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

The paper unifies physical-world understanding, free-viewpoint spatial simulation, and 3D world generation/reconstruction within a single multimodal model built on three "native world states": physics (gravity field, latitude), geometry (depth), and appearance (RGB). Its Omni-Camera representation concatenates a gravity-anchored absolute field (up-vector, latitude) with a relative ray field. Physics propagation rotates the reference-view gravity direction to each future view via the known relative rotation. Appearance and depth share one VAE (depth encoded as an invertible Hilbert-color image) and are generated jointly under a shared flow-matching objective with asymmetric attention. The authors build Puffin-16M (15M vision-language-camera triplets plus 1M trajectories) and annotate 28 public datasets (~44.5M images). Results include state-of-the-art camera-to-world understanding on four benchmarks, camera-controllable generation on Puffin-Cam-Bench (FID 75.93), and 3D world generation on RealEstate10K and Puffin-Traj-Bench.

Critical strengths
  1. A well-motivated, unified camera representation. Combining gravity-anchored absolute orientation with a relative ray field, and using it to drive single-view control, multi-view synthesis, and reconstruction, is a coherent idea supported by strong camera-understanding results.
  2. A substantial potential community resource. Puffin-16M and the annotations for 28 public datasets (~44.5M images) could support camera-grounded modeling beyond this system. Their scale and coverage are valuable, contingent on release and on documenting annotation quality.
Critical weaknesses
  1. Lack of an independent controllability evaluation. Using the model's own perception branch to score all generated images leaves open whether the camera-control margins depend on the chosen evaluator. Re-scoring the existing outputs independently would address this central claim without changing the generation experiment.
  2. Architectural and conditioning effects are not fully isolated. The comparison with Puffin changes both the architecture and the training data. Table A5 is a useful propagation ablation, but the absolute-camera information available to its baseline needs to be specified. Matched-data camera-injection comparisons and a clear conditioning definition would distinguish architectural improvements from data scaling or additional camera information.
Questions for authors
  1. Independent controllability. Please re-score the same generated images (Tables 2 and 3) with independent estimators such as GeoCalib and AnyCalib, to test whether the camera-control margins survive evaluator choice.
  2. Table A5 baseline. What does the no-propagation baseline receive—relative rays only, a repeated initial field, or per-view absolute fields? Please specify the conditioning and whether weights, data, and trajectories are matched.
  3. Matched-data control. With backbone, training data, and update budget fixed, how does replacing the new camera-injection path with Puffin-style Perspective-Field conditioning change camera errors and FID?
5 comments
AuthorsOpenReview ↗

Response to Reviewer ioPk (part 1/3)

W1. Independent validation of camera controllability.

Thank you for this valuable suggestion. We re-evaluated the generated outputs in Sections 3.3, 3.4, and Appendix A0.5 using GeoCalib [1], an independent camera estimator with the state-of-the art performance. The tables below report both evaluators. Specifically, each paired camera-metric cell reports Puffin-World evaluator / underlined GeoCalib evaluator. All angular errors are in degrees; AUC values retain the original 0–1 scale.

Table R3-1. Camera-controllable generation on Puffin-Cam-Bench

MethodUp mean ↓Up median ↓Latitude mean ↓Latitude median ↓Gravity mean ↓Gravity median ↓FID ↓
GPT Image224.13 / 24.19‾\underline{24.19}23.44 / 23.48‾\underline{23.48}16.42 / 16.49‾\underline{16.49}15.98 / 16.02‾\underline{16.02}28.83 / 28.74‾\underline{28.74}29.38 / 29.62‾\underline{29.62}99.36
Nano Banana224.01 / 24.16‾\underline{24.16}23.19 / 23.31‾\underline{23.31}15.98 / 15.84‾\underline{15.84}15.52 / 15.42‾\underline{15.42}28.00 / 27.81‾\underline{27.81}26.92 / 26.62‾\underline{26.62}90.11
Qwen-Image2-Pro24.12 / 24.06‾\underline{24.06}23.51 / 23.46‾\underline{23.46}16.45 / 16.12‾\underline{16.12}15.53 / 15.24‾\underline{15.24}28.57 / 28.25‾\underline{28.25}28.69 / 29.02‾\underline{29.02}98.51
FLUX.2-dev24.28 / 24.37‾\underline{24.37}23.62 / 23.69‾\underline{23.69}15.75 / 15.91‾\underline{15.91}15.24 / 15.26‾\underline{15.26}28.31 / 28.46‾\underline{28.46}28.36 / 28.95‾\underline{28.95}97.06
Z-Image24.64 / 24.33‾\underline{24.33}24.04 / 23.71‾\underline{23.71}16.96 / 16.74‾\underline{16.74}16.09 / 15.70‾\underline{15.70}29.41 / 28.56‾\underline{28.56}28.82 / 27.84‾\underline{27.84}98.33
PreciseCam13.37 / 13.89‾\underline{13.89}12.61 / 12.92‾\underline{12.92}12.31 / 12.98‾\underline{12.98}12.01 / 12.70‾\underline{12.70}17.07 / 17.78‾\underline{17.78}14.01 / 14.37‾\underline{14.37}90.89
Puffin3.86 / 3.49‾\underline{3.49}3.48 / 3.18‾\underline{3.18}4.77 / 4.74‾\underline{4.74}4.43 / 4.57‾\underline{4.57}4.92 / 5.36‾\underline{5.36}2.87 / 2.59‾\underline{2.59}80.29
Puffin-World0.96 / 2.45‾\underline{2.45}0.84 / 2.30‾\underline{2.30}1.34 / 3.48‾\underline{3.48}1.26 / 3.41‾\underline{3.41}1.32 / 3.90‾\underline{3.90}0.79 / 1.27‾\underline{1.27}75.93

Table R3-2. Camera-control evaluation on Puffin-Traj-Bench

MethodRoll error ↓Roll AUC@1° ↑Roll AUC@5° ↑Roll AUC@10° ↑Pitch error ↓Pitch AUC@1° ↑Pitch AUC@5° ↑Pitch AUC@10° ↑FoV error ↓
MotionCtrl3.20 / 1.39‾\underline{1.39}0.25 / 0.45‾\underline{0.45}0.43 / 0.56‾\underline{0.56}0.54 / 0.63‾\underline{0.63}2.15 / 1.83‾\underline{1.83}0.28 / 0.34‾\underline{0.34}0.49 / 0.53‾\underline{0.53}0.59 / 0.62‾\underline{0.62}10.04 / 4.37‾\underline{4.37}
CameraCtrl11.79 / 3.58‾\underline{3.58}0.05 / 0.23‾\underline{0.23}0.11 / 0.40‾\underline{0.40}0.21 / 0.52‾\underline{0.52}10.81 / 3.60‾\underline{3.60}0.05 / 0.17‾\underline{0.17}0.16 / 0.37‾\underline{0.37}0.26 / 0.53‾\underline{0.53}45.84 / 12.03‾\underline{12.03}
ViewCrafter1.92 / 1.95‾\underline{1.95}0.32 / 0.32‾\underline{0.32}0.53 / 0.53‾\underline{0.53}0.66 / 0.66‾\underline{0.66}2.89 / 3.38‾\underline{3.38}0.23 / 0.19‾\underline{0.19}0.43 / 0.39‾\underline{0.39}0.58 / 0.55‾\underline{0.55}6.43 / 7.33‾\underline{7.33}
SEVA1.35 / 0.86‾\underline{0.86}0.38 / 0.56‾\underline{0.56}0.63 / 0.75‾\underline{0.75}0.75 / 0.83‾\underline{0.83}1.76 / 1.41‾\underline{1.41}0.29 / 0.37‾\underline{0.37}0.56 / 0.65‾\underline{0.65}0.71 / 0.79‾\underline{0.79}12.59 / 5.15‾\underline{5.15}
MVGenMaster9.24 / 7.76‾\underline{7.76}0.06 / 0.17‾\underline{0.17}0.14 / 0.28‾\underline{0.28}0.28 / 0.38‾\underline{0.38}9.43 / 8.44‾\underline{8.44}0.08 / 0.09‾\underline{0.09}0.17 / 0.20‾\underline{0.20}0.30 / 0.33‾\underline{0.33}26.98 / 6.38‾\underline{6.38}
Puffin-World0.80 / 0.79‾\underline{0.79}0.57 / 0.58‾\underline{0.58}0.79 / 0.79‾\underline{0.79}0.88 / 0.88‾\underline{0.88}1.10 / 1.34‾\underline{1.34}0.45 / 0.40‾\underline{0.40}0.72 / 0.66‾\underline{0.66}0.83 / 0.79‾\underline{0.79}2.96 / 3.87‾\underline{3.87}
AuthorsOpenReview ↗

Response to Reviewer ioPk (part 2/3)

Table R3-3. Physics-propagation ablation

SettingMotionPSNR ↑SSIM ↑LPIPS ↓Roll error ↓Pitch error ↓
BaselineRoll22.970.690.122.35 / 2.07‾\underline{2.07}6.73 / 6.48‾\underline{6.48}
Pitch18.530.600.262.17 / 1.76‾\underline{1.76}5.60 / 6.04‾\underline{6.04}
Yaw18.120.590.301.31 / 1.68‾\underline{1.68}3.10 / 3.52‾\underline{3.52}
Average19.870.620.231.94 / 1.83‾\underline{1.83}5.14 / 5.35‾\underline{5.35}
Baseline with PPRoll24.020.740.102.04 / 1.95‾\underline{1.95}6.31 / 6.34‾\underline{6.34}
Pitch19.180.650.242.03 / 1.61‾\underline{1.61}5.28 / 5.87‾\underline{5.87}
Yaw18.420.620.301.22 / 1.68‾\underline{1.68}2.67 / 3.45‾\underline{3.45}
Average20.540.670.211.76 / 1.74‾\underline{1.74}4.75 / 5.22‾\underline{5.22}

From the re-evaluated results, we have three observations:

  1. Camera-controllable image generation. On Puffin-Cam-Bench, Puffin-World retains the lowest errors across all six camera metrics. With GeoCalib, its mean up-vector, latitude, and gravity errors are 2.45∘‾\underline{2.45^\circ}, 3.48∘‾\underline{3.48^\circ}, and 3.90∘‾\underline{3.90^\circ}, compared with 3.49∘‾\underline{3.49^\circ}, 4.74∘‾\underline{4.74^\circ}, and 5.36∘‾\underline{5.36^\circ} for the strongest baseline, Puffin.

  2. Multi-view generation. On Puffin-Traj-Bench, Puffin-World retains the lowest median roll, pitch, and FoV errors: 0.79∘‾\underline{0.79^\circ}, 1.34∘‾\underline{1.34^\circ}, and 3.87∘‾\underline{3.87^\circ}. It also achieves the highest or joint-highest roll/pitch AUC at every reported threshold, tying with SEVA at pitch AUC@10° (0.79‾\underline{0.79}).

  3. Physics propagation. GeoCalib also confirms the average benefit of PP: roll error decreases from 1.83∘‾\underline{1.83^\circ} to 1.74∘‾\underline{1.74^\circ}, and pitch error from 5.35∘‾\underline{5.35^\circ} to 5.22∘‾\underline{5.22^\circ}. Across individual motion categories, the camera errors either improve or remain unchanged at the reported precision.

Although absolute scores and some baseline orderings vary between evaluators, the overall performance trends and our main comparative conclusions remain consistent. This independent check supports the credibility of using our physics-perception branch to compare camera-controlled spatial simulation across the tested methods, and shows that our main conclusions are not specific to the original evaluator. We have included both sets of results in the revised manuscript and clarified the evaluator dependence of individual scores and performance margins.

[1] Veicht, Alexander, et al. "Geocalib: Learning single-image calibration with geometric optimization." European Conference on Computer Vision, 2024.

AuthorsOpenReview ↗

Response to Reviewer ioPk (part 3/3)

W2. Separating architectural effects, data scaling, and the absolute information in the PP baseline.

Thank you for highlighting this point. The primary goal of our camera interface is to support scalable, unified generation across tasks while preserving single-view quality. We clarify the comparison and conditioning protocol below.

  1. Matched-backbone and matched-data comparison. We compare Puffin-style camera-token injection with our latent-space fusion using the same SD3.5-Large MMDiT backbone and training data. For this shared single-view task, both interfaces encode the target absolute camera configuration, while the relative ray component of Omni-Camera is held constant. Thus, the comparison does not introduce additional inter-view motion information.

    Table R3-4. Camera-interface comparison under a shared backbone and training dataset.

    Metric or capabilityPuffin-style injectionPuffin-World fusion
    FID ↓78.1275.93
    Gravity median error ↓1.220.79
    Inference time ↓44.5 s19.5 s
    Compute per image ↓7.0 PFLOPs2.9 PFLOPs
    Additional camera tokens✓✗
    Camera-controllable text-to-image✓✓
    Joint multi-view image-to-3D / text-to-3D✗✓
    Native RGB-D generation✗✓

    The two interfaces achieve comparable FID and camera metrics, while our fusion reduces inference time and computation substantially. This comparison controls the backbone and training data and evaluates the camera interface as a whole, including its representation and injection mechanism. We therefore do not attribute the difference to latent-space fusion alone.

  2. Explicit conditioning definition for Table A5. The no-propagation baseline retains the relative ray maps but replaces the absolute Perspective Field with constant values. It therefore receives no informative absolute gravity field through this conditioning component. With physics propagation, the Perspective Field is instead constructed from the reference view’s estimated absolute camera state and the commanded relative rotations. Both settings use the same checkpoint, reference inputs, and camera trajectories; only the Perspective Field changes. During training, we randomly set this field with probability to constant while retaining the relative rays, so the model encounters both informative and constant-field conditions. Table A5 therefore evaluates the benefit of propagated gravity conditioning, separately from the camera-interface comparison above.

  3. The scalability advantage. Puffin’s interface VAE-encodes a three-channel Perspective Field as additional camera tokens. Our fusion module accommodates the nine-channel Omni-Camera representation and view-role mask within existing latent positions, providing a shared interface for absolute camera grounding, relative motion, and different view roles. This provides a practical basis for unified text-to-image, image-to-3D, and text-to-3D generation, including joint multi-view RGB-D outputs.

We have revised Section 2.2 to emphasize this scalability contribution, identify the original Puffin comparison as system-level, and explicitly document the matched comparison and Table A5 conditioning protocol in the manuscript.

Q1. Re-scoring the camera metrics.

Please refer to our response to W1 for more details. The overall performance trends and our main comparative conclusions remain consistent.

Q2. What does the no-propagation baseline receive, and what is held fixed?

The no-propagation baseline receives relative ray maps, with the absolute Perspective Field replaced by constant values; it uses neither a repeated initial field nor per-view absolute fields. Both settings use the same checkpoint, training data, reference inputs, and camera trajectories. The only change at inference is whether the Perspective Field is constant or constructed through physics propagation. Please refer to our response to W2 for the training augmentation and detailed conditioning protocol.

Q3. Matched-data comparison with Puffin-style camera injection.

Please refer to our response to W2 and Table R3-4 for the comparison protocol and results. Our primary objective is efficient camera conditioning that applies across tasks, rather than improving a single-task score alone. As emphasized in W2, our interface integrates absolute camera grounding, relative motion, and view-role information without introducing additional camera tokens. This significantly reduces computational cost while providing a shared conditioning formulation for text-to-image, image-to-3D, and text-to-3D generation, including joint multi-view RGB-D outputs. Its central advantage is therefore the combination of comparable single-view quality, greater efficiency, and broader task applicability within one scalable unified model.

Reviewer_HK4POpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

Puffin-World extends Puffin to unify single-image physical camera perception, camera-controllable text-to-image generation, and multi-view 3D scene generation with dense geometry prediction, all within a single framework. It introduces a 9-channel Omni-Camera representation that combines 3-channel absolute Perspective Fields with 6-channel relative ray maps, injected via pixel-wise addition and sparse cross-layer re-injection. Physics Propagation transfers the estimated initial gravity direction across views during multi-view synthesis. To output dense geometry without modifying the frozen SD3.5 VAE, depth maps are encoded as 3-channel RGB images via a 3D Hilbert curve, trained with asymmetric attention and a linear loss warmup. The authors also construct Puffin-16M and re-annotate 28 public datasets with absolute camera parameters.

Critical strengths
  1. The work is well-motivated. Existing camera-controlled models rely mainly on relative poses, making them easy to orientation drift under large rotations. Omni-Camera combines absolute physical references with relative poses, while Physics Propagation maintains gravity consistency across frames, yielding clear gains on rotational trajectories.

  2. The geometry-appearance integration is efficient. Hilbert depth encoding reuses the frozen SD3.5 VAE, while Asymmetric Attention and depth-loss warmup preserve RGB quality during training.

  3. The data contribution is valuable. Built from 900K panoramas, Puffin-16M improves coverage of large rotations and diverse aspect ratios. Camera re-annotation and DA3-based dense depth completion across 28 public datasets also increase its community value, if it is fully released.

Critical weaknesses
  1. 3D generation and reconstruction lack quantitative validation. Although it is claimed to be a core contribution, neither depth accuracy nor point-cloud quality is evaluated against relevant methods. The point-cloud fusion procedure is also unclear, and the manuscript provides no ablations for Hilbert encoding, Asymmetric Attention, or depth-loss warmup.

  2. Camera-control evaluation risks self-evaluation bias. Tables 2, 3, and A5 use Puffin-World's own understanding branch as the evaluator, potentially favoring its generation branch through shared training biases. Thus, this evaluation requires independent evaluators. Conditioning general-purpose image models with raw camera parameters also makes those comparisons questionable.

Questions for authors
  1. Please report quantitative depth and multi-view reconstruction results. Does Stage IV affect RGB generation quality (PSNR/LPIPS), given that Asymmetric Attention blocks geometry-to-appearance information flow?

  2. Which variant (Base or Pro) is used in Tables 1–3 and A5? Does reducing the LLM from 7B to 1.5B in Pro affect calibration accuracy? A direct comparison of both variants on understanding and generation tasks would clarify this.

  3. What advantage does joint depth generation offer over applying a feed-forward model such as DA3 or CUT3R to generated RGB views?

6 comments
AuthorsOpenReview ↗

Response to Reviewer HK4P (part 1/5)

W1. Quantitative geometry evaluation, the fusion procedure, and geometry-design ablations.

Thank you for your valuable suggestion. We have added quantitative geometry evaluations across four datasets. To be specific, we use held-out scenes with independent geometric references: DL3DV [1], 7Scenes [2], NRGBD [3], and ETH3D [4]. Each clip contains real references and generated views along a ground-truth trajectory with translation, with camera motion matched across datasets. Frozen DUSt3R [5], CUT3R [6], and VGGT [7] reconstruct the same sequence for comparison with our native geometry.

Please note that Puffin-World jointly generates appearance and geometry, whereas the reconstruction baselines recover geometry from provided RGB observations. For regions unseen in the reference, plausible generated content may differ from the particular ground-truth scene, making full-view depth errors insufficient to characterize generative geometry. We therefore combine accuracy evaluation on observed/co-visible regions with cross-view consistency over the generated sequence. The quantitative evaluation results are listed in Table R2-1.

Table R2-1. Native geometry versus post-hoc reconstruction. Reference-view AbsRel is scale-aligned. Co-visible δ1 measures the fraction of valid co-visible pixels whose predicted depth is within a factor of 1.25 of ground truth, using a single reference-fitted scale across all views.. Depth consistency measures reprojected points within 5% / 10% depth error. Point cloud consistency measures the fraction of points confirmed by ≥2 other views within 2 px and 5% depth error.

DatasetMethodRef.-view AbsRel ↓Co-visible δ1 ↑Consistency (depth) <5% / <10% ↑Consistency (point cloud) ↑
DL3DVDUSt3R0.0780.89668% / 83%71.8%
CUT3R0.0340.95771% / 83%75.9%
VGGT0.0560.92779% / 85%80.3%
Puffin-World0.0720.93172% / 86%76.6%
7ScenesDUSt3R0.0400.97479% / 92%79.8%
CUT3R0.0420.96687% / 95%85.2%
VGGT0.0520.96690% / 95%89.2%
Puffin-World0.0520.97689% / 94%87.4%
NRGBDDUSt3R0.0270.96283% / 92%81.2%
CUT3R0.0260.96379% / 90%73.3%
VGGT0.0230.96089% / 93%80.8%
Puffin-World0.0440.96184% / 92%79.5%
ETH3DDUSt3R0.0280.87351% / 67%9.7%
CUT3R0.0290.86940% / 64%11.3%
VGGT0.0330.86955% / 75%15.9%
Puffin-World0.0580.89651% / 72%18.4%

We further provide some analyses in terms of the above evaluations.

  1. Competitive consistency with remaining precision gaps. On DL3DV, 7Scenes, and NRGBD, our native geometry achieves the best depth consistency at the 10% threshold on DL3DV and remains within one percentage point of the best on the other two datasets. At the stricter 5% threshold, it exceeds DUSt3R and CUT3R across all three datasets, although VGGT remains stronger. Co-visible δ1 is highest on 7Scenes and within 0.03 of the best on DL3DV and NRGBD. On wide-baseline ETH3D, Puffin-World achieves the highest co-visible δ1 and point-cloud consistency. These results support coherent geometry across translated viewpoints, while reference-view accuracy and fine-scale consistency remain areas for improvement.

  2. Training scope for geometry. Our depth supervision comes from only five datasets—DL3DV, Hypersim, MVS-Synth, TartanAir, and ScanNet. In comparison, CUT3R uses 29 depth-supervised dataset sources within its 32-dataset training collection, while VGGT lists 17 training datasets, corresponding to 5.8× and 3.4× our source count, respectively. This narrower range of geometric supervision may contribute to the remaining precision gap relative to dedicated reconstruction models.

  3. Value of unification. Our contribution is to natively unify physical, geometric, and appearance world states, enabling understanding, generation, and reconstruction within one model. Compared with applying a separate feed-forward model such as DA3 or CUT3R to generated RGB views, our approach produces depth alongside appearance in the commanded camera frame, making geometry directly available for multi-view fusion without an additional reconstruction network or inference pass. Depth prediction also accesses the shared multi-view appearance, text, and camera context, explicitly conditioning geometry on the intended scene and trajectory. These benefits complement the competitive cross-view consistency demonstrated above. In the future, we plan to expand the scale and diversity of 3D supervision to improve geometric accuracy and advance toward state-of-the-art performance across understanding, generation, and reconstruction.

AuthorsOpenReview ↗

Response to Reviewer HK4P (part 2/5)

For the point cloud fusion procedure, we construct the point cloud by back-projecting each view’s native depth into a shared world coordinate frame using its intrinsics and commanded camera pose, with colors taken from the corresponding generated RGB image. The pipeline involves no TSDF fusion, meshing, or learned refinement.

To evaluate Hilbert encoding, we pass the same ground-truth depth maps through the frozen VAE. Comparing Hilbert and grayscale encodings with the same depth transform f(d)f(d) isolates the contribution of the color mapping.

Table R2-2. Frozen-VAE depth-encoding diagnostic on DL3DV frames. Median error denotes the median per-pixel relative error; Within 1% and Within 5% report the percentages of pixels below these error thresholds.

Depth encodingMedian error ↓Within 1% ↑Within 5% ↑
Hilbert color path of f(d)f(d) (ours)1.18%44.9%95.9%
Same f(d)f(d), grayscale7.79%4.7%33.4%
Turbo colormap of f(d)f(d)2.34%24.2%88.0%
Disparity, grayscale1.94%41.2%80.1%
Log depth, grayscale2.55%22.8%87.3%
Linear depth, grayscale12.50%2.3%16.1%

From the above results, we can observe Hilbert achieves the best results on all three reported metrics. Compared with grayscale using the same f(d)f(d), it reduces median error from 7.79% to 1.18% and increases the fraction within 5% error from 33.4% to 95.9%. These results support its fidelity for typical pixels through the existing VAE.

Additionally, removing the asymmetric attention allows the newly introduced depth tokens to directly influence the RGB pathway, reducing PSNR and SSIM by 1.26 dB and 0.11, respectively, on the RealEstate10K test set. This supports its role in limiting interference with appearance generation when adding geometry. Depth-loss warmup complements this design by gradually increasing the geometry loss weight from zero, limiting early depth gradients on the shared modulation and output layers to stabilize optimization and facilitate convergence. Its effect is best assessed through training stability and convergence trajectories, which final accuracy alone does not fully capture. We will incorporate these analyses and clarify the warmup schedule in the revised manuscript.

[1] Ling, Lu, et al. "Dl3dv-10k: A large-scale scene dataset for deep learning-based 3d vision." IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024.

[2] Shotton, Jamie, et al. "Scene coordinate regression forests for camera relocalization in RGB-D images." IEEE Conference on Computer Vision and Pattern Recognition. 2013.

[3] Azinović, Dejan, et al. "Neural rgb-d surface reconstruction." IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022.

[4] Schops, Thomas, et al. "A multi-view stereo benchmark with high-resolution images and multi-camera videos." IEEE Conference on Computer Vision and Pattern Recognition, 2017.

[5] Wang, Shuzhe, et al. "Dust3r: Geometric 3d vision made easy." IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024.

[6] Wang, Qianqian, et al. "Continuous 3d perception model with persistent state." IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025.

[7] Wang, Jianyuan, et al. "Vggt: Visual geometry grounded transformer." IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025.

AuthorsOpenReview ↗

Response to Reviewer HK4P (part 3/5)

W2. Self-evaluation bias and the conditioning of general-purpose image models.

Thank you for this helpful suggestion. We re-evaluated the generated outputs in Sections 3.3, 3.4, and Appendix A0.5 using GeoCalib [8], an independent camera estimator with the state-of-the art performance. The tables below report both evaluators. Specifically, each paired camera-metric cell reports Puffin-World evaluator / underlined GeoCalib evaluator. All angular errors are in degrees; AUC values retain the original 0–1 scale.

Table R2-3. Camera-controllable generation on Puffin-Cam-Bench

MethodUp mean ↓Up median ↓Latitude mean ↓Latitude median ↓Gravity mean ↓Gravity median ↓FID ↓
GPT Image224.13 / 24.19‾\underline{24.19}23.44 / 23.48‾\underline{23.48}16.42 / 16.49‾\underline{16.49}15.98 / 16.02‾\underline{16.02}28.83 / 28.74‾\underline{28.74}29.38 / 29.62‾\underline{29.62}99.36
Nano Banana224.01 / 24.16‾\underline{24.16}23.19 / 23.31‾\underline{23.31}15.98 / 15.84‾\underline{15.84}15.52 / 15.42‾\underline{15.42}28.00 / 27.81‾\underline{27.81}26.92 / 26.62‾\underline{26.62}90.11
Qwen-Image2-Pro24.12 / 24.06‾\underline{24.06}23.51 / 23.46‾\underline{23.46}16.45 / 16.12‾\underline{16.12}15.53 / 15.24‾\underline{15.24}28.57 / 28.25‾\underline{28.25}28.69 / 29.02‾\underline{29.02}98.51
FLUX.2-dev24.28 / 24.37‾\underline{24.37}23.62 / 23.69‾\underline{23.69}15.75 / 15.91‾\underline{15.91}15.24 / 15.26‾\underline{15.26}28.31 / 28.46‾\underline{28.46}28.36 / 28.95‾\underline{28.95}97.06
Z-Image24.64 / 24.33‾\underline{24.33}24.04 / 23.71‾\underline{23.71}16.96 / 16.74‾\underline{16.74}16.09 / 15.70‾\underline{15.70}29.41 / 28.56‾\underline{28.56}28.82 / 27.84‾\underline{27.84}98.33
PreciseCam13.37 / 13.89‾\underline{13.89}12.61 / 12.92‾\underline{12.92}12.31 / 12.98‾\underline{12.98}12.01 / 12.70‾\underline{12.70}17.07 / 17.78‾\underline{17.78}14.01 / 14.37‾\underline{14.37}90.89
Puffin3.86 / 3.49‾\underline{3.49}3.48 / 3.18‾\underline{3.18}4.77 / 4.74‾\underline{4.74}4.43 / 4.57‾\underline{4.57}4.92 / 5.36‾\underline{5.36}2.87 / 2.59‾\underline{2.59}80.29
Puffin-World0.96 / 2.45‾\underline{2.45}0.84 / 2.30‾\underline{2.30}1.34 / 3.48‾\underline{3.48}1.26 / 3.41‾\underline{3.41}1.32 / 3.90‾\underline{3.90}0.79 / 1.27‾\underline{1.27}75.93

Table R2-4. Camera-control evaluation on Puffin-Traj-Bench

MethodRoll error ↓Roll AUC@1° ↑Roll AUC@5° ↑Roll AUC@10° ↑Pitch error ↓Pitch AUC@1° ↑Pitch AUC@5° ↑Pitch AUC@10° ↑FoV error ↓
MotionCtrl3.20 / 1.39‾\underline{1.39}0.25 / 0.45‾\underline{0.45}0.43 / 0.56‾\underline{0.56}0.54 / 0.63‾\underline{0.63}2.15 / 1.83‾\underline{1.83}0.28 / 0.34‾\underline{0.34}0.49 / 0.53‾\underline{0.53}0.59 / 0.62‾\underline{0.62}10.04 / 4.37‾\underline{4.37}
CameraCtrl11.79 / 3.58‾\underline{3.58}0.05 / 0.23‾\underline{0.23}0.11 / 0.40‾\underline{0.40}0.21 / 0.52‾\underline{0.52}10.81 / 3.60‾\underline{3.60}0.05 / 0.17‾\underline{0.17}0.16 / 0.37‾\underline{0.37}0.26 / 0.53‾\underline{0.53}45.84 / 12.03‾\underline{12.03}
ViewCrafter1.92 / 1.95‾\underline{1.95}0.32 / 0.32‾\underline{0.32}0.53 / 0.53‾\underline{0.53}0.66 / 0.66‾\underline{0.66}2.89 / 3.38‾\underline{3.38}0.23 / 0.19‾\underline{0.19}0.43 / 0.39‾\underline{0.39}0.58 / 0.55‾\underline{0.55}6.43 / 7.33‾\underline{7.33}
SEVA1.35 / 0.86‾\underline{0.86}0.38 / 0.56‾\underline{0.56}0.63 / 0.75‾\underline{0.75}0.75 / 0.83‾\underline{0.83}1.76 / 1.41‾\underline{1.41}0.29 / 0.37‾\underline{0.37}0.56 / 0.65‾\underline{0.65}0.71 / 0.79‾\underline{0.79}12.59 / 5.15‾\underline{5.15}
MVGenMaster9.24 / 7.76‾\underline{7.76}0.06 / 0.17‾\underline{0.17}0.14 / 0.28‾\underline{0.28}0.28 / 0.38‾\underline{0.38}9.43 / 8.44‾\underline{8.44}0.08 / 0.09‾\underline{0.09}0.17 / 0.20‾\underline{0.20}0.30 / 0.33‾\underline{0.33}26.98 / 6.38‾\underline{6.38}
Puffin-World0.80 / 0.79‾\underline{0.79}0.57 / 0.58‾\underline{0.58}0.79 / 0.79‾\underline{0.79}0.88 / 0.88‾\underline{0.88}1.10 / 1.34‾\underline{1.34}0.45 / 0.40‾\underline{0.40}0.72 / 0.66‾\underline{0.66}0.83 / 0.79‾\underline{0.79}2.96 / 3.87‾\underline{3.87}
Reviewer_GTprOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Puffin-World extends the camera-centric Puffin framework with multiview RGB and depth generation. Its Omni-Camera representation combines gravity-aligned orientation with relative camera rays, while physics propagation carries the reference view’s gravity direction along a prescribed trajectory. The paper introduces 15 million image–language–camera triplets and one million trajectories. Experiments evaluate camera estimation, controlled image generation, and multiview synthesis, with qualitative demonstrations of depth prediction and 3D reconstruction.

Critical strengths
  1. The camera representation has a clear motivation. Combining gravity-aligned orientation with relative camera geometry provides a coherent way to connect camera estimation with generation across views.

  2. The dataset targets a useful gap in camera-motion coverage. Its emphasis on substantial roll and pitch variations provides explicit supervision for viewpoints that are underrepresented in several existing trajectory datasets.

Critical weaknesses
  1. The methodological contribution needs to be distinguished more clearly from prior work and the expanded training recipe.

    1. The joint appearance–geometry component in Section 2.3 is not clearly positioned against closely related methods. MVGD learns novel-view image and depth synthesis within a multitask diffusion framework, while IDCNet jointly generates RGB-D sequences under camera-trajectory control. Neither is discussed in the manuscript. Although their input requirements and architectures differ, these precedents make it important to explain what is new in the proposed joint modeling beyond its integration with camera understanding and gravity-aware conditioning.

    2. Section 2.2 identifies the camera-conditioning design as a primary departure from Puffin, but its benefit is not isolated under a matched backbone and training dataset. Table A5 provides useful evidence for physics propagation, including a PSNR improvement from 19.87 to 20.54, but does not separately evaluate the new injection mechanism. A matched comparison on the shared single-view generation task would help establish what this design adds beyond the expanded training recipe.

  2. The quantitative evaluation does not fully support the geometry and camera-control claims.

    1. Table 3 reports image fidelity and camera-control metrics, but no quantitative depth accuracy, reconstruction accuracy, or cross-view geometric consistency. Figures 3 and A14 provide substantial qualitative evidence, yet do not establish how accurately the generated geometry agrees across views. Comparing the native depth output with a frozen depth estimator applied to the same generated RGB images would also clarify the benefit of predicting geometry within the generator.

    2. Sections 3.3 and 3.4, as well as Appendix A0.5, assess camera control using Puffin-World’s own understanding branch. The common evaluator makes the protocol consistent, and Table 1 demonstrates strong camera-estimation performance on benchmark images, but neither establishes that its errors are comparable across different generators. An independent check is needed to determine whether the camera-control rankings are robust to the choice of evaluator.

Questions for authors
  1. Methodological contribution

    1. What specifically distinguishes the joint appearance–geometry design in Section 2.3 from MVGD and IDCNet, accounting for their different input requirements and architectures?

    2. Could you provide a comparison with the original Puffin camera-conditioning design on single-view generation, using the same backbone and training data? This would help isolate the benefit of the proposed conditioning and injection mechanism.

  2. Geometry accuracy

    1. Could you report depth accuracy and cross-view geometric consistency on held-out scenes with independent geometric ground truth, including trajectories with translation?

    2. On the same evaluation set, how does the native depth output compare with a frozen depth estimator applied to the same generated RGB images?

  3. Independent camera evaluation

    Do the camera-control rankings in Tables 2 and 3 remain consistent when evaluated with an independent camera estimator?

  4. Experimental configuration

    1. Which model variants, checkpoints, test subsets, and aggregation procedures were used in Tables 3 and A5? Both use Puffin-Traj-Bench, but Table 3 reports a PSNR of 18.00, whereas the configuration with physics propagation in Table A5 reports 20.54.

    2. How was the baseline without physics propagation in Table A5 constructed? Was it trained without the propagated gravity condition, or was that condition removed only at inference? Please clarify which other settings were held fixed.

7 comments
AuthorsOpenReview ↗

Response to Reviewer GTpr (part 1/6)

W1.1. Positioning joint appearance–geometry modeling against MVGD and IDCNet.

Thank you for mentioning the relevant MVGD [1] and IDCNet [2]. Compared with these works, our contribution concerns how geometry is represented, coupled, and trained within a pretrained unified model. Please refer to the following table for more details, which compares these mechanisms while accounting for the different architectures and input requirements.

Table R1-1. Mechanism-level comparison of joint appearance–geometry generation.

AspectMVGD [1]IDCNet [2]Puffin-World
Geometry integrationA separate prediction task selected by task embeddingsRGB/depth latents concatenated along the channel axisSeparate RGB/depth token blocks paired by view index and camera condition
Cross-modal interactionShared scene representation with task-specific outputsJoint RGB/depth processing within each tokenAsymmetric attention: depth attends to appearance/text; reverse attention is blocked
Depth encodingPixel-space log-depth with scene-scale normalizationShared video VAEHilbert-color encoding through the frozen image VAE; no dedicated depth encoder/decoder
Backbone and adaptationPixel-space RIN diffusion trained from scratchCogVideoX-5B-I2V; RGB-to-RGBD pretraining followed by camera-conditioned joint RGB-D generationPretrained LLM–SD3.5 model extended with a geometry embedding, modality-normalized losses, and depth-loss warmup
Inputs and camera groundingPosed input views and target camera; relative ray encodingInitial RGB-D and trajectory in its stated joint formulation; Plücker conditioningImage/text conditioning and camera controls; reference-image gravity/FoV estimated internally and combined with relative rays

We further highlight three properties that distinguish our integration:

  1. Geometry reads the appearance context through asymmetric attention. Depth tokens attend to appearance and text, while RGB and text tokens do not directly attend to depth. This structure is intended to limit interference when adding a new modality (i.e., geometry) to a converged image generator, guaranteeing RGB quality with a minimum training cost. We compare Stage-III (only appearance as output) and Stage-IV (both appearance and depth as output) PSNR/LPIPS under the same Table 3 protocol: Stage-III: 17.25/0.324 vs. Stage-IV: 17.22/0.318 on the RealEstate10K test set. Our method ensures consistent generation performance while effectively enabling the geometry extension.

  2. Depth reuses the frozen image VAE without a dedicated autoencoder. Our Hilbert-color mapping converts unbounded depth into a bounded three-channel representation. On the reported DL3DV diagnostic, Hilbert encoding reduces the median relative error from 7.79% to 1.18% and increases the fraction of pixels within 5% error from 33.4% to 95.9%, compared with grayscale encoding using the same depth nonlinearity. These results demonstrate that the encoding preserves depth information effectively through the existing frozen VAE, without requiring a dedicated depth encoder or decoder.

  3. Geometry is coupled to native physical-camera understanding. Given a reference image, Puffin-World internally estimates gravity and FoV, propagates the physical camera state along the commanded trajectory, and jointly generates RGB and depth for the target views. This process requires neither an input depth map nor an external calibration module, connecting geometry generation directly to the model’s own physical-camera understanding.

As Reviewer HK4P also noted, “the proposed geometry–appearance integration is efficient: Hilbert depth encoding reuses the frozen VAE, while asymmetric attention and depth-loss warmup preserve RGB quality during training.” In response to your suggestions, we have clarified these contributions and their distinctions from MVGD and IDCNet in Section 2.3 of the revised manuscript.

[1] Guizilini, Vitor, et al. "Zero-shot novel view and depth synthesis with multi-view geometric diffusion." IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025.

[2] Liu, Lijuan, et al. "Idcnet: Guided video diffusion for metric-consistent rgbd scene generation with precise camera control." arXiv preprint arXiv:2508.04147, 2025.

AuthorsOpenReview ↗

Response to Reviewer GTpr (part 2/6)

W1.2. Isolating the new camera-conditioning mechanism from the backbone and training recipe.

We appreciate your insightful comment. The primary goal of the new camera interface is scalability across tasks, rather than an isolated improvement in single-view image quality. Following your suggestion, we provide a detailed comparison using a shared setting in Table R1-2:

Table R1-2. Single-view quality and inference cost.

Metric or capabilityPuffin-style injectionPuffin-World fusion
FID ↓78.1275.93
Inference time ↓44.5 s19.5 s
Compute per image ↓7.0 PFLOP2.9 PFLOP
New token introduced✓✗
Camera-controllable text-to-image✓✓
Joint multi-view image-to-3D / text-to-3D✗✓
Native RGB-D generation✗✓

We further highlight three points that distinguish our design:

  1. Comparable quality with lower cost. Using the same SD3.5-Large MMDiT backbone and training data, the two interfaces achieve similar FIDs, while our latent-space fusion provides approximately 2.3× faster inference on an A800-80GB GPU and 2.4× lower computation.

  2. Richer conditioning without additional camera tokens. Puffin’s three-channel Perspective Field targets absolute single-view control, and each VAE-encoded camera map introduces a separate token stream. Our interface directly fuses the nine-channel Omni-Camera map and role mask into existing latent positions, accommodating relative motion and view roles without extra camera tokens.

  3. The scalability advantage. Our design provides a shared conditioning formulation for text-to-image, image-to-3D, and text-to-3D generation. By combining absolute camera grounding, relative motion, and view-role information, it extends camera control from single-view synthesis to joint multi-view RGB-D generation, allowing one unified model to support different input modalities and generation tasks.

AuthorsOpenReview ↗

Response to Reviewer GTpr (part 3/6)

W2.1. Quantitative evaluation on geometry accuracy and cross-view consistency.

Thank you for your valuable suggestion. We have added quantitative geometry evaluations across four datasets. To be specific, we use held-out scenes with independent geometric references: DL3DV [3], 7Scenes [4], NRGBD [5], and ETH3D [6]. Each clip contains real references and generated views along a ground-truth trajectory with translation, with camera motion matched across datasets. Frozen DUSt3R [7], CUT3R [8], and VGGT [9] reconstruct the same sequence for comparison with our native geometry.

Please note that Puffin-World jointly generates appearance and geometry, whereas the reconstruction baselines recover geometry from provided RGB observations. For regions unseen in the reference, plausible generated content may differ from the particular ground-truth scene, making full-view depth errors insufficient to characterize generative geometry. We therefore combine accuracy evaluation on observed/co-visible regions with cross-view consistency over the generated sequence. The quantitative evaluation results are listed in Table R1-3.

Table R1-3. Native geometry versus post-hoc reconstruction. Reference-view AbsRel is scale-aligned. Co-visible δ1 measures the fraction of valid co-visible pixels whose predicted depth is within a factor of 1.25 of ground truth, using a single reference-fitted scale across all views. Depth consistency measures reprojected points within 5% / 10% depth error. Point cloud consistency measures the fraction of points confirmed by ≥2 other views within 2 px and 5% depth error.

DatasetMethodRef.-view AbsRel ↓Co-visible δ1 ↑Consistency (depth) <5% / <10% ↑Consistency (point cloud) ↑
DL3DVDUSt3R0.0780.89668% / 83%71.8%
CUT3R0.0340.95771% / 83%75.9%
VGGT0.0560.92779% / 85%80.3%
Puffin-World0.0720.93172% / 86%76.6%
7ScenesDUSt3R0.0400.97479% / 92%79.8%
CUT3R0.0420.96687% / 95%85.2%
VGGT0.0520.96690% / 95%89.2%
Puffin-World0.0520.97689% / 94%87.4%
NRGBDDUSt3R0.0270.96283% / 92%81.2%
CUT3R0.0260.96379% / 90%73.3%
VGGT0.0230.96089% / 93%80.8%
Puffin-World0.0440.96184% / 92%79.5%
ETH3DDUSt3R0.0280.87351% / 67%9.7%
CUT3R0.0290.86940% / 64%11.3%
VGGT0.0330.86955% / 75%15.9%
Puffin-World0.0580.89651% / 72%18.4%

We further provide some analyses in terms of the above evaluations.

  1. Competitive consistency with remaining precision gaps. On DL3DV, 7Scenes, and NRGBD, our native geometry achieves the best depth consistency at the 10% threshold on DL3DV and remains within one percentage point of the best on the other two datasets. At the stricter 5% threshold, it exceeds DUSt3R and CUT3R across all three datasets, although VGGT remains stronger. Co-visible δ1 is highest on 7Scenes and within 0.03 of the best on DL3DV and NRGBD. On wide-baseline ETH3D, Puffin-World achieves the highest co-visible δ1 and point-cloud consistency. These results support coherent geometry across translated viewpoints, while reference-view accuracy and fine-scale consistency remain areas for improvement.

  2. Training scope for geometry. Our depth supervision comes from only five datasets—DL3DV, Hypersim, MVS-Synth, TartanAir, and ScanNet. In comparison, CUT3R uses 29 depth-supervised dataset sources within its 32-dataset training collection, while VGGT lists 17 training datasets, corresponding to 5.8× and 3.4× our source count, respectively. This narrower range of geometric supervision may contribute to the remaining precision gap relative to dedicated reconstruction models.

  3. Value of unification. Our contribution is to natively unify physical, geometric, and appearance world states, enabling understanding, generation, and reconstruction within one model. Compared with applying a separate feed-forward model such as DA3 or CUT3R to generated RGB views, our approach produces depth alongside appearance in the commanded camera frame, making geometry directly available for multi-view fusion without an additional reconstruction network or inference pass. Depth prediction also accesses the shared multi-view appearance, text, and camera context, explicitly conditioning geometry on the intended scene and trajectory. These benefits complement the competitive cross-view consistency demonstrated above. In the future, we plan to expand the scale and diversity of 3D supervision to improve geometric accuracy and advance toward state-of-the-art performance across understanding, generation, and reconstruction.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

Dear ACs and Reviewers,

We sincerely thank the Area Chairs and reviewers for their time and constructive feedback. We appreciate the reviewers’ recognition of the following strengths:

  • A well-motivated, unified camera representation. Combining absolute gravity grounding with relative camera geometry provides a coherent connection between camera understanding and generation across views (GTpr, HK4P, ioPk).
  • Efficient appearance–geometry integration. Reusing the frozen image VAE, together with asymmetric attention and depth-loss warmup, provides a practical approach to incorporating native geometry (HK4P).
  • Broader camera-motion coverage and potential community value. Puffin-16M addresses underrepresented viewpoints and challenging rotations, while the associated camera annotations offer a valuable resource for broader camera-grounded modeling (GTpr, HK4P, ioPk).

Following your suggestions, we have expanded the evaluations and clarified the scope of our contributions. The main additions and revisions are:

  1. Independent camera-control evaluation. We re-evaluated the existing image-generation, multi-view generation, and physics-propagation outputs using GeoCalib to examine evaluator dependence. We also clarified the prompting protocol for general-purpose image models. (GTpr: W2.2/Q3; HK4P: W2; ioPk: W1/Q1.)
  2. Quantitative geometry validation. We added depth-accuracy and cross-view depth/point-cloud consistency evaluations on held-out scenes with independent geometric references and translated trajectories, comparing native geometry with frozen DUSt3R, CUT3R, and VGGT applied to the same RGB sequences. We also explained the point-cloud construction procedure and the evaluation scope for generative geometry. (GTpr: W2.1/Q2; HK4P: W1/Q1/Q3.)
  3. Camera-interface comparison and claim clarification. We compared Puffin-style conditioning with our fusion interface under a shared backbone and training dataset, reporting image quality, camera accuracy, and inference cost. We clarified the distinction between interface-level and system-level comparisons, emphasizing efficiency and task scalability. (GTpr: W1.2/Q1.2; ioPk: W2/Q3.)
  4. Geometry-component analysis and RGB preservation. We added a frozen-VAE depth-encoding diagnostic, an asymmetric-attention ablation, and a Stage III–IV comparison of RGB generation quality, and clarified the optimization role of depth-loss warmup. (HK4P: W1/Q1.)
  5. Positioning against related work. We added a mechanism-level comparison with MVGD and IDCNet and clarified the contribution and benefits of integrating native geometry into a pretrained unified model. (GTpr: W1.1/Q1.1; HK4P: Q3.)
  6. Experimental configuration and reproducibility. We specified the model variants used in each table, added Pro calibration results for comparison with Base, and clarified the evaluation subsets, no-propagation baseline, conditioning augmentation, and settings held fixed in the propagation ablation. (GTpr: Q4; HK4P: Q2; ioPk: W2/Q2.)

To support the research community and facilitate future work, we will make all code, models, and datasets publicly available. Please refer to the reviewer-specific, point-by-point responses below for the detailed protocols, results, and discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.