Pith. sign in

REVIEW 3 major objections 6 minor 57 references

Text2Stereo: Repurposing Stable Diffusion for Stereo Generation with Consistency Rewards

T0 review · 3 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Fine-tuning Stable Diffusion with a consistency reward yields wide-baseline stereo image pairs from text prompts alone.

desk verdict Useful engineering contribution with a self-confirming stereo metric; worth review, but the headline number needs an independent check. read the letter →

arxiv 2506.05367 v2 pith:IEI6GS5F submitted 2025-05-27 cs.CV

Aakash Garg , Libing Zeng , Andrii Tsarov , Nima Khademi Kalantari This is my paper
classification cs.CV
keywords text-to-stereogenerationStableDiffusionLoRAfine-tuningAlignProprewardstereoconsistencyPearsoncorrelationwide-baseline
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

The paper proposes Text2Stereo, a way to turn a standard text-to-image diffusion model into a generator of stereo image pairs from a single text prompt. The authors claim that by first fine-tuning Stable Diffusion with LoRA on vertically stacked left-right pairs and then applying AlignProp with a reward that measures agreement between monocular and binocular disparity, the model learns to produce wide-baseline stereo images that are both geometrically consistent and prompt-aligned. This matters because existing stereo-from-prompt methods warp one image using estimated depth and inpaint the missing regions, which causes distorted parallax and wrong occlusions at wide baselines. If the claim holds, VR headsets and light field displays could receive diverse, wide-baseline content directly from prompts, without hand-made 3D assets.

What carries the argument

The load-bearing mechanism is the stereo consistency reward $R_s$, computed as the Pearson correlation between a monocular disparity map and a stereo disparity map: $$R_s=\frac{\sum_p(d_m(p)-\bar d_m)(d_s(p)-\bar d_s)}{\sqrt{\sum_p(d_m(p)-\bar d_m)^2}\sqrt{\sum_p(d_s(p)-\bar d_s)^2}},$$ where $d_m$ comes from DepthAnythingV2 and $d_s$ is taken as the x-coordinate of SEA-RAFT optical flow. This reward, combined with the HPSv2 prompt reward and a negative-disparity penalty, is backpropagated through a randomly truncated denoising trajectory by AlignProp. The vertical stacking of the left and right frames is what lets the single-image Stable Diffusion architecture be reused as a stereo generator.

What would settle it

Render a synthetic wide-baseline stereo scene with known ground-truth disparity, generate many pairs with the trained model, and compare the reward value against true disparity error; if pairs with reward near 1 have large disparity error, or if the reward stays high when the right image is replaced by an unrelated image whose monocular depth structure happens to correlate, the consistency claim would fail.

Watch

Extended reading notes

Core claim

The paper's central claim is that a pretrained latent diffusion model can be repurposed for stereo generation without training from scratch. The authors stack left and right images vertically into one 512x512 RGB image, fine-tune Stable Diffusion v1.5 with LoRA (rank 4) on 30,982 stereo pairs from NeRFStereo and a new MVStereo dataset rendered from 3D Gaussian splatting reconstructions of MVImgNet scenes, then further optimize the model with AlignProp against the reward $R = 0.25R_s + 0.75R_p + 0.25R_c$, where $R_s$ is the Pearson correlation between DepthAnythingV2 monocular disparity and SEA-RAFT optical-flow x-disparity, $R_p$ is HPSv2 prompt alignment, and $R_c$ penalizes negative disparities. In the final configuration the paper reports the stereo score rising from 0.414 to 0.949 and the prompt score from 0.237 to 0.264. The paper further claims that the generated right images are more consistent than those produced by StereoDiffusion or 3D Photography when those methods are given the same left image.

Load-bearing premise

The method depends on the assumption that the correlation between a depth estimate from one image and the left-right pixel shift between the two images is a valid measure of true stereo consistency, and that this same correlation can be used to grade the results.

Editorial extensions

If this is right

  • A text prompt alone can produce a wide-baseline stereo pair, eliminating the depth-warping and inpainting step used by StereoDiffusion and 3D Photography.
  • The generated pairs are consistent enough that Splatt3R can reconstruct a 3D Gaussian representation from them, with clear object boundaries when rendered from a novel view.
  • Optimizing with only the stereo reward improves geometry but hurts prompt alignment; the full three-term reward is needed for both, and using 750 training prompts gives the best measured trade-off.
  • The improvement transfers to held-out test prompts outside the training distribution, suggesting the consistency tuning does not simply memorize the training set.
  • The method does not yet control the stereo baseline or accept a single image as input, leaving those capabilities for future work.

Reading between the lines

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

  • Inference: Because the reward and the reported evaluation metric are the same Pearson correlation, part of the numerical win may reflect the reward optimizing its own measure; an independent consistency check, such as pixel-wise disparity error on synthetic ground truth, would be a stronger test.
  • Inference: The same stacked-pair plus reward-tuning recipe might extend to other binocular or multiview tasks, where agreement between monocular depth and cross-view flow could serve as a generic consistency reward.
  • Inference: A testable extension is conditioning the model on a desired baseline, for example by encoding the baseline into the prompt or cross-attention, which the paper identifies as future work.
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.

Referee Report

3 major / 6 minor

Summary. The manuscript proposes Text2Stereo, a method that adapts Stable Diffusion to generate vertically stacked left-right stereo pairs from text prompts. A first stage fine-tunes SD with LoRA on a combined dataset of NeRFStereo and newly rendered MVStereo pairs. A second stage, based on AlignProp, optimizes a reward R = alpha*Rs + beta*Rp + gamma*Rc, where Rs is the Pearson correlation between a monocular disparity map and the x-component of optical flow, Rp is the HPSv2 prompt alignment score, and Rc penalizes negative disparities. The paper reports improved stereo and prompt scores during optimization, qualitative comparisons to StereoDiffusion and 3D Photography, and ablations over reward components and the number of training prompts.

Significance. If the central claim is accepted, the two-stage LoRA-plus-AlignProp recipe is a practical way to repurpose a text-to-image model for stereo generation, and the MVStereo dataset is a useful addition. The paper is clearly written, the training pipeline is described in detail, and the ablations are internally consistent. However, the main quantitative stereo metric is the same as the optimized reward, so the evidence for geometric consistency is not yet independent; the comparative claim also lacks quantitative support. With an independent geometric evaluation and quantitative baseline comparisons, the method could be a solid contribution.

major comments (3)
  1. [Sec. 5.3, Table 1; Eq. (4)] The 'Stereo Score' reported in Table 1 and Fig. 8 is the same Pearson correlation Rs defined in Eq. (4) that is maximized as a training reward. Therefore, the increase from 0.414 to 0.949 primarily demonstrates that the AlignProp optimization increases its own objective; it does not by itself establish geometric correctness. The same concern applies to the prompt score, because Rp is the HPSv2 reward being optimized. I ask the authors to add an independent geometric evaluation, such as disparity error against ground-truth stereo data, epipolar or vertical-flow error, or a user study, and to report the same independent metric for all baselines.
  2. [Sec. 4.3, Eq. (4)] The stereo consistency reward rests on the assumption that high Pearson correlation between DepthAnythingV2 monocular disparity and the horizontal component of SEA-RAFT optical flow implies correct stereo geometry. Because Pearson correlation is invariant under monotone transformations, a pair can receive a high score even when disparity magnitudes are wrong or when the images are not rectified. In addition, the two estimators can share systematic biases in textureless or fronto-parallel regions, making a correlated pair appear consistent when it is not. I recommend validating the reward by showing that high Rs correlates with low disparity error on pairs with known ground truth, and by checking vertical disparity as a sanity check for rectification.
  3. [Sec. 5.2] The claim to 'outperform existing methods' is currently supported only by qualitative figures (Figs. 5 and 6); no quantitative comparison to StereoDiffusion or 3D Photography is reported. Because Table 1 is an ablation of the proposed method only, the comparative claim is not substantiated. The authors should either add quantitative metrics for the baselines under the same evaluation protocol or soften the claim accordingly.
minor comments (6)
  1. [Sec. 5.2] The text says 'As shown in Figure [40]', but the intended reference is Figure 6; please correct the citation.
  2. [Sec. 6] The name 'LLaV A' should be 'LLaVA'.
  3. [Table 1] The table caption should explicitly define the Stereo Score and Prompt Score and state that they are the same quantities used in the training reward, since this information is currently implicit in the text.
  4. [Sec. 4.3] The reward weights alpha = 0.25, beta = 0.75, and gamma = 0.25 sum to 1.25; if this is intentional, please state why, and if not, please normalize or explain the choice.
  5. [Sec. 4.1] Please provide details of the stereo camera configuration used to render the MVStereo pairs (baseline, focal length, convergence) and statistics over the combined dataset, since 'wide baseline' is a central premise of the paper.
  6. [Fig. 8] Several panels of Fig. 8 have partially clipped axis labels and some panels lack a legend; please improve the figure readability.

Circularity Check

2 steps flagged · score 6.0 of 10

Same Pearson-correlation score serves as both optimized reward and reported metric, making the quantitative consistency claim self-confirming; only qualitative Splatt3R checks are independent.

  1. self definitional [Sec. 4.3 (Eqs. 3-4); Sec. 5.3 (Table 1)]
    "R = αRs + βRp + γRc (3) where Rs, Rp, and Rc refer to the stereo consistency, prompt consistency, and convergence rewards... α = 0.25... we propose measuring their similarity using Pearson correlation, as follows: Rs = ... (4). Table 1. We quantitatively evaluate the impact of various components of our system in terms of the stereo and prompt consistency scores."

    The 'Stereo Score' reported in Table 1 is the same Rs defined in Eq. 4 and optimized through Eq. 3 with weight 0.25. Thus the ablation improvement (e.g., Base 0.414 to Base + Stereo 0.985) is a direct measurement of how well the optimizer increased its own objective component, not an independent geometric test. The paper defines 'stereo consistency' as this Pearson correlation, so reporting that same correlation as the quantitative evidence of consistency improvement is self-definitional.

  2. fitted input called prediction [Sec. 5.3 (Fig. 8)]
    "We show the rewards averaged over 100 training and test prompts during the optimization process. The mean values of both prompt and stereo consistency rewards increase... These trends suggest an improvement in the model's ability to generate images with enhanced prompt alignment and stereo consistency."

    Fig. 8 plots the same reward functions that AlignProp is explicitly maximizing, namely R from Eq. 3. The increase in these curves is the expected outcome of gradient ascent on R; it is a training curve, not an independent evaluation. Using these reward curves as evidence of improved capability is circular because the quantity used to demonstrate improvement is exactly the quantity being optimized.

full rationale

The paper contains no load-bearing self-citations: the only self-citation, [22], is a related-work reference to light-field synthesis and plays no role in the derivation. The central circularity is in the evaluation. Stereo consistency is introduced in Sec. 4.3 as the Pearson correlation Rs between monocular disparity (DepthAnythingV2) and the x-component of SEA-RAFT optical flow. This Rs is inserted into the AlignProp objective R = 0.25Rs + 0.75Rp + 0.25Rc, and Table 1 then reports a 'Stereo Score' that is the same Rs. The ablation numbers therefore measure how successfully the optimizer increased a component of its own loss, rather than validating geometric correctness against an external standard. Likewise, Fig. 8 plots the reward values themselves and presents their increase as evidence of improvement. This is a self-confirming loop for the core 'consistency enhancement' contribution. There is, however, some independent content: the prompt score uses HPSv2, a pre-trained human-preference model not used as a training target for the prompt term, and the Splatt3R novel-view renderings in Fig. 6 are an external, qualitative consistency check. These independent elements prevent the paper from being wholly circular, but the Splatt3R evidence is qualitative and shown for a single prompt, while the only quantitative table is the self-confirming one. Score 6 reflects partial circularity: the headline quantitative evidence reduces to the optimized reward by construction, while the method itself still has an independent generative pipeline and some external, though limited, validation.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The core training pipeline depends on five domain assumptions about pretrained priors, disparity proxies, rectification, reward optimization, and caption quality. Free parameters include reward weights, LoRA rank, and the number of training prompts. No new physical entities such as particles, forces, dimensions, or conserved quantities are introduced.

free parameters (3)
  • Reward weights alpha, beta, gamma = 0.25, 0.75, 0.25
    Hand-chosen weights in Eq. 3 balancing stereo, prompt, and convergence rewards; no sensitivity analysis is provided.
  • LoRA rank = 4
    LoRA rank in Sec. 5.1 chosen without reported sensitivity study; it controls adapter capacity and overfitting.
  • Training prompt count = 750
    Final model uses 750 prompts selected via ablation in Table 1; scores plateau after 750, so the choice is data-dependent.
assumptions (5)
  • domain assumption Stable Diffusion v1.5 prior can be adapted to stereo generation by LoRA fine-tuning on vertically stacked left-right images.
    In Sec. 4.2 the entire adaptation strategy assumes the pretrained prior retains generalization while learning stereo layout; if false, the model would overfit the small dataset.
  • domain assumption Pearson correlation between monocular disparity and stereo disparity is a valid measure of geometric stereo consistency.
    In Sec. 4.3 Eq. 4 this correlation is both the training reward and the main evaluation metric; it assumes correlated disparity estimates imply correct geometry.
  • domain assumption The generated left-right pairs are approximately rectified, so the x-coordinate of SEA-RAFT optical flow equals stereo disparity.
    In Sec. 4.3 ds is defined as the x-coordinate of optical flow; vertical misalignment or rotation in generated pairs would make this proxy inaccurate.
  • domain assumption AlignProp's truncated backpropagation through denoising optimizes the reward without collapsing the model.
    Sec. 3.2 relies on the prior method [30]; the paper does not independently verify that mode collapse is avoided beyond the reported scores.
  • domain assumption BLIP captions adequately describe the content of training stereo pairs.
    Sec. 4.1 uses BLIP for text prompts; the limitation section admits captions are short and sometimes inaccurate.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Text2Stereo: Repurposing Stable Diffusion for Stereo Generation with Consistency Rewards." pith.science (2026). https://pith.science/paper/IEI6GS5F

@misc{pith2026250605367,
  author       = {Pith},
  title        = {Pith review of: Text2Stereo: Repurposing Stable Diffusion for Stereo Generation with Consistency Rewards},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IEI6GS5F}},
  note         = {Machine review of arXiv:2506.05367}
}
read the original abstract

In this paper, we propose a novel diffusion-based approach to generate stereo images given a text prompt. Since stereo image datasets with large baselines are scarce, training a diffusion model from scratch is not feasible. Therefore, we propose leveraging the strong priors learned by Stable Diffusion and fine-tuning it on stereo image datasets to adapt it to the task of stereo generation. To improve stereo consistency and text-to-image alignment, we further tune the model using prompt alignment and our proposed stereo consistency reward functions. Comprehensive experiments demonstrate the superiority of our approach in generating high-quality stereo images across diverse scenarios, outperforming existing methods.

Figures

Figures reproduced from arXiv: 2506.05367 by the authors.

Figure 1
Figure 1. Given an input text prompt, our method synthesizes a stereo pair of left and right images. We use the generated left image [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. We show the overview of our approach comprising two stages (left and right). In the first stage, we fine-tune the pretrained Stable [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. We present example images from our dataset. In addition [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Given a stereo image pair, we estimate the stereo dispar [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Qualitative comparison of our approach against StereoDiffusion [ [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Evaluation of stereo consistency using Splatt3R [ [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]
Figure 7
Figure 7. Figure 7: We demonstrate the results through various stages of the training process. The numbers in the first and third rows correspond [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: We show the rewards averaged over 100 training and [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: We show the impact of different component of our sys [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

57 extracted references · 33 canonical work pages

  1. [1]

    Training diffusion models with reinforce- ment learning

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforce- ment learning. arXiv preprint arXiv:2305.13301, 2023. 3

    arXiv 2023

  2. [2]

    Chan, Connor Z

    Eric R. Chan, Connor Z. Lin, Matthew A. Chan, Koki Nagano, Boxiao Pan, Shalini De Mello, Orazio Gallo, Leonidas Guibas, Jonathan Tremblay, Sameh Khamis, Tero Karras, and Gordon Wetzstein. Efficient geometry-aware 3D generative adversarial networks. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. 2

    work page 2022

  3. [3]

    Luciddreamer: Domain-free gen- eration of 3d gaussian splatting scenes

    Jaeyoung Chung, Suyoung Lee, Hyeongjin Nam, Jaerin Lee, and Kyoung Mu Lee. Luciddreamer: Domain-free gen- eration of 3d gaussian splatting scenes. arXiv preprint arXiv:2311.13384, 2023. 2

    arXiv 2023

  4. [4]

    Directly fine-tuning diffusion models on differentiable re- wards

    Kevin Clark, Paul Vicol, Kevin Swersky, and David J Fleet. Directly fine-tuning diffusion models on differentiable re- wards. arXiv preprint arXiv:2309.17400, 2023. 3

    arXiv 2023

  5. [5]

    Objaverse: A universe of annotated 3d objects

    Matt Deitke, Dustin Schwenk, Jordi Salvador, Luca Weihs, Oscar Michel, Eli VanderBilt, Ludwig Schmidt, Kiana Ehsani, Aniruddha Kembhavi, and Ali Farhadi. Objaverse: A universe of annotated 3d objects. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13142–13153, 2023. 2

    work page 2023

  6. [6]

    Optimizing ddpm sampling with shortcut fine-tuning

    Ying Fan and Kangwook Lee. Optimizing ddpm sampling with shortcut fine-tuning. arXiv preprint arXiv:2301.13362,

    arXiv

  7. [7]

    Re- inforcement learning for fine-tuning text-to-image diffusion models

    Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Moham- mad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Re- inforcement learning for fine-tuning text-to-image diffusion models. Advances in Neural Information Processing Sys- tems, 36, 2024. 3

    work page 2024

  8. [8]

    Get3d: A generative model of high quality 3d tex- tured shapes learned from images

    Jun Gao, Tianchang Shen, Zian Wang, Wenzheng Chen, Kangxue Yin, Daiqing Li, Or Litany, Zan Gojcic, and Sanja Fidler. Get3d: A generative model of high quality 3d tex- tured shapes learned from images. Advances In Neural In- formation Processing Systems, 35:31841–31854, 2022. 2

    2022

Show all 57 references
  1. [9]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014. 2

    2014

  2. [10]

    Memory-efficient backpropagation through time

    Audrunas Gruslys, R ´emi Munos, Ivo Danihelka, Marc Lanc- tot, and Alex Graves. Memory-efficient backpropagation through time. Advances in neural information processing systems, 29, 2016. 3

    2016

  3. [11]

    Stylenerf: A style-based 3d aware generator for high- resolution image synthesis

    Jiatao Gu, Lingjie Liu, Peng Wang, and Christian Theobalt. Stylenerf: A style-based 3d aware generator for high- resolution image synthesis. In International Conference on Learning Representations, 2022. 2

    2022

  4. [12]

    Denoising dif- fusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising dif- fusion probabilistic models. Advances in neural information processing systems, 33:6840–6851, 2020. 1, 2, 3

    2020

  5. [13]

    LoRA: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen- Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA: Low-rank adaptation of large language models. InIn- ternational Conference on Learning Representations , 2022. 4

    2022

  6. [14]

    Holopix50k: A large-scale in-the-wild stereo image dataset

    Yiwen Hua, Puneet Kohli, Pritish Uplavikar, Anand Ravi, Saravana Gunaseelan, Jason Orozco, and Edward Li. Holopix50k: A large-scale in-the-wild stereo image dataset. In CVPR, 2020. 3

    2020

  7. [15]

    3d gaussian splatting for real-time radiance field rendering

    Bernhard Kerbl, Georgios Kopanas, Thomas Leimk ¨uhler, and George Drettakis. 3d gaussian splatting for real-time radiance field rendering. ACM Trans. Graph., 42(4):139–1,

  8. [16]

    Auto-encoding variational bayes

    Diederik P Kingma. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013. 3

    2013 arXiv

  9. [17]

    Simple and effective synthesis of in- door 3d scenes

    Jing Yu Koh, Harsh Agrawal, Dhruv Batra, Richard Tucker, Austin Waters, Honglak Lee, Yinfei Yang, Jason Baldridge, and Peter Anderson. Simple and effective synthesis of in- door 3d scenes. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 1169–1178, 2023. 1

    2023

  10. [18]

    Aligning text- to-image models using human feedback

    Kimin Lee, Hao Liu, Moonkyung Ryu, Olivia Watkins, Yuqing Du, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, and Shixiang Shane Gu. Aligning text- to-image models using human feedback. arXiv preprint arXiv:2302.12192, 2023. 3

    2023 arXiv

  11. [19]

    Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation, 2022. 4

    2022

  12. [20]

    Practical stereo matching via cascaded recurrent net- work with adaptive correlation

    Jiankun Li, Peisen Wang, Pengfei Xiong, Tao Cai, Ziwei Yan, Lei Yang, Jiangyu Liu, Haoqiang Fan, and Shuaicheng Liu. Practical stereo matching via cascaded recurrent net- work with adaptive correlation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern R...

    2022

  13. [21]

    Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model

    Jiahao Li, Hao Tan, Kai Zhang, Zexiang Xu, Fujun Luan, Yinghao Xu, Yicong Hong, Kalyan Sunkavalli, Greg Shakhnarovich, and Sai Bi. Instant3d: Fast text-to-3d with sparse-view generation and large reconstruction model. In The Twelfth International Conference on Learning Represe...

    2024

  14. [22]

    Synthesizing light field from a single image with variable mpi and two network fusion

    Qinbo Li and Nima Khademi Kalantari. Synthesizing light field from a single image with variable mpi and two network fusion. ACM Transactions on Graphics, 39(6), 2020. 2

    2020

  15. [23]

    Visual instruction tuning, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning, 2023. 8

    2023

  16. [24]

    Mayer, E

    N. Mayer, E. Ilg, P. H ¨ausser, P. Fischer, D. Cremers, A. Dosovitskiy, and T. Brox. A large dataset to train convo- lutional networks for disparity, optical flow, and scene flow estimation. In CVPR, 2016. 3

    2016

  17. [25]

    Realfusion: 360deg reconstruction of any object from a single image

    Luke Melas-Kyriazi, Iro Laina, Christian Rupprecht, and Andrea Vedaldi. Realfusion: 360deg reconstruction of any object from a single image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages 8446–8455, 2023. 2

    2023

  18. [26]

    Srinivasan, Matthew Tancik, Jonathan T

    Ben Mildenhall, Pratul P. Srinivasan, Matthew Tancik, Jonathan T. Barron, Ravi Ramamoorthi, and Ren Ng. Nerf: Representing scenes as neural radiance fields for view syn- thesis. In Proceedings of the European Conference on Com- puter Vision (ECCV), 2020. 2

    2020

  19. [27]

    Giraffe: Represent- ing scenes as compositional generative neural feature fields

    Michael Niemeyer and Andreas Geiger. Giraffe: Represent- ing scenes as compositional generative neural feature fields. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2021. 2

    2021

  20. [28]

    Text2immersion: Generative immersive scene with 3d gaussians

    Hao Ouyang, Kathryn Heal, Stephen Lombardi, and Tiancheng Sun. Text2immersion: Generative immersive scene with 3d gaussians. arXiv preprint arXiv:2312.09242,

    arXiv

  21. [29]

    Dreamfusion: Text-to-3d using 2d diffusion

    Ben Poole, Ajay Jain, Jonathan T Barron, and Ben Milden- hall. Dreamfusion: Text-to-3d using 2d diffusion. arXiv preprint arXiv:2209.14988, 2022. 2

    2022 arXiv

  22. [30]

    Aligning text-to-image diffusion models with reward backpropagation

    Mihir Prabhudesai, Anirudh Goyal, Deepak Pathak, and Katerina Fragkiadaki. Aligning text-to-image diffusion models with reward backpropagation. arXiv preprint arXiv:2310.03739, 2023. 2, 3, 4, 5

    2023 arXiv

  23. [31]

    Sinmpi: Novel view synthesis from a single image with expanded multiplane images

    Guo Pu, Peng-Shuai Wang, and Zhouhui Lian. Sinmpi: Novel view synthesis from a single image with expanded multiplane images. In SIGGRAPH Asia 2023 Conference Papers, pages 1–10, 2023. 2

    2023

  24. [32]

    Learning transferable visual models from natural language supervision, 2021

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision, 2021. 5

    2021

  25. [33]

    Pixel- synth: Generating a 3d-consistent experience from a single image

    Chris Rockwell, David F Fouhey, and Justin Johnson. Pixel- synth: Generating a 3d-consistent experience from a single image. In Proceedings of the IEEE/CVF International Con- ference on Computer Vision, pages 14104–14113, 2021. 1

    2021

  26. [34]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj ¨orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684–10695, 2022. 1, 2, 3, 4

    2022

  27. [35]

    Graf: Generative radiance fields for 3d-aware im- age synthesis

    Katja Schwarz, Yiyi Liao, Michael Niemeyer, and Andreas Geiger. Graf: Generative radiance fields for 3d-aware im- age synthesis. Advances in Neural Information Processing Systems, 33:20154–20166, 2020. 2

    2020

  28. [36]

    A picture is worth a thousand words: Principled recaptioning improves image generation, 2023

    Eyal Segalis, Dani Valevski, Danny Lumen, Yossi Matias, and Yaniv Leviathan. A picture is worth a thousand words: Principled recaptioning improves image generation, 2023. 8

    2023

  29. [37]

    Layered depth images

    Jonathan Shade, Steven Gortler, Li-wei He, and Richard Szeliski. Layered depth images. In Proceedings of the 25th annual conference on Computer graphics and interac- tive techniques, pages 231–242, 1998. 2

    1998

  30. [38]

    Mvdream: Multi-view diffusion for 3d gen- eration

    Yichun Shi, Peng Wang, Jianglong Ye, Mai Long, Kejie Li, and Xiao Yang. Mvdream: Multi-view diffusion for 3d gen- eration. arXiv preprint arXiv:2308.16512, 2023. 2

    2023 arXiv

  31. [39]

    3d photography using context-aware layered depth inpainting

    Meng-Li Shih, Shih-Yang Su, Johannes Kopf, and Jia-Bin Huang. 3d photography using context-aware layered depth inpainting. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , pages 8028–8038, 2020. 1, 2, 6, 7

    2020

  32. [40]

    Splatt3r: Zero-shot gaussian splat- ting from uncalibarated image pairs

    Brandon Smart, Chuanxia Zheng, Iro Laina, and Vic- tor Adrian Prisacariu. Splatt3r: Zero-shot gaussian splat- ting from uncalibarated image pairs. arXiv preprint arXiv:2408.13912, 2024. 6, 7

    2024 arXiv

  33. [41]

    Pushing the boundaries of view extrapolation with multiplane images

    Pratul P Srinivasan, Richard Tucker, Jonathan T Barron, Ravi Ramamoorthi, Ren Ng, and Noah Snavely. Pushing the boundaries of view extrapolation with multiplane images. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 175–184, 2019. 1

    2019

  34. [42]

    Dreamgaussian: Generative gaussian splatting for effi- cient 3d content creation

    Jiaxiang Tang, Jiawei Ren, Hang Zhou, Ziwei Liu, and Gang Zeng. Dreamgaussian: Generative gaussian splatting for effi- cient 3d content creation. arXiv preprint arXiv:2309.16653,

    arXiv

  35. [43]

    Nerf-supervised deep stereo

    Fabio Tosi, Alessio Tonioni, Daniele De Gregorio, and Mat- teo Poggi. Nerf-supervised deep stereo. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 855–866, 2023. 1, 3, 4

    2023

  36. [44]

    Single-view view syn- thesis with multiplane images

    Richard Tucker and Noah Snavely. Single-view view syn- thesis with multiplane images. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 551–560, 2020. 1, 2

    2020

  37. [45]

    Layer-structured 3d scene inference via view synthesis

    Shubham Tulsiani, Richard Tucker, and Noah Snavely. Layer-structured 3d scene inference via view synthesis. In Proceedings of the European Conference on Computer Vi- sion (ECCV), pages 302–317, 2018. 2

    2018

  38. [46]

    Stereodiffusion: Training-free stereo image generation using latent diffusion models

    Lezhong Wang, Jeppe Revall Frisvad, Mark Bo Jensen, and Siavash Arjomand Bigdeli. Stereodiffusion: Training-free stereo image generation using latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 7416–7425, 2024...

    2024

  39. [47]

    Sea-raft: Simple, efficient, accurate raft for optical flow

    Yihan Wang, Lahav Lipson, and Jia Deng. Sea-raft: Simple, efficient, accurate raft for optical flow. In European Confer- ence on Computer Vision, pages 36–54. Springer, 2025. 5

    2025

  40. [48]

    Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distilla- tion

    Zhengyi Wang, Cheng Lu, Yikai Wang, Fan Bao, Chongxuan Li, Hang Su, and Jun Zhu. Prolificdreamer: High-fidelity and diverse text-to-3d generation with variational score distilla- tion. Advances in Neural Information Processing Systems , 36, 2024. 2

    2024

  41. [49]

    Synsin: End-to-end view synthesis from a sin- gle image

    Olivia Wiles, Georgia Gkioxari, Richard Szeliski, and Justin Johnson. Synsin: End-to-end view synthesis from a sin- gle image. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7467–7477,

  42. [50]

    Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis

    Xiaoshi Wu, Yiming Hao, Keqiang Sun, Yixiong Chen, Feng Zhu, Rui Zhao, and Hongsheng Li. Human preference score v2: A solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341 ,

    arXiv

  43. [51]

    Desai Xie, Jiahao Li, Hao Tan, Xin Sun, Zhixin Shu, Yi Zhou, Sai Bi, S ¨oren Pirk, and Arie E. Kaufman. Carve3d: Improving multi-view reconstruction consistency for dif- fusion models with rl finetuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern ...

    2024

  44. [52]

    Depth any- thing v2, 2024

    Lihe Yang, Bingyi Kang, Zilong Huang, Zhen Zhao, Xiao- gang Xu, Jiashi Feng, and Hengshuang Zhao. Depth any- thing v2, 2024. 5

    2024

  45. [53]

    Wonderjourney: Going from anywhere to everywhere

    Hong-Xing Yu, Haoyi Duan, Junhwa Hur, Kyle Sargent, Michael Rubinstein, William T Freeman, Forrester Cole, De- qing Sun, Noah Snavely, Jiajun Wu, et al. Wonderjourney: Going from anywhere to everywhere. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern R...

    2024

  46. [54]

    Mvimgnet: A large-scale dataset of multi-view images

    Xianggang Yu, Mutian Xu, Yidan Zhang, Haolin Liu, Chongjie Ye, Yushuang Wu, Zizheng Yan, Chenming Zhu, Zhangyang Xiong, Tianyou Liang, et al. Mvimgnet: A large-scale dataset of multi-view images. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognit...

    2023

  47. [55]

    Text2nerf: Text-driven 3d scene generation with neu- ral radiance fields

    Jingbo Zhang, Xiaoyu Li, Ziyu Wan, Can Wang, and Jing Liao. Text2nerf: Text-driven 3d scene generation with neu- ral radiance fields. IEEE Transactions on Visualization and Computer Graphics, 2024. 2

    2024

  48. [56]

    Stereo magnification: Learning view synthesis using multiplane images

    Tinghui Zhou, Richard Tucker, John Flynn, Graham Fyffe, and Noah Snavely. Stereo magnification: Learning view synthesis using multiplane images. arXiv preprint arXiv:1805.09817, 2018. 2

    2018 arXiv

  49. [57]

    Fsgs: Real-time few-shot view synthesis using gaussian splatting

    Zehao Zhu, Zhiwen Fan, Yifan Jiang, and Zhangyang Wang. Fsgs: Real-time few-shot view synthesis using gaussian splatting. In European Conference on Computer Vision , pages 145–163. Springer, 2024. 4

    2024

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.