Pith. sign in

REVIEW 4 major objections 7 minor 49 references

A two-stage cascade plus material masks can turn a phone photo and a few microscope close-ups into a seamless, materially grounded gigapixel image at up to 350×.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-31 06:36 UTC pith:QR53MOJV

load-bearing objection Solid next-step systems paper on extreme reference SR: cascade + seg masks help coherence at 65–350×, but the “materially grounded” claim is only weakly measured outside the training patches. the 4 major comments →

arxiv 2607.24729 v1 pith:QR53MOJV submitted 2026-07-27 cs.CV

MicroZoom: Structure-Preserving Detail Synthesis at Extreme Scale

Huy Huynh , Jingwei Ma , Brian Curless , Ira Kemelmacher-Shlizerman , Steven M. Seitz This is my paper
classification cs.CV
keywords extreme-scale super-resolutionreference-based SRgigapixel synthesiscascaded diffusionmicroscopysegmentation-guided generationper-instance fine-tuningstructure-preserving detail
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

MicroZoom claims that extreme-scale, full-image super-resolution—from ordinary phone photos up to roughly 65–350×—is feasible if you ground synthesis in sparse real microscope references rather than pure hallucination. At those factors the low-resolution input loses almost all texture identity, and millions of local patches drift into globally broken patterns (for example, a fabric weave that no longer lines up). The authors answer with a deliberate two-stage cascade: the first stage restores large-scale pattern regularity where context still exists; the second refines local texture on that canvas. A segmentation mask further pins material boundaries so one texture does not bleed into another. On self-captured everyday objects, human raters prefer the results over prior extreme and reference-based methods, and the outputs stay visually tied to the real microscope character of each object. The point is exploratory visualization of microscopic surface detail across an entire object, not pixel-exact reconstruction.

Core claim

At full-image magnifications of about 65–350×, separating global pattern recovery from local texture refinement in a cascaded, per-instance diffusion pipeline—and conditioning on both structure and segmentation masks—produces seamless gigapixel imagery that is more globally coherent and more preferred by people than single-stage or ungrounded extreme super-resolution, while remaining grounded in sparse real microscope references.

What carries the argument

Cascaded per-instance fine-tuning: the total scale factor is split into two sequential models (coarse then refine), each adapted with LoRA on focus-stacked microscope patches; ControlNet carries structural alignment from the low-res input, and channel-concatenated segmentation masks enforce material boundaries, with a conditioning warm-up that learns texture before structure.

Load-bearing premise

A few focus-stacked microscope patches plus a phone photo and optional masks are enough to keep the vast unobserved regions of a gigapixel canvas materially faithful to that specific object, not just globally coherent but wrong.

What would settle it

On multi-material objects, compare full-canvas outputs against held-out microscope patches from regions never used in fine-tuning: if weave geometry, grain, or material identity systematically mismatch those patches—or if human preference collapses once judges see those held-out references—the claim of materially grounded full-object synthesis fails.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Phone photos plus a handful of consumer microscope shots become a practical route to interactive, full-object microscopic visualization.
  • Extreme SR pipelines should treat global pattern coherence as a separate stage rather than relying on single-shot upscaling.
  • Segmentation (or other region labels) is a necessary regularizer at material boundaries when the low-res signal is nearly ambiguous.
  • Plausible, reference-grounded synthesis—not exact reconstruction—is the realistic objective once magnification exceeds tens of times.
  • Inference cost and per-instance training remain the main barriers to casual use; distillation or texture-primitive tiling would directly unlock broader access.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same cascade-plus-boundary idea could transfer to other domains where local generators destroy large-scale regularity (histology slides, aerial fabric/terrain, printed security patterns).
  • If a general microscopic-texture prior replaced per-instance LoRA, zero-shot MicroZoom-style exploration might become possible without capturing close-ups of every new object.
  • Train/test color and exposure gaps highlighted in the failures suggest that joint capture calibration or appearance transfer is as load-bearing as the generative architecture itself.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The manuscript presents MicroZoom, a per-instance generative framework that upscales a single smartphone photograph of an object to gigapixel resolution at magnification factors of roughly 65–350×, conditioned on a sparse set of focus-stacked consumer-microscope close-ups. The method combines (i) a two-stage cascade in which a coarser SR stage (e.g., 20×) establishes large-scale pattern regularity before a refinement stage (e.g., 5×) adds texture detail, (ii) ControlNet structural conditioning plus channel-concatenated SAM segmentation masks to resolve material-boundary ambiguity, and (iii) a conditioning warm-up schedule, all trained via LoRA on top of a frozen FLUX.1-dev backbone with a flow-matching objective, and deployed at inference with MultiDiffusion-style sliding-window aggregation. Evaluation on 21 self-captured objects reports the best DISTS (0.213) and competitive LSD (1.014) against UltraZoom, Chain-of-Zoom, ContinuousSR, and IP-Adapter+ControlNet on a synthetic-LR split with pixel-aligned ground truth, plus a 79-participant preference study (60.5% quality / 52.6% similarity votes) on real smartphone inputs. The authors are explicit that the goal is plausible synthesis, not exact reconstruction, and document failure modes (train/test domain gap, generative shift) in §4.6.

Significance. If the central claim holds, this is a useful contribution to extreme-scale, reference-based super-resolution: a working per-instance pipeline that produces globally coherent gigapixel imagery at 65–347× from casual captures, with a clear modular design (cascade, segmentation conditioning, warm-up), an honest failure-case section (§4.6), additive ablations (Table 2, Fig. 4), and a non-trivial 79-participant human study. The scoped framing — "plausible synthesis, not exact reconstruction" — is appropriate and well maintained in most of the text. However, the "materially grounded" half of the claim currently rests on an evaluation protocol in which the only pixel-aligned fidelity numbers are computed against the same data used for per-instance fine-tuning, and the real-input setting has no fidelity measurement at all; until that gap is closed, the fidelity claims are supported mainly by human preference rather than by measurement.

major comments (4)
  1. [§4.4.1 / Table 1 (protocol), with §3.2.1] Train/evaluation overlap in the only pixel-aligned fidelity protocol. Per §3.2.1, the per-instance LoRA is fine-tuned on random crops of X_GT (the focus-stacked microscope captures); per §4.4.1, the Syn-LR split — the sole source of the DISTS/LSD numbers in Table 1 and all of Table 2 — is constructed by downsampling those same X_GT captures and scoring against them. No microscope view or spatial region is stated to be held out from fine-tuning before evaluation. A per-instance generative model scored against its own training signal can achieve low DISTS by reconstructing memorized content, so the 0.213-vs-0.232 margin over UltraZoom and the absolute fidelity values cannot currently distinguish 'synthesis grounded in references that generalizes across the canvas' from 'good fit to the observed patches.' This is load-bearing for the claim of 'materially grounded' imagery. The head-to-head
  2. [§4.4.2 (Ref-LR evaluation gap)] No fidelity measurement exists for the actual use case. The Ref-LR setting (smartphone LR input) — the setting the paper targets and the one used for the user study — has zero quantitative fidelity evaluation; it is assessed only by preference votes against a single shown ground-truth patch (§4.4.2). Preference votes of this form cannot detect a globally coherent but materially incorrect texture, which is precisely the failure mode the paper itself documents (§4.6.2 generative shift, texture bleeding despite segmentation). Given that the abstract claims outputs are 'grounded in the material character of the real references,' at least one quantitative grounding measurement on Ref-LR outputs is needed — e.g., re-photographing a small set of objects at microscope scale in regions covered by the full-image output and scoring those regions, or reporting DISTS/LSD of Ref-LR outputs against the
  3. [§3.4 (inference) with §3.1.3 / §3.2.2] Provenance of the segmentation mask at gigapixel inference is unspecified. §3.1.3 defines M via SAM on the microscopic captures and downsamples it to M_LR for training; §3.2.2 makes the mask a channel-concatenated conditioning input that the flow-matching loss treats as a hard constraint. But at inference (§3.4) the model is applied over sliding windows of the full image F upsampled to gigapixel resolution, which requires a segmentation mask over the entire canvas at the output resolution — a mask that was never observed microscopically over most of the surface. The text does not state how this full-canvas mask is produced (SAM on the bicubic-upsampled F? SAM on F then upsampled? per-window?). This is load-bearing for both reproducibility and for the boundary-bleeding claims, and it interacts with §4.6.2, where ControlNet 'overpowers' the segmentation cue: without knowing the inference-t
  4. [Table 2 / §4.5] Ablation rows are not mutually comparable. The '+ Segmentation Map' row (DISTS 0.224, LSD 1.076) is computed only on the 11 multi-material objects, while the adjacent rows ('+ Cascaded Modules' 0.220/1.021 and '+ Cond. Warm Up' 0.213/1.014) are computed on all 21 objects. The table therefore does not demonstrate an additive improvement from segmentation: 0.224 (11-object subset) is worse than 0.220 (full set) at the preceding stage, and the final improvement could reflect the evaluation-set change rather than the added module. Relatedly, the cascade stage regresses LSD (0.951→1.021) and the full method never recovers the single-stage LSD (1.014 vs 0.951), which the text acknowledges only briefly. Please re-run Table 2 with every stage evaluated on both the full set and the 11-object subset, and discuss the persistent LSD regression relative to single-stage fine-tuning more directly, sinc
minor comments (7)
  1. [§3.3] The flow-matching objective in §3.3 is unnumbered and its notation is inconsistent: the target field is written u_t(x|x_1) but x_1 is never defined (the path is defined with data point x and noise x_0). Please number the equation and align the notation.
  2. [§4.4.2] User study design: one trial per object (21 trials) with a single ground-truth patch shown may not represent consistency across the full canvas. Reporting per-criterion agreement statistics (e.g., inter-rater agreement or bootstrap CIs on the 60.5%/52.6% vote shares) would strengthen Table 1's right half.
  3. [Fig. 3] Fig. 3 caption states the compared crops 'are not exactly aligned, but have been selected from similar regions.' Since the qualitative comparison is a key piece of evidence, consider including at least one exactly aligned crop (e.g., from the Syn-LR protocol) in the main figure; the supplement's Syn-LR figures partially address this.
  4. [§4.2] Training budget is ambiguous: §3.3 specifies 1000 + 1000 warm-up steps, while §4.2 reports '~30 minutes per epoch, batch size 1.' Please state total optimization steps per stage and per object so the per-instance cost is clear.
  5. [§4.5 / Supplement S-1] Typos/style: §4.5 'Table 2 report' → 'reports'; §S-1 'we employ Log Spectral (LSD) metric' → 'the Log Spectral Distance (LSD) metric'; Eq. (4) does not define K or the windowing used to estimate the power spectra P̄.
  6. [Abstract / §4.1] The stated magnification range '65–350×' is supported by the data (max shown 347.11×, Fig. 7), but the abstract and §1 would be more precise stating the observed range 65–347×.
  7. [§3.1.2 / §4.6.1] The color-matching step in Eq. (3) matches X_GT to the phone image F, yet the train/test gap (§4.6.1) still produces color shifts. A brief note on why histogram matching in gamma space is insufficient here (e.g., spatially varying phone-camera processing) would help readers assess generality to other capture devices.

Circularity Check

1 steps flagged

Empirical SR paper with only mild train/eval overlap on Syn-LR; no derivation-by-construction circularity.

specific steps
  1. fitted input called prediction [§3.1.2–3.2.1 and Table 1 / §4.4.1 (Syn-LR metrics)]
    "To construct paired dataset for supervised per-instance super-resolution, we derive training pairs solely from the microscopic ground truth X_GT. ... Y=Down_s(ColorMatch(X_GT,F)) ... At each training step, we sample (x,y,m)∼D(X_GT,Y,M_LR) ... Note that the automated metrics (DISTS, LSD) are evaluated on the synthetic dataset (Syn-LR) to allow for pixel-aligned ground truth comparisons."

    The LoRA is fit to random crops of X_GT; Syn-LR inputs are downsamples of those same X_GT captures and are scored against the same X_GT. With no stated hold-out of captures or spatial regions, the reported DISTS/LSD (and the 0.213 vs 0.232 margin) partly re-measure reconstruction of the fine-tuning signal rather than independent material grounding. Mild only: same protocol applies to UltraZoom, and Ref-LR preference is separate.

full rationale

MicroZoom is a methods/empirical generation paper, not a first-principles derivation. The cascade (s=s1×s2), ControlNet structure path, segmentation channel-concat, conditioning warm-up, and flow-matching objective are architectural choices checked by ablations and by comparison to external baselines (ContinuousSR, Chain-of-Zoom, IP-Adapter+ControlNet, UltraZoom) plus a 79-person Ref-LR preference study. None of those steps reduces an equation to its own fitted input by definition, imports a uniqueness theorem from overlapping authors, or renames a known law. The sole mild circularity is evaluation hygiene: Syn-LR DISTS/LSD in Table 1 are computed on bicubic downsamples of the same microscope X_GT crops used for per-instance LoRA fine-tuning, with no stated spatial or capture hold-out, so absolute fidelity partly measures fit to the training signal. Relative ranking under the same protocol and the Ref-LR user study remain independent evidence. Score 2, not higher: the central coherence/preference claims are not forced by that loop.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 1 invented entities

Load-bearing content is mostly standard generative-SR tooling plus engineering choices. The central claim rests on pretrained flow-matching priors, per-instance adaptation sufficiency, cascade error tradeoffs, segmentation quality, and sparse-reference coverage—not on new physical entities. Free parameters are training/inference hyperparameters and per-object scale splits that shape outcomes but are not presented as universal constants.

free parameters (4)
  • cascade scale split (s = s1 × s2) = asymmetric; e.g. factors with s2 limited by 1024 crops / min network input (often ≤5×)
    Total magnification is hand-decomposed into coarse then fine stages (e.g., ~20× then ~5×) under architecture input-size constraints; choice directly affects structure vs. detail tradeoff claimed as central.
  • LoRA rank and training schedule = rank r=32; 1000 + 1000 steps; batch size 1; Prodigy lr=1.0
    Adapter capacity and the 1000-step mask-only warm-up plus 1000-step full conditioning are chosen hyperparameters that control how strongly instance texture and structure bind.
  • relative scale factor s from pixel density ρ = per-object, reported range ~65×–350×
    s = ρ_XGT / ρ_F is measured per capture with a tool in-frame; calibration error propagates into paired supervision and claimed magnification.
  • focus-stack burst count B and focal planes K
    Noise averaging and all-in-focus fusion quality depend on capture counts that are pipeline parameters, not derived.
axioms (6)
  • domain assumption A pretrained flow-matching text-to-image backbone (Flux.1-dev) plus lightweight LoRA/ControlNet adaptation can be steered into consumer-microscope texture domains.
    Sec. 3.2–3.3 treat this prior as given; without it the per-instance recipe has no generator.
  • ad hoc to paper First-stage coarser SR recovers large-scale pattern regularity well enough that second-stage refinement improves perceptual coherence more than it amplifies structural errors.
    Explicit design bet in Sec. 3.2; paper admits cascade drift but claims it is less harmful than single-stage global incoherence.
  • domain assumption SAM multi-class masks (when used) are accurate enough to act as hard material-boundary constraints under extreme LR ambiguity.
    Sec. 3.1.3 and mask channel-concat in 3.2.2; failures in Sec. 4.6 show ControlNet can still overpower masks.
  • ad hoc to paper Sparse local microscope references characterize material appearance for the entire object surface under generative synthesis.
    Problem formulation and dataset pipeline assume N close-ups ground full-canvas texture; this is required for ‘materially grounded’ full-image claims.
  • domain assumption MultiDiffusion with variable-stride window aggregation yields seamless gigapixel outputs without destructive boundary artifacts.
    Sec. 3.4 relies on prior MultiDiffusion/UltraZoom tiling results as operational fact.
  • standard math Flow matching / standard squared vector-field regression is a valid training objective for this conditioned SR task.
    Sec. 3.3 writes the usual FM loss; not re-derived.
invented entities (1)
  • MicroZoom cascaded structure-then-detail pipeline with segmentation-channel conditioning and conditioning warm-up no independent evidence
    purpose: Name the specific system that separates global pattern recovery from local texture refinement and hard-constrains materials at boundaries for extreme-scale RefSR.
    Not a new physical object; a composed method identity. Independent evidence is only the paper’s own ablations and user study, not an external measurable entity.

pith-pipeline@v1.2.0-grok45-kimik3 · 18312 in / 4160 out tokens · 103496 ms · 2026-07-31T06:36:02.414664+00:00 · methodology

0 comments
read the original abstract

We introduce MicroZoom, a generative framework for gigapixel image synthesis at the microscopic scale. Given a standard photograph and a sparse set of consumer-grade microscope close-ups, MicroZoom synthesizes a seamless, gigapixel-resolution image grounded in the material character of the real references, enabling exploratory visualization of microscopic texture across the full spatial extent of an object. Our goal is plausible synthesis, not exact reconstruction. We focus on full-image, reference-based, extreme-scale super-resolution at magnification levels of up to 350x, a setting that introduces two major challenges: (1) recovering texture-specific detail from highly lossy inputs near ambiguous material boundaries, and (2) preserving correct large-scale pattern structure, such as the repeating geometry of a fabric weave, across millions of local predictions. We address these with a two-stage cascaded design, where the first stage recovers global pattern coherence and the second refines local texture detail, supplemented by a segmentation mask to guide synthesis at ambiguous boundaries. We verify our approach on a collection of self-captured everyday objects and demonstrate globally coherent, materially grounded gigapixel imagery.

Figures

Figures reproduced from arXiv: 2607.24729 by Brian Curless, Huy Huynh, Ira Kemelmacher-Shlizerman, Jingwei Ma, Steven M. Seitz.

Figure 1
Figure 1. Figure 1: Given a single standard resolution image from a smartphone and a few microscope images, MicroZoom performs extreme-scale super-resolution to [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Method Overview. (a) Dataset Construction: We curate synthetic training data by capturing a full-view smartphone image and multiple high￾magnification digital microscope views. To resolve the shallow depth of field in microscopy, we implement a focus-stacking pipeline that fuses multiple focal planes into a single, sharp ground truth X𝐺𝑇 . We then generate aligned training pairs (X𝐺𝑇 , Y ) via downsampling… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative Comparison. Rows are ordered from lowest to highest magnification. For each example, we compare a 6000 × 6000 patch across methods. From left to right, we show a microscope capture of the object and the full image with the region of the input patch highlighted. Then, we compare the low-resolution input (upscaled using bicubic interpolation), four baselines, and our final result. These patches a… view at source ↗
Figure 4
Figure 4. Figure 4: Ablations. We visualize the additive contribution of each module on the final output. Naive super-resolution on a pretrained ControlNet fails to resolve meaningful detail due to the domain gap. Full fine-tuning sharpens textures but suffers from structural hallucinations and material bleeding. Adding segmentation maps resolves these semantic conflicts, guiding boundaries between textures. Finally, our casc… view at source ↗
Figure 5
Figure 5. Figure 5: Failure Cases. We visualize two distinct failure cases in our method. The left example showcases the train/test domain gap that exists even after color correction, which leads to inaccurate color shifts. The right example showcases semantic drift, where ambiguous low-frequency information in the input image can generate incorrect textures, even after being labeled. 5 Discussion and Future Work We presented… view at source ↗
Figure 6
Figure 6. Figure 6: Additional Multi-Material Examples. We present 8 more examples of objects with multiple materials, demonstrating MicroZoom’s ability to synthesize plausible details for scenes with varying textures and boundaries. For more results, see our project page [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Additional Single Material Examples. We present 8 more examples of objects with a single material, demonstrating MicroZoom’s ability to synthesize plausible textures for objects containing a single pattern type. For more results, see our project page [PITH_FULL_IMAGE:figures/full_fig_p010_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Consistent Synthesis. We present multiple distinct crops sampled from various locations within a single super-resolution output. Our method consistently generates plausible structures across the entire image, demon￾strating robust performance across regions. S-3 Synthetic LR Examples In this section, we present the comprehensive set of generated results using synthetic low-resolution (Syn-LR) inputs, which… view at source ↗
Figure 9
Figure 9. Figure 9: Syn-LR Multi-Material Examples. Full qualitative examples generated from synthetic LR inputs for objects containing multiple materials, corre￾sponding to the quantitative evaluation in [PITH_FULL_IMAGE:figures/full_fig_p012_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Syn-LR Single Material Examples. Full qualitative examples generated from synthetic LR inputs for objects containing a single material, corresponding to the quantitative evaluation in [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

49 extracted references · 1 canonical work pages

  1. [1]

    Omer Bar-Tal, Lior Yariv, Yaron Lipman, and Tali Dekel. 2023. MultiDiffusion: Fusing Diffusion Paths for Controlled Image Generation. arXiv:2302.08113 [cs.CV] https://arxiv.org/abs/2302.08113

    Pith/arXiv arXiv 2023

  2. [2]

    Black Forest Labs. 2024. FLUX.1: A family of flow matching models. https: //blackforestlabs.ai

    2024

  3. [3]

    Peter J Burt and Edward H Adelson. 1983. The Laplacian pyramid as a compact image code.IEEE Transactions on Communications31, 4 (1983), 532–540

    1983

  4. [4]

    Jiezhang Cao, Jingyun Liang, Kai Zhang, Yawei Li, Yulun Zhang, Wenguan Wang, and Luc Van Gool. 2022. Reference-based Image Super-Resolution with De- formable Attention Transformer. arXiv:2207.11938 [cs.CV] https://arxiv.org/abs/ 2207.11938

    Pith/arXiv arXiv 2022

  5. [5]

    Jin Cao, Hong-Xing Yu, and Jiajun Wu. 2025. WonderZoom: Multi-Scale 3D World Generation. arXiv:2512.09164 [cs.CV] https://arxiv.org/abs/2512.09164

    arXiv 2025

  6. [6]

    Xiangyu Chen, Xintao Wang, Wenlong Zhang, Xiangtao Kong, Yu Qiao, Jiantao Zhou, and Chao Dong. 2025. HAT: Hybrid Attention Transformer for Image Restoration. arXiv:2309.05239 [cs.CV] https://arxiv.org/abs/2309.05239

    arXiv 2025

  7. [7]

    Simoncelli

    Keyan Ding, Kede Ma, Shiqi Wang, and Eero P. Simoncelli. 2020. Image Quality Assessment: Unifying Structure and Texture Similarity.CoRRabs/2004.07728 (2020). https://arxiv.org/abs/2004.07728

    Pith/arXiv arXiv 2020

  8. [8]

    Chao Dong, Chen Change Loy, Kaiming He, and Xiaoou Tang. 2014. Image super- resolution using deep convolutional networks.arXiv preprint arXiv:1501.00092 (2014)

    Pith/arXiv arXiv 2014

  9. [9]

    Freeman, T.R

    W.T. Freeman, T.R. Jones, and E.C. Pasztor. 2002. Example-based super-resolution. IEEE Computer Graphics and Applications22, 2 (2002), 56–65. doi:10.1109/38.988747

    work page doi:10.1109/38.988747 2002

  10. [10]

    Bermano, Gal Chechik, and Daniel Cohen-Or

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit H. Bermano, Gal Chechik, and Daniel Cohen-Or. 2022. An Image is Worth One Word: Personalizing Text-to-Image Generation using Textual Inversion. arXiv:2208.01618 [cs.CV] https://arxiv.org/abs/2208.01618

    Pith/arXiv arXiv 2022

  11. [11]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. 2021. LoRA: Low-Rank Adaptation of Large Language Models. arXiv:2106.09685 [cs.CL] https://arxiv.org/abs/2106.09685

    Pith/arXiv arXiv 2021

  12. [12]

    Yuming Jiang, Kelvin C. K. Chan, Xintao Wang, Chen Change Loy, and Zi- wei Liu. 2021. Robust Reference-based Super-Resolution via C2-Matching. arXiv:2106.01863 [cs.CV] https://arxiv.org/abs/2106.01863

    Pith/arXiv arXiv 2021

  13. [13]

    Bryan Sangwoo Kim, Jeongsol Kim, and Jong Chul Ye. 2025. Chain-of-Zoom: Extreme Super-Resolution via Scale Autoregression and Preference Alignment. arXiv:2505.18600 [cs.CV] https://arxiv.org/abs/2505.18600

    Pith/arXiv arXiv 2025

  14. [14]

    Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick

    Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C. Berg, Wan-Yen Lo, Piotr Dollár, and Ross Girshick. 2023. Segment Anything. arXiv:2304.02643 [cs.CV] https://arxiv.org/abs/2304.02643

    Pith/arXiv arXiv 2023

  15. [15]

    Black Forest Labs. 2024. FLUX. https://github.com/black-forest-labs/flux

    2024

  16. [16]

    Christian Ledig, Lucas Theis, Ferenc Huszár, et al. 2017. Photo-realistic single image super-resolution using a generative adversarial network. InCVPR

    2017

  17. [17]

    Yuseung Lee, Kunho Kim, Hyunjin Kim, and Minhyuk Sung. 2023. SyncDiffusion: Coherent Montage via Synchronized Joint Diffusions. arXiv:2306.05178 [cs.CV] https://arxiv.org/abs/2306.05178

    Pith/arXiv arXiv 2023

  18. [18]

    Jingyun Liang, Jiezhang Cao, Guolei Sun, Kai Zhang, Luc Van Gool, and Radu Timofte. 2021. SwinIR: Image restoration using swin transformer. InICCV Work- shops

    2021

  19. [19]

    Bee Lim, Sanghyun Son, Heewon Kim, Seungjun Nah, and Kyoung Mu Lee. 2017. Enhanced deep residual networks for single image super-resolution. InCVPR Workshops

    2017

  20. [20]

    Xinqi Lin, Jingwen He, Ziyan Chen, et al. 2023. DiffBIR: Towards blind image restoration with generative diffusion prior.arXiv preprint arXiv:2308.15070(2023)

    Pith/arXiv arXiv 2023

  21. [21]

    David G Lowe. 2004. Distinctive image features from scale-invariant keypoints. International Journal of Computer Vision60, 2 (2004), 91–110

    2004

  22. [22]

    Simian Luo, Yiqin Tan, Longbo Huang, Jian Li, and Hang Zhao. 2023. Latent Con- sistency Models: Synthesizing High-Resolution Images with Few-Step Inference. arXiv:2310.04378 [cs.CV] https://arxiv.org/abs/2310.04378

    Pith/arXiv arXiv 2023

  23. [23]

    Jingwei Ma, Vivek Jayaram, Brian Curless, Ira Kemelmacher-Shlizerman, and Steven M. Seitz. 2025. UltraZoom: Generating Gigapixel Images from Regular Photos. arXiv:2506.13756 [cs.CV] https://arxiv.org/abs/2506.13756

    Pith/arXiv arXiv 2025

  24. [24]

    Taesung Park, Ming-Yu Liu, Ting-Chun Wang, and Jun-Yan Zhu. 2019. Semantic Image Synthesis with Spatially-Adaptive Normalization. arXiv:1903.07291 [cs.CV] https://arxiv.org/abs/1903.07291

    Pith/arXiv arXiv 2019

  25. [25]

    William Peebles and Saining Xie. 2023. Scalable Diffusion Models with Trans- formers. arXiv:2212.09748 [cs.CV] https://arxiv.org/abs/2212.09748

    Pith/arXiv arXiv 2023

  26. [26]

    Long Peng, Anran Wu, Wenbo Li, Peizhe Xia, Xueyuan Dai, Xinjie Zhang, Xin Di, Haoze Sun, Renjing Pei, Yang Wang, Yang Cao, and Zheng-Jun Zha. 2025. Pixel to Gaussian: Ultra-Fast Continuous Super-Resolution with 2D Gaussian Modeling. arXiv:2503.06617 [cs.CV] https://arxiv.org/abs/2503.06617

    Pith/arXiv arXiv 2025

  27. [27]

    Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen. 2023. Improving Image Generation with Better Captions.arXiv preprint arXiv:2309.16609 (2023)

    Pith/arXiv arXiv 2023

  28. [28]

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. 2022. High-Resolution Image Synthesis with Latent Diffusion Models. arXiv:2112.10752 [cs.CV] https://arxiv.org/abs/2112.10752

    Pith/arXiv arXiv 2022

  29. [29]

    Nataniel Ruiz, Yuanzhen Li, Varun Jampani, Yael Pritch, Michael Rubinstein, and Kfir Aberman. 2023. DreamBooth: Fine Tuning Text-to-Image Diffusion Models for Subject-Driven Generation. arXiv:2208.12242 [cs.CV] https://arxiv.org/abs/ 2208.12242

    Pith/arXiv arXiv 2023

  30. [30]

    Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Den- ton, Seyed Kamyar Seyed Ghasemipour, Burcu Karagol Ayan, S. Sara Mahdavi, Rapha Gontijo Lopes, Tim Salimans, Jonathan Ho, David J Fleet, and Mohammad Norouzi. 2022. Photorealistic Text-to-Image Diffusion Models with Deep Lan- guage Understanding. arXiv:2205.11487 [cs.CV] https:/...

    Pith/arXiv arXiv 2022

  31. [31]

    Axel Sauer, Dominik Lorenz, Andreas Blattmann, and Robin Rombach. 2023. Adversarial Diffusion Distillation. arXiv:2311.17042 [cs.CV] https://arxiv.org/ abs/2311.17042

    Pith/arXiv arXiv 2023

  32. [32]

    Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. 2023. Consistency Models. arXiv:2303.01469 [cs.LG] https://arxiv.org/abs/2303.01469

    Pith/arXiv arXiv 2023

  33. [33]

    Libin Sun and James Hays. 2012. Super-resolution from internet-scale scene matching.2012 IEEE International Conference on Computational Photography (ICCP)(2012), 1–12. https://api.semanticscholar.org/CorpusID:2311801

    2012

  34. [34]

    Jianyi Wang, Zongsheng Yue, Shangchen Zhou, Kelvin CK Chan, and Chen Change Loy. 2023. Exploiting diffusion prior for real-world image super- resolution.arXiv preprint arXiv:2305.07015(2023)

    Pith/arXiv arXiv 2023

  35. [35]

    Xiaojuan Wang, Janne Kontkanen, Brian Curless, Steve Seitz, Ira Kemelmacher, Ben Mildenhall, Pratul Srinivasan, Dor Verbin, and Aleksander Holynski. 2024. Generative Powers of Ten. arXiv:2312.02149 [cs.CV] https://arxiv.org/abs/2312. 02149

    Pith/arXiv arXiv 2024

  36. [36]

    Xintao Wang, Liangbin Xie, Chao Dong, and Ying Shan. 2021. Real-ESRGAN: Training real-world blind super-resolution with pure synthetic data. InICCV Workshops

    2021

  37. [37]

    Xintao Wang, Ke Yu, Chao Dong, and Chen Change Loy. 2018. Recovering Realistic Texture in Image Super-resolution by Deep Spatial Feature Transform. arXiv:1804.02815 [cs.CV] https://arxiv.org/abs/1804.02815

    Pith/arXiv arXiv 2018

  38. [38]

    Xintao Wang, Ke Yu, Shixiang Wu, et al . 2018. ESRGAN: Enhanced super- resolution generative adversarial networks. InECCV Workshops

    2018

  39. [39]

    Zihan Wang, Ziliang Xiong, Hongying Tang, and Xiaobing Yuan. 2024. Detail-Enhancing Framework for Reference-Based Image Super-Resolution. arXiv:2405.00431 [cs.CV] https://arxiv.org/abs/2405.00431

    Pith/arXiv arXiv 2024

  40. [40]

    Rongyuan Wu, Tao Yang, Lingchen Sun, Zhengqiang Zhang, Shuai Li, and Lei Zhang. 2024. SeeSR: Towards Semantics-Aware Real-World Image Super- Resolution. arXiv:2311.16518 [cs.CV] https://arxiv.org/abs/2311.16518

    Pith/arXiv arXiv 2024

  41. [41]

    Jiahua Xiao, Jiawei Zhang, Dongqing Zou, Xiaodan Zhang, Jimmy Ren, and Xing Wei. 2024. Semantic Segmentation Prior for Diffusion-Based Real-World Super-Resolution. arXiv:2412.02960 [cs.CV] https://arxiv.org/abs/2412.02960

    Pith/arXiv arXiv 2024

  42. [42]

    Fuzhi Yang, Huan Yang, Jianlong Fu, Hongtao Lu, and Baining Guo

  43. [43]

    Tao Yang, Rongyuan Wu, Peiran Ren, Xuansong Xie, and Lei Zhang. 2024. Pixel- Aware Stable Diffusion for Realistic Image Super-resolution and Personalized Stylization. arXiv:2308.14469 [cs.CV] https://arxiv.org/abs/2308.14469

    Pith/arXiv arXiv 2024

  44. [44]

    Hu Ye, Jun Zhang, Sibo Liu, Xiao Han, and Wei Yang. 2023. IP-Adapter: Text Compatible Image Prompt Adapter for Text-to-Image Diffusion Models. (2023)

    2023

  45. [45]

    Kai Zhang, Jingyun Liang, Luc Van Gool, and Radu Timofte. 2021. Design- ing a Practical Degradation Model for Deep Blind Image Super-Resolution. arXiv:2103.14006 [eess.IV] https://arxiv.org/abs/2103.14006

    Pith/arXiv arXiv 2021

  46. [46]

    Lin Zhang, Xin Li, Dongliang He, Errui Ding, and Zhaoxiang Zhang. 2023. LMR: A Large-Scale Multi-Reference Dataset for Reference-based Super-Resolution. arXiv:2303.04970 [cs.CV] https://arxiv.org/abs/2303.04970

    Pith/arXiv arXiv 2023

  47. [47]

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. 2023. Adding Conditional Control to Text-to-Image Diffusion Models. arXiv:2302.05543 [cs.CV] https: //arxiv.org/abs/2302.05543

    Pith/arXiv arXiv 2023

  48. [48]

    Zhifei Zhang, Zhaowen Wang, Zhe Lin, and Hairong Qi. 2019. Image Super- Resolution by Neural Texture Transfer. arXiv:1903.00834 [cs.CV] https://arxiv. org/abs/1903.00834 MicroZoom: Structure-Preserving Detail Synthesis at Extreme Scale•9 IPAdapter Chain-of-ZoomContinuousSRInput (Bicubic)Full Image (green box = crop) Close-up (Reference) UltraZoom Broccoli...

    Pith/arXiv arXiv 2019

  49. [2020]

    arXiv:2006.04139 [cs.CV] https://arxiv.org/abs/2006.04139

    Learning Texture Transformer Network for Image Super-Resolution. arXiv:2006.04139 [cs.CV] https://arxiv.org/abs/2006.04139

    Pith/arXiv arXiv 2006