Pith. sign in

REVIEW 5 major objections 4 minor 2 cited by

UltraZoom: Generating Gigapixel Images from Regular Photos

T0 review · 5 major / 4 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 UltraZoom argues that one standard-resolution photo of an object, a few handheld close-ups, and short bridging videos are enough to synthesize a coherent, photorealistic gigapixel image of the whole object.

desk verdict A genuinely useful integration of registration, per-instance fine-tuning, and gigapixel inference, but the evaluation does not actually test generalization beyond the trained close-up regions. read the letter →

arxiv 2506.13756 v1 pith:KDK6DKP2 submitted 2025-06-16 cs.CV

Jingwei Ma , Vivek Jayaram , Brian Curless , Ira Kemelmacher-Shlizerman , Steven M. Seitz This is my paper
classification cs.CV
keywords gigapixelimagingreference-basedsuper-resolutionper-instancefine-tuningimageregistrationtexturesynthesisdiffusionmodelslow-rankadaptationarbitrary-scale
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

UltraZoom argues that gigapixel-scale imaging of an object does not require a capture rig or thousands of stitched photos: a single full-view phone photo, a few macro close-ups, and short videos bridging them are enough to synthesize a zoomable gigapixel image. The core move is to turn the close-ups into a per-object training set, degrade them to match the full view's blur, color, and compression, and fine-tune a pretrained generative model on the resulting low-to-high-resolution patch pairs. If this is right, everyday handheld captures could produce zoomable images of textiles, food, products, and other objects with detail faithful to the actual surfaces.

What carries the argument

The load-bearing machinery is the per-instance paired dataset. From each color-corrected close-up, high-resolution patches are paired with degraded versions: bicubic downsampling by the estimated scale, an extra 2x downsampling to mimic distant-capture blur, and conditional JPEG compression, after color matching to the registered full-image region. The scale and color statistics come from registration that tracks grids of points through the bridging videos with a point tracker, fits a 2D similarity transform per video segment with RANSAC (random-sample consensus), and chains the transforms; the scale is read off the transform's determinant. The paired data then drives low-rank (LoRA) fine-tuning of a frozen pretrained generative upscaler with a flow-matching loss, and sliding-window inference blends outputs using latent-space averaging with stride variation.

What would settle it

Capture an object with a printed fiducial grid of known physical spacing, run the full UltraZoom capture protocol, and compare the estimated close-up location and scale factor against the grid's known position; if registration drifts by more than one close-up patch width on repetitive or low-texture material, the paired-data premise fails and the synthesized details will not align with the real object.

Watch

Extended reading notes

Core claim

The central claim is that arbitrary-material objects can be upscaled to 6–30x, or up to 5 gigapixels, by adapting a pretrained generative model to a per-instance dataset built from the close-ups alone. Exact pixel alignment between the close-up and the full image is never required; instead, coarse registration through the bridging video supplies the scale factor and color statistics, and degraded close-up patches simulate the full view's appearance. At inference the fine-tuned model is applied in sliding windows, with overlapping latents and pixels blended and the stride varied across denoising steps to suppress seam artifacts. On 15 captured objects the paper reports the strongest patch-level fidelity among the compared super-resolution methods, and a user study preferred its quality 96% of the time and its consistency 79% of the time.

Load-bearing premise

Everything depends on the bridging video letting the point tracker chain similarity transforms from each close-up to the full image accurately enough to estimate scale and color statistics; if the video loses track during rapid zoom or on repetitive texture, the paired training data is corrupted and the gigapixel output would place plausible-looking details in the wrong locations.

Editorial extensions

If this is right

  • A phone and a few seconds of bridging video become a viable alternative to specialized gigapixel capture systems for objects, not just scenes.
  • Because the training pairs come from the object's own close-ups, the method transfers to arbitrary materials without requiring pre-existing gigapixel datasets.
  • Registration by chained similarity transforms removes the dependence on SIFT-style keypoint matching or dense optical flow between macro and wide-angle views.
  • Per-instance adaptation of a frozen pretrained generative model keeps detail faithful to the exemplar while limiting hallucinated content.
  • The sliding-window design with stride variation gives seamless pan-and-zoom over outputs spanning 0.25 to 5 gigapixels.

Reading between the lines

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

  • Editorial extension: the bridging-video registration recipe could serve other paired macro/wide tasks, such as archiving paintings or inspecting manufactured surfaces, where only scale and color statistics are needed.
  • Editorial extension: the most likely failure regime is spatially varying texture density, such as a printed logo on plain fabric, because the local-window model has no global position information; adding positional or layout conditioning would be a direct fix.
  • Editorial extension: the reported 3.82-hour inference for one gigapixel output on a single A100 means interactive consumer use would require the patch-retrieval caching the paper names as future work.
  • Editorial extension: the separate full-shot capture could be replaced by keyframe selection from the final bridging video, turning the whole pipeline into a single continuous phone capture.
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

5 major / 4 minor

Summary. The paper presents UltraZoom, a system for generating gigapixel-resolution images from a single full-view photograph plus a small number of close-up photographs. The method first captures a bridging video between each close-up and the full view, uses point tracking and chained similarity transforms to register the close-ups within the full image, and estimates the relative scale and color statistics. It then constructs a paired dataset by degrading color-corrected close-up patches to simulate the appearance of the corresponding full-image regions. A pretrained text-to-image diffusion model is fine-tuned per instance using LoRA and a flow-matching objective on these pairs. At inference, the model is applied in a sliding-window manner with stride variation to produce a seamless gigapixel output. The paper reports quantitative comparisons using LR-MAE, Patch-FID/KID, and a user study, along with qualitative results and ablations.

Significance. UltraZoom addresses a practical and appealing problem: producing gigapixel imagery from casual handheld captures. The core idea—constructing a per-instance paired dataset by degrading close-ups and fine-tuning a large pretrained diffusion model—is reasonable and the qualitative results are visually impressive, with a project page that promises code and full-resolution interactive results. The registration approach using bridging videos and point tracking is a clever workaround for the challenging macro-to-regular alignment problem. However, the evaluation does not yet substantiate the central claim that the method generalizes across the entire object. All perceptual metrics and the user study compare outputs against the same close-ups used for per-instance fine-tuning, so the reported numbers may largely reflect memorization of the exemplar textures rather than successful synthesis in unvisited regions. If a held-out evaluation confirms the generalization, this would be a strong contribution; as it stands, the paper is promising but needs additional evidence.

major comments (5)
  1. [§4.2, Table 1] The evaluation in Table 1 is circular with respect to the paper's central claim of generalization. The Patch-FID and KID metrics compare patches sampled from the generated output against patches from the captured close-ups, which are exactly the images used to construct the paired training set in Section 3.2. The user study similarly presents a real close-up patch alongside a nearby output patch, with the output sampled near the trained close-up region. Because the model is fine-tuned to map degraded versions of these close-ups to the clean close-ups, high perceptual similarity to these references is at least partially enforced by the training objective. Neither metric provides evidence about regions of the object distant from the close-ups, where the model must infer unseen texture. I recommend a held-out evaluation: for each object, withhold one close-up region from training and evaluate Patch-FID/KID and user preference on patches sampled from areas far from any trained close-up, comparing against baselines on the same patches.
  2. [§3.1 (Capture Process); Abstract] The abstract and the introduction state that the input is 'a full-shot image' and 'one or more close-ups,' but the capture process described in Section 3.1 requires a sequence of bridging videos connecting each close-up to the full image. The registration pipeline in Section 3.1 and Figure 3 depends entirely on these videos; without them, the scale factor and color statistics needed for dataset construction cannot be estimated. This discrepancy is load-bearing because a reader cannot reproduce the method from the stated inputs, and it also changes the practical claim of 'minimal input.' Please either revise the contributions to acknowledge the video requirement, or demonstrate that registration can be performed from the still images alone.
  3. [Table 1] No variance or statistical significance is reported for any quantitative metric. With 18 close-up instances across 15 objects, the Patch-FID gap between Ours (134.986) and the best baseline (310.371) may be significant, but the user-study percentages (96.08% vs. 1.96% top-1 quality) and the LR-MAE comparisons (0.040 vs. 0.007 for Thera) need per-example standard deviations and paired significance tests before the claims of 'significantly outperforms' and 'competitive LR consistency' can be accepted. Reporting only aggregate numbers leaves open the possibility that a few easy examples drive the results.
  4. [§4.4 Ablations] The ablation study is purely qualitative, showing only two examples in Figure 5. It demonstrates that fine-tuning and degradation alignment improve fidelity near the reference, but it does not test whether the fine-tuned model propagates details correctly across the full object at locations distant from the close-ups. A quantitative ablation with held-out patches would directly address the core generalization claim and would also contextualize the contribution of each component (scale, fine-tuning, degradation alignment, stride variation) with numbers rather than visual inspection alone.
  5. [§3.1, Image Registration] The accuracy of the chained similarity transforms is never quantified. The method relies on CoTracker3 point tracking across rapidly changing fields of view and repetitive textures, and any registration error would corrupt the paired training data and, in turn, the fine-tuning objective. I recommend reporting registration error on at least a few sequences, for example by comparing the estimated scale against a manually measured or fiducial-based scale, or by measuring reprojection error of tracked points after the chained transform. Without such validation, the robustness claim for the registration step is not supported.
minor comments (4)
  1. [Figure 4 caption] The caption states that 'the output patch may not be perfectly aligned with the reference' and that it is 'sampled near the reference.' It would be helpful to quantify the approximate distance between the output patch and the reference patch, since the user study relies on this 'nearby' relationship to judge consistency.
  2. [§3.3, Gigapixel Inference] The inference details such as the sliding-window size, the overlap amount, and the exact stride-variation schedule are only referenced to the supplemental material. Please provide the default values in the main text to make the system reproducible without needing the supplement.
  3. [References] Several bibliographic entries have formatting issues, including a broken URL in the Fischler and Bolles entry and overlong URL fragments that may not render correctly. Please clean up the reference list in the camera-ready version.
  4. [§4.1, Baseline Methods] The description of ContinuousSR as 'trained on 4-8x continuous scales' is not obviously supported by the cited paper (Peng et al. 2025), which is about 2D Gaussian modeling for arbitrary-scale super-resolution. Please verify that the baseline description matches the cited work.

Circularity Check

1 steps flagged · score 6.0 of 10

Evaluation is circular: Patch-FID/KID and the user study compare generated outputs against the same close-ups used to construct the per-instance training set, so reported fidelity scores contain a self-comparison component.

  1. fitted input called prediction [Section 3.2 Per-Instance Fine-tuning; Section 4.2 Quantitative Comparison; Table 1]
    "we fine-tune a pretrained generative model on the instance-specific dataset we construct from the close-ups ... [Eq. 3: L_FM = E ||u_theta(c_t,t,d,y) - u(c_t,t)||^2] ... Since the close-ups and corresponding regions in the generated high-resolution output are not pixel-aligned, we compute Fréchet Inception Distance (FID) and Kernel Inception Distance (KID) on randomly sampled patches from the real and generated images ... [Table 1: Patch-FID/KID to evaluate perceptual similarity to the captured close-ups]"

    The per-instance fine-tuning objective (Eq. 3) trains the model to map degraded patches d ~ P(D(C~)) to clean patches c ~ P(C~), where C~ are the color-corrected captured close-ups. The Table 1 metrics then evaluate the model by comparing generated output patches to patches from 'the captured close-ups' — the very same images that supplied the training targets. High Patch-FID/KID is therefore partly enforced by construction: the model is explicitly optimized to reproduce close-up patch statistics, so the metric measures fit to the training distribution rather than generalization to unvisited object regions.

full rationale

The derivation chain for the system itself — capture, video-based registration, scale estimation, degradation alignment, per-instance fine-tuning, sliding-window inference — is not circular: the full-view image is not used to construct the paired training data beyond color statistics, and the inference-time mapping is applied to image regions not seen in training. No self-citation is load-bearing; the method rests on standard components (CoTracker3, FLUX, DreamBooth, LoRA) with independent origins. The significant circularity is confined to the evaluation. The quantitative metrics (Patch-FID/KID) compare generated output patches to patches from the captured close-ups, which are also the source of the per-instance training pairs. Because the training objective explicitly minimizes a reconstruction/flow-matching loss between degraded close-up patches and clean close-up patches, a low Patch-FID/KID relative to those same close-ups is a partially self-referential success measure. The user study's 'consistency with the exemplar' question is likewise a comparison against the exemplars used for fine-tuning, and the output patch is chosen 'nearby' the reference close-up. Thus the reported 96.08% preference and Patch-FID 134.986 demonstrate that the model can reproduce the trained close-up appearance in their vicinity, but they do not establish faithful detail synthesis across the full object, where the model must infer unseen texture. This is a partial, evaluation-level circularity rather than a derivation-level equivalence: the method is not simply renaming its input, and the gigapixel output is not identical to the close-ups by construction. But the central claim of 'faithful, photorealistic gigapixel images from minimal input' rests heavily on these self-comparison metrics, warranting a score of 6 rather than a lower score.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

No new physical entities are introduced. The central claim rests on object self-similarity, the availability and trackability of bridge videos, per-instance adaptation of a pretrained generative model, and hand-chosen degradation constants.

free parameters (2)
  • JPEG quality for degradation simulation = 75
    Hand-chosen in Section 3.1 (Degradation Alignment) to simulate iPhone JPEG artifacts; changes the training pair distribution.
  • Additional blur downsampling factor = 2x
    An extra 2x bicubic downsampling is applied to mimic optical blur; hand-set constant in Section 3.1.
assumptions (4)
  • domain assumption A single object is self-similar and covered by a limited set of material patterns, so a small number of close-ups suffices to infer missing details across the entire object.
    Stated in Section 1 as the motivation; if false, sparse close-ups cannot represent the object and the method fails.
  • domain assumption Bridging videos allow reliable accumulation of 2D similarity transforms that register close-ups to the full image despite rapid FOV change and repetitive texture.
    Used in Section 3.1 Image Registration; depends on CoTracker3 point tracking and RANSAC over tracked points.
  • domain assumption A pretrained text-to-image model (FLUX) can be adapted with DreamBooth/LoRA to learn object-specific LR-to-HR mappings from a small per-instance paired dataset.
    Invoked in Section 3.2; the central generative capability is inherited from prior work.
  • ad hoc to paper The degradation model (downsampling, blur, JPEG) sufficiently closes the train-test gap between close-ups and the full-view image.
    Section 3.1 Degradation Alignment; manual choice, acknowledged as a limitation in Section 5.

how reviews work

0 comments
Cite this review

Pith. "Pith review of UltraZoom: Generating Gigapixel Images from Regular Photos." pith.science (2026). https://pith.science/paper/KDK6DKP2

@misc{pith2026250613756,
  author       = {Pith},
  title        = {Pith review of: UltraZoom: Generating Gigapixel Images from Regular Photos},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KDK6DKP2}},
  note         = {Machine review of arXiv:2506.13756}
}
read the original abstract

We present UltraZoom, a system for generating gigapixel-resolution images of objects from casually captured inputs, such as handheld phone photos. Given a full-shot image (global, low-detail) and one or more close-ups (local, high-detail), UltraZoom upscales the full image to match the fine detail and scale of the close-up examples. To achieve this, we construct a per-instance paired dataset from the close-ups and adapt a pretrained generative model to learn object-specific low-to-high resolution mappings. At inference, we apply the model in a sliding window fashion over the full image. Constructing these pairs is non-trivial: it requires registering the close-ups within the full image for scale estimation and degradation alignment. We introduce a simple, robust method for getting registration on arbitrary materials in casual, in-the-wild captures. Together, these components form a system that enables seamless pan and zoom across the entire object, producing consistent, photorealistic gigapixel imagery from minimal input.

Figures

Figures reproduced from arXiv: 2506.13756 by the authors.

Figure 1
Figure 1. Given a standard-resolution image and a close-up reference of an object, our method faithfully restores object-specific details to each patch and merges [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Method Overview. (a) Dataset Construction: For each scene, we capture a close-up, a full-view image, and a bridging video that connects the two views. We track the close-up region across the video to register it within the full image, estimating the relative scale 𝑠 and the color statistics 𝐻 of the matched region. These are used to construct a dataset of paired high-resolution and degraded image patches, designed t… view at source ↗
Figure 3
Figure 3. Close-up-to-full registration. Given a close-up, the full image, and a connecting video, we first split the video into shorter segments to improve point tracking, as the field of view changes rapidly. At the start of each segment, we initialize a grid of points and track them across frames within the segment. A 2D similarity transform is then estimated for each segment using RANSAC over the tracked points. These tra… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Qualitative Comparison. Rows are ordered from low to high scale. For each example, we compare 1024 × 1024 patches across methods. From left to right: a 1024 × 1024 crop from the captured close-up (reference), full image with patch location (green box), low-resolution i…
Figure 5
Figure 5. Figure 5: Ablations. We show the effect of individual method components on two examples. (1) Pretrained model at default 4x scale produces results with hallucinated details. (2) Using the correct scale without the close-up reference still results in hallucination and introduces …
Figure 6
Figure 6. Figure 6: Additional Results. We present further qualitative comparisons between baseline methods and ours on additional captured examples. These results continue to demonstrate the strong visual quality and exemplar consistency achieved by our approach. Visit the supplemental v…

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. GarmentZoom: Generating Zoomable Images from Garment Listings

    cs.CV 2026-06 unverdicted novelty 6.0 of 10

    GarmentZoom trains one model to synthesize unaligned close-up details into full-view garment images across continuous scales 3-20x without per-instance tuning.

  2. MicroZoom: Structure-Preserving Detail Synthesis at Extreme Scale

    cs.CV 2026-07 conditional novelty 5.0 of 10

    A cascaded, segmentation-conditioned, per-instance diffusion method synthesizes globally coherent gigapixel microscopic detail from a phone photo and sparse microscope references at up to 350×.

Reference graph

Works this paper leans on

35 extracted references · 17 canonical work pages · cited by 2 Pith papers

  1. [1]

    https://doi.org/10.1109/ CVPR.2004.1315043 Chang Chen, Zhiwei Xiong, Xinmei Tian, Zheng-Jun Zha, and Feng Wu

    I–I. https://doi.org/10.1109/ CVPR.2004.1315043 Chang Chen, Zhiwei Xiong, Xinmei Tian, Zheng-Jun Zha, and Feng Wu

    arXiv 2004

  2. [2]

    https://doi.org/10.1109/ICCV.1999.790383 Alexei A

    1033–1038 vol.2. https://doi.org/10.1109/ICCV.1999.790383 Alexei A. Efros and William T. Freeman

    work page arXiv 1999

  3. [10]

    arXiv:1505.07376 [cs.CV] https://arxiv.org/abs/ 1505.07376 Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans

    Texture Synthesis Using Convolutional Neural Networks. arXiv:1505.07376 [cs.CV] https://arxiv.org/abs/ 1505.07376 Jonathan Ho, Chitwan Saharia, William Chan, David J Fleet, Mohammad Norouzi, and Tim Salimans

    arXiv

  4. [12]

    arXiv:2106.09685 [cs.CL] https://arxiv.org/abs/2106.09685 Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun

    LoRA: Low-Rank Adaptation of Large Language Models. arXiv:2106.09685 [cs.CL] https://arxiv.org/abs/2106.09685 Xuecai Hu, Haoyuan Mu, Xiangyu Zhang, Zilei Wang, Tieniu Tan, and Jian Sun

    arXiv

  5. [16]

    https://doi.org/10.1023/B: VISI.0000029664.99615.94 Liying Lu, Wenbo Li, Xin Tao, Jiangbo Lu, and Jiaya Jia

    Distinctive Image Features from Scale-Invariant Keypoints.Inter- national Journal of Computer Vision60 (11 2004), 91–. https://doi.org/10.1023/B: VISI.0000029664.99615.94 Liying Lu, Wenbo Li, Xin Tao, Jiangbo Lu, and Jiaya Jia

    arXiv 2004

  6. [17]

    MASA-SR: Matching Acceleration and Spatial Adaptation for Reference-Based Image Super-Resolution

    MASA-SR: Matching Acceleration and Spatial Adaptation for Reference-Based Image Super-Resolution. arXiv:2106.02299 [cs.CV] https://arxiv.org/abs/2106.02299 Long Peng, Anran Wu, Wenbo Li, Peizhe Xia, Xueyuan Dai, Xinjie Zhang, Xin Di, Haoze Sun, Renjing Pei, Yang Wang, et al

    work page Pith review arXiv

  7. [18]

    Marco Pesavento, Marco Volino, and Adrian Hilton

    Pixel to Gaussian: Ultra-Fast Continuous Super-Resolution with 2D Gaussian Modeling.arXiv preprint arXiv:2503.06617 (2025). Marco Pesavento, Marco Volino, and Adrian Hilton

    arXiv 2025

  8. [19]

    Attention-based Multi-Reference Learning for Image Super-Resolution

    Attention-based Multi- Reference Learning for Image Super-Resolution. arXiv:2108.13697 [cs.CV] https: //arxiv.org/abs/2108.13697 Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman

    work page Pith review arXiv

Show all 35 references
  1. [20]

    arXiv:2208.12242 [cs.CV] https://arxiv.org/abs/2208

    DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. arXiv:2208.12242 [cs.CV] https://arxiv.org/abs/2208. 12242 Tamar Rott Shaham, Tali Dekel, and Tomer Michaeli

    arXiv

  2. [21]

    arXiv:1905.01164 [cs.CV] https: //arxiv.org/abs/1905.01164 Taizhang Shang, Qiuju Dai, Shengchen Zhu, Tong Yang, and Yandong Guo

    SinGAN: Learning a Generative Model from a Single Natural Image. arXiv:1905.01164 [cs.CV] https: //arxiv.org/abs/1905.01164 Taizhang Shang, Qiuju Dai, Shengchen Zhu, Tong Yang, and Yandong Guo

    1905 arXiv

  3. [26]

    arXiv:2403.08436 [cs.CV] https://arxiv.org/abs/2403.08436 Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C.K

    PFStorer: Personalized Face Restoration and Super-Resolution. arXiv:2403.08436 [cs.CV] https://arxiv.org/abs/2403.08436 Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin C.K. Chan, and Chen Change Loy. 2024b. Exploiting Diffusion Prior for Real-World Image Super-Resolution. (...

    2024 arXiv

  4. [27]

    InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Dual-Camera Super-Resolution With Aligned Attention Modules. InProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). 2001–2010. Yifan Wang, Aleksander Holynski, Brian L. Curless, and Steven M. Seitz. 2024a. Infinite Texture: Text-guided High Resolutio...

    2001 arXiv

  5. [30]

    arXiv:2006.04139 [cs.CV] https://arxiv.org/abs/2006.04139 Jianchao Yang, John Wright, Thomas S

    Learning Texture Transformer Network for Image Super-Resolution. arXiv:2006.04139 [cs.CV] https://arxiv.org/abs/2006.04139 Jianchao Yang, John Wright, Thomas S. Huang, and Yi Ma

    2006 arXiv

  6. [33]

    arXiv:1912.00515 [eess.IV] https://arxiv.org/abs/1912.00515 Zhifei Zhang, Zhaowen Wang, Zhe Lin, and Hairong Qi

    Texture Hallucination for Large-Factor Painting Super-Resolution. arXiv:1912.00515 [eess.IV] https://arxiv.org/abs/1912.00515 Zhifei Zhang, Zhaowen Wang, Zhe Lin, and Hairong Qi. 2019b. Image Super-Resolution by Neural Texture Transfer. arXiv:1903.00834 [cs.CV] https://arxiv.o...

    1912 arXiv

  7. [34]

    arXiv:1807.10547 [cs.CV] https://arxiv.org/abs/1807.10547 Jun-Yan Zhu, Richard Zhang, Deepak Pathak, Trevor Darrell, Alexei A

    CrossNet: An End-to-end Reference-based Super Resolution Network using Cross-scale Warping. arXiv:1807.10547 [cs.CV] https://arxiv.org/abs/1807.10547 Jun-Yan Zhu, Richard Zhang, Deepak Pathak, Trevor Darrell, Alexei A. Efros, Oliver Wang, and Eli Shechtman

    arXiv

  8. [35]

    Toward Multimodal Image-to-Image Translation. arXiv:1711.11586 [cs.CV] https://arxiv.org/abs/1711.11586 UltraZoom: Generating Gigapixel Images from Regular Photos•9 Close-up (reference) Full Image (crop: green box) Input (bicubic) 1.0x 1.0x 1.0x 6.141x 29.790x 6.155x Thera (+ ...

    arXiv

  9. [1981]

    ACM24, 6 (1981), 381–395

    Random Sample Consensus: A Paradigm for Model Fitting with Applications to Image Analysis and Automated Cartography.Commun. ACM24, 6 (1981), 381–395. /brokenurl#http://publication.wilsonwong.me/load. php?id=233282275 William T. Freeman, Thouis R. Jones, and Egon C. Pasztor

    1981

  10. [2001]

    Philipp Fischer, Alexey Dosovitskiy, Eddy Ilg, Philip Häusser, Caner Hazırbaş, Vladimir Golkov, Patrick van der Smagt, Daniel Cremers, and Thomas Brox

    Image Quilting for Texture Synthesis and Transfer.Proceedings of SIGGRAPH 2001(August 2001), 341–346. Philipp Fischer, Alexey Dosovitskiy, Eddy Ilg, Philip Häusser, Caner Hazırbaş, Vladimir Golkov, Patrick van der Smagt, Daniel Cremers, and Thomas Brox

    2001

  11. [2003]

    Black Forest Labs

    Graphcut Textures: Image and Video Synthesis Using Graph Cuts.ACM Transactions on Graphics, SIGGRAPH 200322, 3 (July 2003), 277–286. Black Forest Labs

    2003

  12. [2004]

    InProceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition,

    Super-resolution through neighbor embedding. InProceedings of the 2004 IEEE Computer Society Conference on Computer Vision and Pattern Recognition,

    2004

  13. [2007]

    Vivek Kwatra, Arno Schödl, Irfan Essa, Greg Turk, and Aaron Bobick

    Captur- ing and Viewing Gigapixel Images.ACM Transactions on Graphics (Proceedings of SIGGRAPH 2007)26, 3 (2007), to appear. Vivek Kwatra, Arno Schödl, Irfan Essa, Greg Turk, and Aaron Bobick

    2007

  14. [2010]

    Image Super-Resolution Via Sparse Representation.IEEE Transactions on Image Processing19, 11 (2010), 2861–

    2010

  15. [2014]

    https://doi.org/10.7717/peerj.453 Tuomas Varanka, Tapani Toivonen, Soumya Tripathy, Guoying Zhao, and Erman Acar

    scikit-image: image processing in Python.PeerJ2 (June 2014), e453. https://doi.org/10.7717/peerj.453 Tuomas Varanka, Tapani Toivonen, Soumya Tripathy, Guoying Zhao, and Erman Acar

    2014 doi

  16. [2015]

    arXiv:1504.06852 [cs.CV] https://arxiv.org/abs/1504.06852 M

    FlowNet: Learning Optical Flow with Convolutional Networks. arXiv:1504.06852 [cs.CV] https://arxiv.org/abs/1504.06852 M. Fischler and R. Bolles

    arXiv

  17. [2016]

    arXiv:1603.03417 [cs.CV] https://arxiv.org/abs/1603.03417 Stéfan van der Walt, Johannes L

    Texture Networks: Feed-forward Synthesis of Textures and Stylized Images. arXiv:1603.03417 [cs.CV] https://arxiv.org/abs/1603.03417 Stéfan van der Walt, Johannes L. Schönberger, Juan Nunez-Iglesias, François Boulogne, Joshua D. Warner, Neil Yager, Emmanuelle Gouillart, Tony Yu...

    arXiv

  18. [2017]

    arXiv:1705.06566 [cs.CV] https://arxiv.org/abs/1705

    Learning Texture Manifolds with the Periodic Spatial GAN. arXiv:1705.06566 [cs.CV] https://arxiv.org/abs/1705. 06566 Jianrui Cai, Hui Zeng, Hongwei Yong, Zisheng Cao, and Lei Zhang

    arXiv

  19. [2018]

    arXiv:1706.02823 [cs.CV] https://arxiv.org/abs/1706.02823 Ruikang Xu, Mingde Yao, and Zhiwei Xiong

    TextureGAN: Controlling Deep Image Synthesis with Texture Patches. arXiv:1706.02823 [cs.CV] https://arxiv.org/abs/1706.02823 Ruikang Xu, Mingde Yao, and Zhiwei Xiong

    arXiv

  20. [2019]

    arXiv:1903.00875 [cs.CV] https://arxiv.org/abs/1903.00875 Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht

    Meta-SR: A Magnification-Arbitrary Network for Super-Resolution. arXiv:1903.00875 [cs.CV] https://arxiv.org/abs/1903.00875 Nikita Karaev, Iurii Makarov, Jianyuan Wang, Natalia Neverova, Andrea Vedaldi, and Christian Rupprecht

    1903 arXiv

  21. [2020]

    arXiv:2005.12597 [eess.IV] https://arxiv.org/abs/2005.12597 Ansh Sharma, Albert Xiao, Praneet Rathi, Rohit Kundu, Albert Zhai, Yuan Shen, and Shenlong Wang

    Perceptual Extreme Super Resolution Network with Receptive Field Block. arXiv:2005.12597 [eess.IV] https://arxiv.org/abs/2005.12597 Ansh Sharma, Albert Xiao, Praneet Rathi, Rohit Kundu, Albert Zhai, Yuan Shen, and Shenlong Wang

    2005 arXiv

  22. [2021]

    arXiv preprint arXiv:2106.15282(2021)

    Cascaded Diffusion Models for High Fidelity Image Generation. arXiv preprint arXiv:2106.15282(2021). Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    2021 arXiv

  23. [2022]

    arXiv:2103.12716 [cs.CV] https://arxiv.org/abs/2103.12716 Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo

    UltraSR: Spatial Encoding is a Missing Key for Implicit Image Function-based Arbitrary-Scale Super-Resolution. arXiv:2103.12716 [cs.CV] https://arxiv.org/abs/2103.12716 Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo

    arXiv

  24. [2023]

    MultiDiffusion: Fusing Diffusion Paths for Controlled Image Generation. arXiv:2302.08113 [cs.CV] https: //arxiv.org/abs/2302.08113 Alexander Becker, Rodrigo Caye Daudt, Dominik Narnhofer, Torben Peters, Nando Met- zger, Jan Dirk Wegner, and Konrad Schindler

    arXiv

  25. [2024]

    arXiv:2409.01491 [cs.CV] https://arxiv.org/abs/2409.01491 Dmitry Ulyanov, Vadim Lebedev, Andrea Vedaldi, and Victor Lempitsky

    EarthGen: Generating the World from Top-Down Views. arXiv:2409.01491 [cs.CV] https://arxiv.org/abs/2409.01491 Dmitry Ulyanov, Vadim Lebedev, Andrea Vedaldi, and Victor Lempitsky

    arXiv

  26. [2025]

    Urs Bergmann, Nikolay Jetchev, and Roland Vollgraf

    Thera: Aliasing-Free Arbitrary- Scale Super-Resolution with Neural Heat Fields.arXiv preprint arXiv:2311.17643 (2025). Urs Bergmann, Nikolay Jetchev, and Roland Vollgraf

    2025

  27. [2873]

    https://doi.org/10.1109/TIP.2010.2050625 Xuaner Zhang, Qifeng Chen, Ren Ng, and Vladlen Koltun. 2019a. Zoom to Learn, Learn to Zoom. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Yulun Zhang, Zhifei Zhang, Stephen DiVerdi, Zhaowen ...

    2010

Pith tools

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