Solar Altitude Guided Scene Illumination

Alfred Sch\"ottl; Maximilian Hoh; Nicolas Rodriguez Pe\~na; Nico Leuze; Samed Do\u{g}an

REVIEW 4 major objections 6 minor 24 references

Solar Altitude Guided Scene Illumination

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that solar altitude, a scalar computed from ego-pose and capture time, is a sufficient global conditioning variable for diffusion-based camera image generation, capturing both lighting and illumination-dependent noise…

desk verdict Solar altitude is a neat, label-free conditioning idea, but Table I's 64-sample evidence is too thin to substantiate the 'accurate capture' claim; still worth a serious referee. read the letter →

arxiv 2507.05812 v2 pith:V5AJKDTG submitted 2025-07-08 cs.CV cs.AI

Samed Doğan , Maximilian Hoh , Nico Leuze , Nicolas Rodriguez Peña , Alfred Schöttl This is my paper

classification cs.CVcs.AI

keywords solaraltitudesceneilluminationdiffusionmodelstextualinversioncamerasensordataautonomousdrivingimagenoisenuScenes

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Solar altitude, the angle of the sun above the horizon, is computed from ego-position and local time, and the paper argues that this single scalar carries enough information about global illumination to condition a diffusion model that generates autonomous-driving camera images. The authors identify a gap: datasets like nuScenes only label "Night," so daytime variation is untrained and underreported. They introduce a quantization-plus-residual normalization that allocates extra numerical resolution to twilight, where perceived brightness changes fastest, and wrap it into a dynamically generated text-inversion token. If the claim holds, synthetic camera data can be produced under finely controlled lighting without any manual labeling, which would strengthen simulation-based safety validation.

What carries the argument

The load-bearing mechanism is a scalar normalization called quantization with residual encoding: a solar altitude $a$ is mapped to a bin index $Q(a)$ plus an intra-bin residual $R(a)$, combined as $\hat{a} = (Q(a)+R(a))/K$, with bin edges placed densely near twilight ($[\min,-6,-4,-2,\max]$). This keeps global structure across the full altitude range while preserving fine local variation where daylight changes rapidly. The normalized value is enriched by an RBF encoding layer with non-trainable centers and then modulates static text embeddings through style-adaptive layer normalization, so the resulting context token changes smoothly with altitude. Partial sampling uses the context token in the first half of denoising and the structure token in the second half.

What would settle it

Resample generation from held-out nuScenes frames split by cloud cover or by auto-exposure metadata at matched altitudes: if per-altitude FID or $\Delta\sigma$ varies strongly across subsets, altitude alone fails as a surrogate. Separately, collect real dawn and dusk captures with altitudes between $-6^\circ$ and $-2^\circ$ and compare generated brightness and noise against them, since the paper's own validation set has no twilight samples.

Watch

Extended reading notes

Core claim

The central claim is that a scalar solar altitude, normalized to preserve local variation in perceptually sensitive regimes, is an effective global conditioner for image diffusion in the driving domain. The paper trains a structure token S* on day-only nuScenes frames and a separate dynamic context token D* whose embedding is produced by a small network from the normalized altitude. Generation conditioned on D* and S* reaches an FID of 144.789 at normalized altitude 1.0, whereas the structure-only token reaches 151.433, and the estimated noise rate-of-change $\Delta\sigma$ tracks ground truth (-1.223 versus -1.389). The authors conclude that altitude captures both global lighting characteristics and illumination-dependent sensor noise.

Load-bearing premise

Solar altitude is treated as a sufficient global surrogate for scene illumination and camera noise after dropping rain samples, with clouds, weather, auto-exposure, and gain assumed constant or irrelevant.

Editorial extensions

If this is right

Manually labeled day and night categories become unnecessary: any dataset that logs ego-pose and timestamps can be conditioned on solar altitude directly.
The twilight-focused binning should let generators produce sunrise and sunset scenes even when training data has few such samples, provided validation data covers that range.
Because $\Delta\sigma$ follows ground truth, generated images can be used to test perception systems under illumination-dependent noise without collecting new sensor data.
The method is orthogonal to spatial conditioners such as 3D geometry or bird's-eye-view control, so it can be layered onto existing street-scene generators.
With the proposed $2^\circ$ bin scheme starting at nautical twilight, outdoor datasets beyond driving could adopt the same scalar conditioning.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: since the validation set contains no twilight frames, the claimed twilight sensitivity is a plausible extrapolation rather than a measured result; a dedicated dawn and dusk validation set would test it directly.
Editorial inference: the constant-weather assumption implies altitude alone will mispredict illumination under heavy cloud cover; conditioning on cloud fraction or weather text would be a natural extension and would reveal how much variance altitude actually explains.
Editorial inference: because the noise-rate metric is computed with wavelet shrinkage, matching $\Delta\sigma$ suggests altitude could serve as a proxy for sensor-gain regimes, which might transfer to image-denoising or robustness benchmarks.
Editorial inference: the method's portability to other datasets depends only on pose and time metadata, so a cross-city or cross-camera test would show whether the learned embedding generalizes beyond nuScenes.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: the solar-altitude conditioning idea is a legitimate small contribution, but the experiments don't yet support the "accurate capture" claim. The paper is worth engaging with, not because it's polished, but because the core idea is new and cheap to test.

What's actually new: using solar altitude as a global conditioning variable for camera generation in driving, computed from pose and time without manual labels. I haven't seen that in the cited literature. The residual quantization with bin edges tuned to twilight is a sensible response to the nonlinearity of perceived light, and the SALN/RBF encoder is a reasonable way to make the token continuous. The delta-sigma noise check is a good instinct: it tries to anchor generation to an external physical measure rather than just eyeballing images.

The soft spots are mostly in the evidence. Table I rests on 64 generated images per condition with no error bars and no repeated seeds; FID at 122.9 vs 144.8 can easily move by that amount with sampling noise. The non-monotonic trend across altitude is never explained. The noise comparison comes down to three single-point delta-sigma values, which is suggestive but not statistically meaningful. And the paper states the validation set lacks twilight samples, so the exact regime the binning was designed for is never tested. On top of that, solar altitude is a weak surrogate when clouds, auto-exposure, and gain vary freely; the authors only filter rain.

One definite error: reference [11] is "Early reading assessment", which has nothing to do with pseudo-word tokens. That should be corrected.

None of this kills the idea. The conditioning variable is physically grounded and the implementation is lightweight. But the abstract's "accurately capture lighting characteristics and illumination-dependent image noise" is ahead of the data. With error bars, a proper baseline (unconditional or night/day condition), and an evaluation that matches generated and real images at the same altitude while controlling for weather, the claim would be testable.

Who this is for: people building synthetic camera data pipelines for driving, and anyone thinking about continuous illumination conditioning in diffusion models. It's a workshop-to-conference level paper that deserves a real referee: the idea is worth putting into the literature, but the current evidence is a proof-of-concept, not a demonstration.

Recommendation: accept for peer review with a request for major revision. The reviewer should ask for more samples, error bars, baselines, and a twilight-aware evaluation.

Referee Report

4 major / 6 minor

Summary. The paper introduces solar altitude as a global conditioning variable for diffusion-based generation of driving camera images, motivated by the lack of fine-grained daytime labels in datasets such as nuScenes. The method encodes solar altitude through a nonlinear scalar quantization with residual encoding, designed to emphasize perceptually sensitive twilight intervals, and injects the condition through a learned context token used during partial sampling. The authors report FID scores and wavelet-based noise-rate estimates (Δσ) in Table I, claiming that the conditioned model captures both global illumination and illumination-dependent sensor noise.

Significance. If the central claim were established, the idea would be a useful, label-free conditioning mechanism for synthetic camera data generation in autonomous driving, orthogonal to existing geometry- and layout-based controls. The solar-altitude prior is cheap to compute from ego-pose and time, and the proposed normalization is an intuitive response to the nonlinearity of perceived illumination near twilight. The paper also provides a falsifiable quantitative anchor by comparing induced noise-rate changes against ground-truth estimates. However, the empirical support is currently too thin to substantiate the abstract's claim of 'accurately capture' illumination and noise, and several load-bearing evaluation choices are either unvalidated or confounded.

major comments (4)

[Section IV-B, Table I] The quantitative evaluation rests on 64 generated images per condition (Section IV-A) with no repeated seeds, bootstrap intervals, or error bars. At this sample size, FID has substantial variance, and the reported trend across normalized altitudes (145.312, 122.933, 135.778, 144.789) is non-monotonic: the model improves from 0.0 to 0.33 but then worsens through 0.66 and 1.0. The paper does not explain this behavior. The Δσ rows are single-point estimates of a wavelet-based noise measure, so the reader cannot assess whether the agreement with ground truth (e.g., -1.223 versus -1.389) is statistically meaningful. These limitations directly undermine the central claim that the model 'accurately captures lighting characteristics and illumination-dependent image noise.'
[Section III-A and IV-B] The proposed normalization is specifically designed to emphasize twilight via the bin edges [amin, -6, -4, -2, amax], yet the validation set contains no twilight samples, as the paper itself states ('The validation set does not contain samples captured during twilight conditions, rendering direct comparison infeasible'). Consequently, the very regime that motivates the tailored normalization is untested quantitatively; Fig. 2 provides only qualitative examples and cannot substitute for a direct test. The claim of twilight sensitivity is therefore not supported by the presented evidence.
[Section III-B and Table I] The comparison between the structure-only token (S*) and the conditioned token (D* & S*) in Table I is confounded. The structure token s* is trained on the mini subset and explicitly excludes night scenes, while the context token fθ is trained on the full dataset (Eq. 4). Additionally, the inference protocol differs: D* & S* uses partial sampling with the context token in the first half and the structure token in the second half, whereas the S*-only baseline appears to use the structure token alone. Thus the FID improvement at altitude 1.0 (151.433 versus 144.789) cannot be attributed solely to solar-altitude conditioning.
[Section III-A] The paper assumes 'constant weather' when using solar altitude as a surrogate for illumination and noise, but the dataset filtering only removes rain-labeled samples. Clouds, overcast conditions, and camera auto-exposure or gain variations are not controlled, so the same solar altitude can correspond to very different illumination levels in nuScenes. The Δσ agreement with ground truth could therefore arise from other covariates correlated with altitude rather than from a causal lighting effect. The authors should analyze the residual illumination variance within altitude bins or condition on additional metadata to support the surrogate claim.

minor comments (6)

[References] Reference [11] cites 'Rathvon, Early reading assessment', which is unrelated to the textual inversion pseudo-word 'V*' notation; the intended source appears to be the textual inversion paper [9] or a related follow-up.
[Fig. 1 and Section III-B] The dynamic embedding pipeline in Fig. 1 is not fully described in the text; the roles of the RBF centers, the trainable gammas, and the SALN modulation in producing the context token d* should be specified in enough detail to be reproducible.
[Section IV-B] The quantity Δσ is not formally defined. The paper should state how the wavelet-based noise estimate is computed, on which image regions, and how many ground-truth and generated images are used for each reported value.
[Section III-A] The resampling procedure is described only as 'resampling with replacement, such that the bin count across all bins is approximately uniform.' Please specify whether this applies to the context-token training set, the structure-token training set, or both, and report the resulting per-bin sample counts.
[Fig. 2 and Table I] The x-axis of Fig. 2 and the rows of Table I use normalized altitude values (0.0, 0.33, 0.66, 1.0). Reporting the corresponding solar altitude values in degrees would make the qualitative and quantitative comparisons interpretable for readers.
[Section IV-A] The selection of the 'best structure embedding' via the dual-score approach is not reported; please state the chosen learning rate, number of embeddings, and the resulting FID/BLIP scores for the selected configuration.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: solar altitude is externally computed, and the noise evaluation uses an independent ground-truth estimator.

full rationale

The paper's derivation chain is self-contained and does not reduce to its inputs. The conditioning variable (solar altitude) is computed from ego-pose and capture time via the astropy library, not from the target illumination or noise statistics; Eq. (2) defines a quantile/residual normalization whose bin edges are a modeling choice, not fitted to the evaluation metric. The context token is optimized with the standard LDM objective (Eq. 4) over the full dataset, and the evaluation separately compares generated-image noise deltas (Δσ) against ground-truth deltas computed with an external wavelet estimator [24]. No fitted parameter is renamed as a prediction, and the paper contains no self-citations that carry load-bearing weight. The absence of twilight samples in the validation set and the small sample size (64 images per condition, no error bars) limit the strength of the empirical claims, but these are correctness and generalization concerns, not circularity. Consequently the appropriate circularity score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the meteorological assumption that solar altitude alone predicts illumination and noise, plus a hand-defined quantization. These are reasonable starting points but are not validated with controls or a twilight split. No invented physical entities are proposed.

free parameters (2)

solar altitude bin edges = [amin, -6, -4, -2, amax] degrees
Hand-chosen thresholds for twilight stages; central to the quantization in Equation (2) and to the claimed sensitivity around sunrise and sunset. No justification or sensitivity analysis is provided.
recommended 2 degree bin step = 2 degrees from -12 to +6 degrees
Suggested for future use as a design choice; not derived from data or validated.

assumptions (3)

domain assumption Solar altitude is a valuable proxy for expected lighting characteristics under constant weather
Stated in Section III-A; the paper filters only rain labels and otherwise assumes weather and camera exposure are constant over the dataset.
ad hoc to paper Perceived-light nonlinearity is concentrated in twilight and is adequately captured by the specified bin edges
Introduced in Section III-A and Section IV-A; no direct validation on twilight samples because the validation set lacks them.
domain assumption A single learned embedding token plus a small network can represent global illumination within the Stable Diffusion latent space
Assumed throughout Section III-B and only indirectly checked via FID and delta-sigma.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Solar Altitude Guided Scene Illumination." pith.science (2026). https://pith.science/paper/V5AJKDTG

@misc{pith2026250705812,
  author       = {Pith},
  title        = {Pith review of: Solar Altitude Guided Scene Illumination},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/V5AJKDTG}},
  note         = {Machine review of arXiv:2507.05812}
}

read the original abstract

The development of safe and robust autonomous driving functions is heavily dependent on large-scale, high-quality sensor data. However, real-world data acquisition requires extensive human labor and is strongly limited by factors such as labeling cost, driver safety protocols and scenario coverage. Thus, multiple lines of work focus on the conditional generation of synthetic camera sensor data. We identify a significant gap in research regarding daytime variation, presumably caused by the scarcity of available labels. Consequently, we present solar altitude as global conditioning variable. It is readily computable from latitude-longitude coordinates and local time, eliminating the need for manual labeling. Our work is complemented by a tailored normalization approach, targeting the sensitivity of daylight towards small numeric changes in altitude. We demonstrate its ability to accurately capture lighting characteristics and illumination-dependent image noise in the context of diffusion models.

Figures

Figures reproduced from arXiv: 2507.05812 by the authors.

**Figure 1.** Dynamic Embedding Pipeline B. Textual Inversion In general, text-to-image models employ a pre-trained text encoder in order to guide the image generation process towards some desired text prompt. A tokenizer maps the input prompt to unique tokens on a per-word or sub-word basis. The discrete tokens index a corresponding learned embedding from an internal lookup table, forming the input embeddings. These are then fur… view at source ↗

**Figure 2.** Comparison of ground truth samples (top) with generated images (bottom) for varying altitude conditions, normalized across the training bin range. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

24 extracted references · 12 canonical work pages

[11]

Rathvon, Early reading assessment

N. Rathvon, Early reading assessment . New York, NY: Guilford Publications, Jun. 2004

work page 2004
[1]

Denoising diffusion probabilistic models,

J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems , vol. 33, pp. 6840– 6851, 2020

2020
[2]

Deep unsupervised learning using nonequilibrium thermodynamics,

J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” in Proceedings of the 32nd International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, F. Bach and D. Blei, Eds., vol. 37. Lille, France: PMLR, 07–09 Jul 2015, pp. 2256–2265. [Online]. Availabl...

work page 2015
[3]

Magicdrive: Street view generation with diverse 3d geometry control,

R. Gao, K. Chen, E. Xie, H. Lanqing, Z. Li, D.-Y . Yeung, and Q. Xu, “Magicdrive: Street view generation with diverse 3d geometry control,” in The Twelfth International Conference on Learning Representations , 2023

work page 2023
[4]

Street-view image generation from a bird’s-eye view layout,

A. Swerdlow, R. Xu, and B. Zhou, “Street-view image generation from a bird’s-eye view layout,” IEEE Robotics and Automation Letters , 2024

2024
[5]

Holodrive: Holistic 2d-3d multi-modal street scene generation for autonomous driving,

Z. Wu, J. Ni, X. Wang, Y . Guo, R. Chen, L. Lu, J. Dai, and Y . Xiong, “Holodrive: Holistic 2d-3d multi-modal street scene generation for autonomous driving,” arXiv preprint arXiv:2412.01407 , 2024

arXiv 2024
[6]

X-drive: Cross-modality consistent multi-sensor data synthesis for driving scenarios,

Y . Xie, C. Xu, C. Peng, S. Zhao, N. Ho, A. T. Pham, M. Ding, M. Tomizuka, and W. Zhan, “X-drive: Cross-modality consistent multi-sensor data synthesis for driving scenarios,” arXiv preprint arXiv:2411.01123, 2024

arXiv 2024
[7]

Drivedreamer: Towards real-world-drive world models for autonomous driving,

X. Wang, Z. Zhu, G. Huang, X. Chen, J. Zhu, and J. Lu, “Drivedreamer: Towards real-world-drive world models for autonomous driving,” in European Conference on Computer Vision . Springer, 2024, pp. 55– 72

work page 2024

Show all 24 references

[8]

nuscenes: A multimodal dataset for autonomous driving,

H. Caesar, V . Bankiti, A. H. Lang, S. V ora, V . E. Liong, Q. Xu, A. Kr- ishnan, Y . Pan, G. Baldan, and O. Beijbom, “nuscenes: A multimodal dataset for autonomous driving,” in CVPR, 2020

2020
[9]

An image is worth one word: Personalizing text-to-image generation using textual inversion,

R. Gal, Y . Alaluf, Y . Atzmon, O. Patashnik, A. H. Bermano, G. Chechik, and D. Cohen-or, “An image is worth one word: Personalizing text-to-image generation using textual inversion,” in The Eleventh International Conference on Learning Representations , 2023. [Online]. Availa...

2023
[10]

High- resolution image synthesis with latent diffusion models,

R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer, “High- resolution image synthesis with latent diffusion models,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2022, pp. 10 684–10 695

2022
[12]

The Astropy Project: Sustaining and Growing a Community- oriented Open-source Project and the Latest Major Release (v5.0) of the Core Package,

Astropy Collaboration, A. M. Price-Whelan, P. L. Lim, N. Earl, N. Stark- man, L. Bradley, D. L. Shupe, A. A. Patil, L. Corrales, C. E. Brasseur et al. , “The Astropy Project: Sustaining and Growing a Community- oriented Open-source Project and the Latest Major Release (v5.0) o...

2022
[13]

Attention is all you need,

A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems , vol. 30, 2017

2017
[14]

Fourier features let networks learn high frequency functions in low dimensional domains,

M. Tancik, P. Srinivasan, B. Mildenhall, S. Fridovich-Keil, N. Raghavan, U. Singhal, R. Ramamoorthi, J. Barron, and R. Ng, “Fourier features let networks learn high frequency functions in low dimensional domains,” Advances in neural information processing systems , vol. 33, pp...

2020
[15]

T2i- adapter: Learning adapters to dig out more controllable ability for text- to-image diffusion models,

C. Mou, X. Wang, L. Xie, Y . Wu, J. Zhang, Z. Qi, and Y . Shan, “T2i- adapter: Learning adapters to dig out more controllable ability for text- to-image diffusion models,” in Proceedings of the AAAI conference on artificial intelligence, vol. 38, no. 5, 2024, pp. 4296–4304

2024
[16]

Dreamstyler: Paint by style inversion with text-to-image diffusion mod- els,

N. Ahn, J. Lee, C. Lee, K. Kim, D. Kim, S.-H. Nam, and K. Hong, “Dreamstyler: Paint by style inversion with text-to-image diffusion mod- els,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 38, no. 2, 2024, pp. 674–681

2024
[17]

Learning transferable visual models from natural language supervision,

A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark et al., “Learning transferable visual models from natural language supervision,” in International conference on machine learning . PmLR, 2021, pp. 8748–8763

2021
[18]

Networks for approximation and learning,

T. Poggio and F. Girosi, “Networks for approximation and learning,” Proceedings of the IEEE , vol. 78, no. 9, pp. 1481–1497, 1990

1990
[19]

Meta-stylespeech: Multi- speaker adaptive text-to-speech generation,

D. Min, D. B. Lee, E. Yang, and S. J. Hwang, “Meta-stylespeech: Multi- speaker adaptive text-to-speech generation,” in International Conference on Machine Learning . PMLR, 2021, pp. 7748–7759

2021
[20]

Diffusers: State-of-the-art diffusion models,

P. von Platen, S. Patil, A. Lozhkov, P. Cuenca, N. Lambert, K. Rasul, M. Davaadorj, D. Nair, S. Paul, S. Liu, W. Berman, Y . Xu, and T. Wolf, “Diffusers: State-of-the-art diffusion models,” 2022. [Online]. Available: https://github.com/huggingface/diffusers

2022
[21]

Decoupled weight decay regularization,

I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101 , 2017

2017 arXiv
[22]

Denoising diffusion implicit models,

J. Song, C. Meng, and S. Ermon, “Denoising diffusion implicit models,” arXiv preprint arXiv:2010.02502 , 2020

2010 arXiv
[23]

Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation,

J. Li, D. Li, C. Xiong, and S. Hoi, “Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation,” in International conference on machine learning . PMLR, 2022, pp. 12 888–12 900

2022
[24]

Ideal spatial adaptation by wavelet shrinkage,

D. L. Donoho and I. M. Johnstone, “Ideal spatial adaptation by wavelet shrinkage,” Biometrika, vol. 81, no. 3, pp. 425–455, 09 1994. [Online]. Available: https://doi.org/10.1093/biomet/81.3.425

1994 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [11]

Rathvon, Early reading assessment

N. Rathvon, Early reading assessment . New York, NY: Guilford Publications, Jun. 2004

work page 2004

[2] [1]

Denoising diffusion probabilistic models,

J. Ho, A. Jain, and P. Abbeel, “Denoising diffusion probabilistic models,” Advances in neural information processing systems , vol. 33, pp. 6840– 6851, 2020

2020

[3] [2]

Deep unsupervised learning using nonequilibrium thermodynamics,

J. Sohl-Dickstein, E. Weiss, N. Maheswaranathan, and S. Ganguli, “Deep unsupervised learning using nonequilibrium thermodynamics,” in Proceedings of the 32nd International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, F. Bach and D. Blei, Eds., vol. 37. Lille, France: PMLR, 07–09 Jul 2015, pp. 2256–2265. [Online]. Availabl...

work page 2015

[4] [3]

Magicdrive: Street view generation with diverse 3d geometry control,

R. Gao, K. Chen, E. Xie, H. Lanqing, Z. Li, D.-Y . Yeung, and Q. Xu, “Magicdrive: Street view generation with diverse 3d geometry control,” in The Twelfth International Conference on Learning Representations , 2023

work page 2023

[5] [4]

Street-view image generation from a bird’s-eye view layout,

A. Swerdlow, R. Xu, and B. Zhou, “Street-view image generation from a bird’s-eye view layout,” IEEE Robotics and Automation Letters , 2024

2024

[6] [5]

Holodrive: Holistic 2d-3d multi-modal street scene generation for autonomous driving,

Z. Wu, J. Ni, X. Wang, Y . Guo, R. Chen, L. Lu, J. Dai, and Y . Xiong, “Holodrive: Holistic 2d-3d multi-modal street scene generation for autonomous driving,” arXiv preprint arXiv:2412.01407 , 2024

arXiv 2024

[7] [6]

X-drive: Cross-modality consistent multi-sensor data synthesis for driving scenarios,

Y . Xie, C. Xu, C. Peng, S. Zhao, N. Ho, A. T. Pham, M. Ding, M. Tomizuka, and W. Zhan, “X-drive: Cross-modality consistent multi-sensor data synthesis for driving scenarios,” arXiv preprint arXiv:2411.01123, 2024

arXiv 2024

[8] [7]

Drivedreamer: Towards real-world-drive world models for autonomous driving,

X. Wang, Z. Zhu, G. Huang, X. Chen, J. Zhu, and J. Lu, “Drivedreamer: Towards real-world-drive world models for autonomous driving,” in European Conference on Computer Vision . Springer, 2024, pp. 55– 72

work page 2024