Pith. sign in

REVIEW 3 major objections 6 minor 65 references

Blueprint-guided diffusion parallelizes artwork outpainting 2.4×

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

2026-07-08 14:35 UTC pith:MQTGPLSD

load-bearing objection Blueprint-guided parallel outpainting is a solid architectural contribution, but headline FID gains are inflated by an evaluation asymmetry the authors acknowledge but underweight. the 3 major comments →

arxiv 2607.06162 v1 pith:MQTGPLSD submitted 2026-07-07 cs.CV

High-Resolution Artwork Outpainting with Global Blueprint Guidance and Layout Control

Junha Kim , Hyunjoon Park , Donghyeon Cho This is my paper
classification cs.CV
keywords image outpaintingdiffusion modelsartwork generationlayout controlparallel synthesisglobal blueprintlow-frequency preservation
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.

The paper proposes a two-stage diffusion framework for extending artwork images beyond their original borders at high resolution. The core idea is to decouple global planning from local detail synthesis: a low-resolution global blueprint is first generated with layout conditions (bounding boxes and object descriptions), then high-resolution patches are synthesized in parallel rather than sequentially. This decoupling addresses three problems the authors identify in existing progressive outpainting methods: error accumulation from sequential generation, lack of spatial controllability beyond text prompts, and high inference latency. The mechanism that enables parallel synthesis is the exploitation of a property called low-frequency preservation in forward diffusion: when noise is added to the blueprint image, coarse structural information (layout, object placement) survives even as fine details are destroyed. Each high-resolution patch is initialized from the corresponding region of this noisy blueprint, so all patches begin denoising from structurally aligned states without needing to wait for neighboring patches to finish. Additionally, the blueprint's intermediate U-Net features are cached and injected into each patch's denoising process as global guidance, and a position token tells each patch where it sits on the full canvas. The authors demonstrate improved FID scores (9.31 vs 10.30 for the strongest baseline), a 2.4× inference speedup on 8 GPUs, and unique layout controllability where users can specify object locations in the expanded region.

Core claim

The central discovery is that a low-resolution global blueprint can serve as a shared structural guide for parallel high-resolution patch synthesis, replacing the sequential generation chain used by prior progressive outpainting methods. The key technical insight enabling this is that forward diffusion applied to the blueprint preserves low-frequency structural information (global layout and object placement) while suppressing high-frequency details (textures and edges), so each patch can be initialized from a structurally aligned noisy state derived from the blueprint and then denoised independently. Combined with cached guidance features from the blueprint generation pass and a position-编码

What carries the argument

low-frequency preservation of forward diffusion

Load-bearing premise

The main comparison table gives the proposed method layout annotations (bounding boxes and object descriptions) derived from the full target image, while baselines receive only text prompts; the supplementary acknowledges this is a conditional comparison with additional structural priors, not a strictly source-only comparison.

What would settle it

If parallel patches initialized from the blueprint via forward diffusion produced visibly incoherent results across patch boundaries without overlap blending, the core claim that the blueprint enables parallel synthesis would fail.

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

Share X LinkedIn Reddit HN

If this is right

  • The blueprint-then-parallelize paradigm could extend to other spatial generation tasks where sequential dependency currently limits throughput, such as panorama synthesis or large-scale scene generation.
  • The low-frequency preservation property exploited here suggests that any coarse structural plan can be transferred into a diffusion model's initial noise, potentially enabling more flexible initialization strategies in other generation pipelines.
  • The layout-conditioning mechanism, currently unique among artwork outpainting methods, could become a standard interface for creative tools where users need fine-grained control over what appears where in extended canvas regions.
  • The framework's demonstrated transferability to DiT-based backbones (SD3) suggests the two-stage blueprint-guided approach is architecture-agnostic and could be adopted as newer diffusion architectures emerge.

Where Pith is reading between the lines

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

  • The 2.4× speedup claim is inherently multi-GPU-dependent; on a single GPU the method is only modestly faster than the strongest baseline (15.73s vs 18.31s), meaning the practical efficiency benefit requires hardware investment that may not be available in all deployment scenarios.
  • The comparison where the proposed method receives layout annotations derived from the full target image while baselines receive only text prompts means the headline image-quality gap partially reflects input asymmetry rather than purely architectural superiority; the authors' own ablation without layout conditions shows a smaller but still present advantage (FID 9.71 vs 10.30).
  • The low-frequency preservation mechanism creates a natural trade-off parameter (forward-init strength) between global structural fidelity and local detail quality, as shown in Table S9; this suggests the method occupies a point on a Pareto frontier rather than dominating all objectives simultaneously.

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

3 major / 6 minor

Summary. The paper proposes a two-stage diffusion framework for high-resolution artwork outpainting. Stage 1 generates a low-resolution global blueprint using a layout adapter (bounding boxes + object descriptions) integrated into a frozen SD Inpainting U-Net via a Gated Fuser, while caching timestep-wise guidance features. Stage 2 synthesizes high-resolution local patches in parallel, each initialized from the blueprint via forward diffusion (exploiting low-frequency preservation) and guided by the cached features plus a patch position token. The method is evaluated on IconArt against PQDiff, SD Inpainting, PowerPaint, and ProOut, reporting improved FID, layout accuracy (AP/IoU), and 2.4× inference speedup on 8 GPUs. The supplementary provides extensive ablations, sensitivity analyses, and additional evaluation settings.

Significance. The architectural contribution—decoupling global planning from parallel local synthesis via blueprint-derived guidance features and forward-diffusion initialization—is sound and well-motivated. The ablation study (Table 3) properly isolates each component, and supplementary tables (S1, S7–S10) provide useful sensitivity analyses. The framework's transferability to DiT backbones (Table S5) and arbitrary canvas sizes (Table S3) adds practical value. The parallel synthesis design with a shared global blueprint is a reasonable approach to the error-accumulation problem in progressive outpainting. However, the headline quantitative claims are partially compromised by an evaluation asymmetry (see major comments), which reduces the evidential strength of the main comparison.

major comments (3)
  1. Table 1 (main comparison) and Sec. S2.2: The proposed method receives bounding boxes and object descriptions derived from the full target image via Grounded-SAM, while all baselines receive only text prompts. This input asymmetry is acknowledged in Sec. S2.2 ('the main tables should be interpreted as a conditional comparison with additional structural priors, not as a strictly source-only comparison'), but the main paper's Table 1 caption and abstract do not surface this caveat. Table S1 shows that without layout conditions, FID rises from 9.31 to 9.71—still better than ProOut's 10.30, but with a narrowed gap (~0.6 vs ~1.0). The headline FID of 9.31 is therefore partially attributable to the additional structural prior rather than architectural superiority alone. The authors should either (a) promote the without-layout FID (9.71) as the primary headline number in the main paper, or (b)显著
  2. Table 1, AP/IoU columns: The layout accuracy comparison (AP 0.4336, IoU 0.6382 for the proposed method vs. ~0.36/~0.56 for baselines) is inherently uninformative as a method comparison because no baseline has a layout conditioning mechanism. Reporting these columns alongside image-quality metrics in the same table invites a misleading apples-to-oranges reading. The authors should either move these metrics to a separate table framed as a capability demonstration (not a competitive comparison), or explicitly state in the table caption that baselines lack layout conditioning and that these columns are not a head-to-head comparison.
  3. Sec. S2.6 / Table 1: The ProOut baseline is an unofficial ControlNet-based reproduction ('No official training code is publicly available for ProOut. We therefore implement an unofficial ControlNet-based reproduction'). The 10.30 FID reference point is therefore uncertain. Table S11 shows ProOut FID fluctuating between 10.28 and 10.44 across epochs 30–60, suggesting the baseline is near convergence, but the absence of an official reference makes it impossible to assess whether the reproduction is faithful. The authors should explicitly acknowledge this as a limitation in the main paper (not only the supplementary) and qualify the magnitude of the FID improvement accordingly.
minor comments (6)
  1. Sec. 3.3, Eq. (5): The Gated Fuser is described as injecting instance tokens C into the U-Net, but the equation also produces F^g_t as its output. It would help to clarify whether F^g_t is the same as the fused output H^g_t + alpha_g * Attn(...), or a separate cached quantity, since the notation suggests they are identical but the text describes them as serving different purposes (injection vs. guidance export).
  2. Table 2: The CLIP-A score for the proposed method at 600% ratio (6.31) is lower than at 400% (6.53) and 200% (6.76), and also lower than ProOut at 600% (6.41). The text states the method 'maintains stable visual quality even at extreme masking ratios,' but this decline in CLIP-A is not discussed. A brief note acknowledging this would be appropriate.
  3. Sec. 4.1: The batch size is stated as 512, but Sec. S2.4 states 480 for Stage 1 and 520 for Stage 2. The main paper should be consistent with the supplementary.
  4. Table S9: The forward-init strength of 0.50 yields better FID (9.03), AP (0.4979), and IoU (0.7013) than the default 1.0, yet 1.0 is used as the default. The text explains this as a trade-off favoring patch-level fidelity, but the justification is brief. A clearer rationale for why 1.0 is preferred despite worse global metrics would help readers.
  5. Fig. S14: The comparison with Nano Banana Pro uses different conditioning interfaces (layout map image + text for Nano Banana Pro vs. explicit bounding boxes for the proposed method). The caption should note that this is an informal qualitative comparison rather than a controlled experiment.
  6. Sec. S2.8: The paper notes that Grounded-SAM is used for annotation while Grounding DINO is used for layout metric evaluation, and that both belong to the same detector family. This is a valid concern. A sentence acknowledging this as a potential bias in the AP/IoU metrics would strengthen the discussion.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the careful and constructive review. The referee correctly identifies that the main comparison table involves an input asymmetry (layout conditions for our method vs. text-only baselines), that the AP/IoU columns are not a head-to-head comparison since no baseline has layout conditioning, and that the ProOut baseline is an unofficial reproduction. We agree with all three points and will revise the manuscript to surface these caveats in the main paper rather than only in the supplementary. We provide point-by-point responses below.

read point-by-point responses

  1. Referee: Table 1 input asymmetry: our method receives bounding boxes and object descriptions from the full target image via Grounded-SAM while baselines receive only text prompts. The caveat is in Sec. S2.2 but not in the main paper's Table 1 caption or abstract. The referee suggests either promoting the without-layout FID (9.71) as the primary headline or clearly disclosing the asymmetry in the main paper.

    Authors: The referee is correct that the main paper does not adequately surface this caveat. We acknowledge that the headline FID of 9.31 benefits from the additional structural prior, and that the without-layout FID of 9.71 narrows the gap with ProOut from ~1.0 to ~0.6 FID points. We will revise the main paper as follows: (1) add an explicit note in the Table 1 caption stating that our method receives layout conditions derived from the target image while baselines receive text prompts only; (2) mention the without-layout FID (9.71) in the main results discussion so readers can assess performance under matched input conditions; (3) add a brief caveat in the abstract noting that the comparison includes layout-conditioned results. We note that even without layout conditions, our method still outperforms all baselines in FID (9.71 vs. 10.30 for ProOut), and the 2.4x speedup from parallel synthesis is independent of the layout condition. However, we agree the magnitude of the improvement should be qualified in the main text. revision: yes

  2. Referee: AP/IoU columns in Table 1 are inherently uninformative as a method comparison because no baseline has a layout conditioning mechanism. Reporting these alongside image-quality metrics invites a misleading apples-to-oranges reading. The referee suggests moving these to a separate table or explicitly stating in the caption that baselines lack layout conditioning.

    Authors: We agree that the AP/IoU columns are not a head-to-head comparison since no baseline supports layout conditioning. We considered moving them to a separate table, but we believe the more transparent solution is to retain them in Table 1 with an explicit caption note, because removing them would obscure the fact that these metrics are included to demonstrate a capability unique to our framework rather than to claim superiority over baselines on layout accuracy. We will add a clear statement in the Table 1 caption: 'AP and IoU are reported only for our method and baselines that produce detectable objects; since no baseline incorporates a layout conditioning mechanism, these columns demonstrate a capability of our framework rather than a head-to-head comparison.' We will also add a corresponding sentence in the main text discussion of Table 1. revision: yes

  3. Referee: ProOut baseline is an unofficial ControlNet-based reproduction with no official reference, making the 10.30 FID reference point uncertain. Table S11 shows FID fluctuating between 10.28 and 10.44 across epochs 30-60, but without an official reference it is impossible to assess faithfulness. The referee asks to acknowledge this as a limitation in the main paper and qualify the magnitude of the FID improvement.

    Authors: This is a fair and accurate concern. We will add an explicit limitation note in the main paper (Section 4.1, Implementation Details) stating that ProOut is reproduced via an unofficial ControlNet-based implementation because no official training code is publicly available, and that the FID reference point is therefore uncertain. We will also qualify the FID improvement magnitude in the results discussion, noting that the ProOut reproduction may not reflect the true performance of an official implementation. We would like to note that Table S11 shows ProOut FID is stable across epochs 30-60 (10.28-10.44), suggesting our reproduction is at least internally consistent, but we agree that without an official reference the absolute faithfulness cannot be verified. We also note that our method's improvement holds across multiple independent evaluation settings (Tables 2, S3, S4, S7) where ProOut serves as a baseline, which provides some robustness to the uncertainty in any single ProOut reproduction. revision: yes

Circularity Check

0 steps flagged

No significant circularity. The framework is built from externally published components with standard training objectives; no prediction reduces to a fitted input by construction. The only concern is an evaluation asymmetry (layout priors given to the proposed method but not baselines), which is an experimental fairness issue, not a circularity issue.

full rationale

The paper proposes a two-stage diffusion framework for artwork outpainting. Stage 1 (Eq. 6) and Stage 2 (Eq. 11) use standard noise-prediction L2 losses with frozen U-Net backbones and trainable adapter modules—no objective is defined in terms of the metric it claims to predict. The architectural components (SD Inpainting, InstanceDiffusion's Gated Fuser, AlignNoise, FreeInit, MultiDiffusion) are all externally published and independently validated. The low-frequency preservation property (Sec. 3.1) is cited to FreeInit [49] and prior analyses [29], not to the authors' own prior work. The blueprint-to-patch initialization (Eq. 10) is a straightforward application of the forward diffusion equation (Eq. 1) to the Stage 1 output, not a self-referential definition. The layout accuracy metrics (AP/IoU in Table 1) are computed via Grounding DINO detection against conditioning boxes, which is an external evaluation pipeline, not a tautological comparison. The supplementary (Sec. S2.2) transparently acknowledges that the main comparison gives the proposed method target-derived layout annotations while baselines receive only text prompts, and Table S1 provides the without-layout ablation (FID 9.71 vs 9.31). This is an evaluation fairness concern—properly flagged under correctness risk—but it is not circularity: the FID metric is computed against real ground-truth images, not against the layout input itself, so the headline number is not 'equivalent to the input by construction.' No step in the derivation chain reduces to its own inputs by definition, no fitted parameter is renamed as a prediction, and no self-citation chain forces the conclusion. The 2-point score reflects the minor concern that the layout-conditioned results in Table 1 are partially attributable to the extra input signal, but this is a proportionate experimental-design issue rather than a structural circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 2 invented entities

The framework introduces 5 learned/chosen parameters and relies on 4 domain assumptions from prior work (FreeInit, AlignNoise, MultiDiffusion, SD Inpainting). The two invented entities (guidance feature bank and position token) are both supported by ablation evidence. No ad-hoc axioms or entities are introduced without falsifiable handles.

free parameters (5)
  • alpha_g (Stage 1 gate) = zero-initialized, learned
    Zero-initialized learnable gate parameter for the Gated Fuser in Stage 1 (Eq. 5), trained on artwork data.
  • alpha_p (Stage 2 gate) = zero-initialized, learned
    Zero-initialized learnable gate parameter for the Gated Fuser in Stage 2 (Eq. 8), trained on artwork data.
  • n_null (null token) = learnable
    Learnable null token replacing CLIP embedding in the Position Network (Eq. 7), trained during Stage 2.
  • CFG scale = 3.0
    Classifier-free guidance scale, chosen by sensitivity analysis (Table S8).
  • forward-init strength = 1.0
    Forward diffusion strength for blueprint initialization, chosen by sensitivity analysis (Table S9).
axioms (4)
  • domain assumption Low-frequency components of an image survive forward diffusion to timestep T under standard noise schedules.
    Invoked in Sec. 3.1 based on FreeInit [29,49]. This property is the basis for blueprint-based initial noise construction (Eq. 10).
  • domain assumption Self-attention maps from the first denoising step can be optimized to ensure source-target semantic consistency.
    Invoked in Sec. 3.3 based on AlignNoise [24]. Used for attention-guided noise optimization before Stage 1 denoising.
  • domain assumption Latent-space averaging of overlapping patches at each denoising step produces seamless composition.
    Invoked in Sec. 3.5 (Eq. 13) based on MultiDiffusion [2]. Used for overlap blending during parallel synthesis.
  • domain assumption Stable Diffusion v1.5 Inpainting U-Net provides a sufficient frozen backbone for artwork outpainting.
    Both stages freeze the SD v1.5 Inpainting backbone and train only adapter modules. Implicit assumption that this backbone generalizes to artwork domains.
invented entities (2)
  • Global blueprint guidance feature bank F independent evidence
    purpose: Stores timestep-wise Gated Fuser output features from Stage 1 for injection into Stage 2 U-Net blocks.
    Ablation in Table 3 shows removing F degrades FID from 9.31 to 22.23, providing falsifiable evidence of its contribution.
  • Patch position token p_i independent evidence
    purpose: Encodes spatial coordinates of each patch on the full canvas to coordinate parallel synthesis with the blueprint.
    Ablation in Table 3 shows removing p_i degrades FID from 9.31 to 17.96, providing falsifiable evidence of its contribution.

reviewed 2026-07-08 · how reviews work

0 comments
Cite this review

Pith. "Pith review of High-Resolution Artwork Outpainting with Global Blueprint Guidance and Layout Control." pith.science (2026). https://pith.science/paper/MQTGPLSD

@misc{pith2026260706162,
  author       = {Pith},
  title        = {Pith review of: High-Resolution Artwork Outpainting with Global Blueprint Guidance and Layout Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MQTGPLSD}},
  note         = {Machine review of arXiv:2607.06162}
}
Share X LinkedIn Reddit HN
read the original abstract

Image outpainting extends an image beyond its original borders, requiring seamless style integration and globally coherent scene completion. Building on the success of diffusion models, recent methods have achieved substantial improvements in visual quality. In practice, however, high-resolution outpainting is commonly performed via progressive expansion around a fixed source image, particularly in artwork scenarios. Despite this progress, existing approaches still suffer from three key limitations: (i) the absence of a reliable global planning mechanism, which leads to structural instability and error accumulation at high resolutions; (ii) limited spatial controllability beyond text prompts, making it difficult to place objects at user-specified locations; and (iii) high inference latency caused by inherently sequential patch generation. To address these issues, we propose a global blueprint-guided two-stage diffusion framework for layout-controllable high-resolution outpainting with efficient parallel synthesis. In Stage 1, we generate a low-resolution global blueprint using a layout adapter that injects bounding-box conditions into a Stable Diffusion inpainting backbone, producing a globally consistent structural plan while extracting global guidance features. In Stage 2, we synthesize high-resolution local patches in parallel by injecting the blueprint-derived global guidance and initializing each patch from the blueprint using the low-frequency preservation property of forward diffusion. This design eliminates sequential dependency while maintaining global coherence. Extensive experiments on large-scale artwork datasets demonstrate improved visual fidelity, stronger semantic consistency, and substantially reduced inference time compared to prior baselines, while uniquely supporting explicit layout control for artwork outpainting.

Figures

Figures reproduced from arXiv: 2607.06162 by Donghyeon Cho, Hyunjoon Park, Junha Kim.

Figure 1
Figure 1. Figure 1: Challenges in High-resolution Artwork Outpainting. (a) [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of Our Two-stage Framework. In Stage 1, the Attention Align￾ment Measure (AAM) first optimizes the initial noise to prevent semantic disconnection from the source image. The Layout Adapter then encodes user-specified bounding boxes bk and object descriptions dk into instance tokens C. The Gated Fuser injects C into the global U-Net to generate a low-resolution blueprint xˆ g while exporting guidan… view at source ↗
Figure 3
Figure 3. Figure 3: Parallel Synthesis of High-resolution Local Patches. (Left) [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative Results. The red dashed box in the GT indicates the source region. The area outside the box is masked during inference and serves as the tar￾get region to be generated. PQDiff produces blurry results and SD-Inpainting creates visible seams, while PowerPaint and ProOut often introduce artificial frames or out￾of-context. In contrast, our method synthesizes seamless, high-fidelity extensions that… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative Results on Layout-guided Generation. [PITH_FULL_IMAGE:figures/full_fig_p014_5.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

65 extracted references · 65 canonical work pages · 4 internal anchors

  1. [1]

    Art Institute of Chicago: Open access collection.https : / / www . artic . edu / collection, accessed: 2026-02-11

    work page 2026

  2. [2]

    In: International Conference on Machine Learning

    Bar-Tal, O., Yariv, L., Lipman, Y., Dekel, T.: Multidiffusion: Fusing diffusion paths for controlled image generation. In: International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 202, pp. 1737–1752. PMLR (2023)

    work page 2023

  3. [3]

    In: European Conference on Computer Vi- sion

    Chai, L., Gharbi, M., Shechtman, E., Isola, P., Zhang, R.: Any-resolution train- ing for high-resolution image synthesis. In: European Conference on Computer Vi- sion. Lecture Notes in Computer Science, vol. 13676, pp. 170–188. Springer (2022). https://doi.org/10.1007/978-3-031-19787-1_10

    work page doi:10.1007/978-3-031-19787-1_10 2022

  4. [4]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Chen, Q., Ma, Y., Wang, H., Yuan, J., Zhao, W., Tian, Q., Wang, H., Min, S., Chen, Q., Liu, W.: Infinite-Canvas: Higher-resolution video outpainting with ex- tensive content generation. In: Proceedings of the AAAI Conference on Artificial Intelligence. pp. 2150–2158 (2025).https://doi.org/10.1609/AAAI.V39I2.32213

    work page doi:10.1609/aaai.v39i2.32213 2025

  5. [5]

    Cross-dimension affinity distillation for 3d em neuron segmentation,

    Chen, Y., Wang, O., Zhang, R., Shechtman, E., Wang, X., Gharbi, M.: Image neural field diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 8007–8017 (2024).https://doi. org/10.1109/CVPR52733.2024.00765

    work page doi:10.1109/cvpr52733.2024.00765 2024

  6. [6]

    In: 2022 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR)

    Cheng,Y.C.,Lin,C.H.,Lee,H.Y.,Ren,J.,Tulyakov,S.,Yang,M.H.:Inout:Diverse image outpainting via gan inversion. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 11421–11430 (2022).https: //doi.org/10.1109/CVPR52688.2022.01114

    work page doi:10.1109/cvpr52688.2022.01114 2022

  7. [7]

    In: European Conference on Computer Vi- sion

    Choi, Y., Kwak, S., Lee, K., Choi, H., Shin, J.: Improving diffusion models for authentic virtual try-on in the wild. In: European Conference on Computer Vi- sion. Lecture Notes in Computer Science, vol. 15144, pp. 206–235. Springer (2024). https://doi.org/10.1007/978-3-031-73016-0_13

    work page doi:10.1007/978-3-031-73016-0_13 2024

  8. [8]

    com / ClevelandMuseumArt/openaccess, accessed: 2026-02-11

    Cleveland Museum of Art: Open access collection.https : / / github . com / ClevelandMuseumArt/openaccess, accessed: 2026-02-11

    work page 2026

  9. [9]

    In: International Conference on Machine Learn- ing

    Esser, P., Kulal, S., Blattmann, A., Entezari, R., Müller, J., Saini, H., Levi, Y., Lorenz, D., Sauer, A., Boesel, F., et al.: Scaling rectified flow transformers for high-resolution image synthesis. In: International Conference on Machine Learn- ing. Proceedings of Machine Learning Research, vol. 235, pp. 12606–12633. PMLR (2024)

    work page 2024

  10. [10]

    Exhibition Hub, Fever: Van gogh: The immersive experience.https : / / vangoghexpo.com, accessed: 2026-02-11

    work page 2026

  11. [11]

    In: Proceedings of the 31st ACM International Conference on Multimedia

    Fan, F., Guo, C., Gong, L., Wang, B., Ge, T., Jiang, Y., Luo, C., Zhan, J.: Hi- erarchical masked 3d diffusion model for video outpainting. In: Proceedings of the 31st ACM International Conference on Multimedia. pp. 7890–7900 (2023). https://doi.org/10.1145/3581783.3612478

    work page doi:10.1145/3581783.3612478 2023

  12. [12]

    Neural Networks162, 1–10 (2023).https://doi.org/10.1016/J.NEUNET.2023.02.021

    Gao, P., Yang, X., Zhang, R., Goulermas, J.Y., Geng, Y., Yan, Y., Huang, K.: Generalized image outpainting with u-transformer. Neural Networks162, 1–10 (2023).https://doi.org/10.1016/J.NEUNET.2023.02.021

    work page doi:10.1016/j.neunet.2023.02.021 2023

  13. [13]

    In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops

    Gonthier, N., Gousseau, Y., Ladjal, S., Bonfait, O.: Weakly supervised object de- tection in artworks. In: Proceedings of the European Conference on Computer Vision (ECCV) Workshops. Lecture Notes in Computer Science, vol. 11130, pp. 692–709. Springer (2018).https://doi.org/10.1007/978-3-030-11012-3_53

    work page doi:10.1007/978-3-030-11012-3_53 2018

  14. [14]

    google/models/gemini-image/pro/, accessed: 2026-06-29 Blueprint-Guided Artwork Outpainting 17

    Google DeepMind: Nano Banana Pro (Gemini 3 Pro Image).https://deepmind. google/models/gemini-image/pro/, accessed: 2026-06-29 Blueprint-Guided Artwork Outpainting 17

    work page 2026

  15. [15]

    In: Advances in Neural Information Processing Systems

    Heusel, M., Ramsauer, H., Unterthiner, T., Nessler, B., Hochreiter, S.: GANs trained by a two time-scale update rule converge to a local nash equilibrium. In: Advances in Neural Information Processing Systems. vol. 30, pp. 6626–6637 (2017)

    work page 2017

  16. [16]

    Journal of Machine Learning Research23(47), 1–33 (2022)

    Ho, J., Saharia, C., Chan, W., Fleet, D.J., Norouzi, M., Salimans, T.: Cascaded diffusion models for high fidelity image generation. Journal of Machine Learning Research23(47), 1–33 (2022)

    work page 2022

  17. [17]

    Are NeRFs Ready for Autonomous Driving? Towards Closing the Real-to-Simulation Gap

    Hsu, C.C., Lee, C.M., Chou, Y.S.: DRCT: Saving image super-resolution away from information bottleneck. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops. pp. 6133–6142 (2024). https://doi.org/10.1109/CVPRW63382.2024.00618

    work page doi:10.1109/cvprw63382.2024.00618 2024

  18. [19]

    VideoMAE V2: Scaling Video Masked Autoencoders with Dual Masking

    Ju, X., Zeng, A., Wang, J., Xu, Q., Zhang, L.: Human-Art: A versatile human- centric dataset bridging natural and artificial scenes. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 618– 629 (2023).https://doi.org/10.1109/CVPR52729.2023.00067

    work page doi:10.1109/cvpr52729.2023.00067 2023

  19. [20]

    Emerging

    Khurana, B., Dash, S.R., Bhatia, A., Mahapatra, A., Singh, H., Kulkarni, K.: Semie: Semantically-aware image extrapolation. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 14880–14889 (2021).https: //doi.org/10.1109/ICCV48922.2021.01463

    work page doi:10.1109/iccv48922.2021.01463 2021

  20. [21]

    Berg, Wan-Yen Lo, Piotr Doll ´ar, and Ross B

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., et al.: Segment anything. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3992–4003 (2023).https://doi.org/10.1109/ICCV51070.2023.00371

    work page doi:10.1109/iccv51070.2023.00371 2023

  21. [22]

    In: International Conference on Machine Learning

    Li, J., Li, D., Savarese, S., Hoi, S.: BLIP-2: Bootstrapping language-image pre- training with frozen image encoders and large language models. In: International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 202, pp. 19730–19742. PMLR (2023)

    work page 2023

  22. [23]

    In: International Con- ference on Machine Learning

    Li, J., Li, D., Xiong, C., Hoi, S.: BLIP: bootstrapping language-image pre-training for unified vision-language understanding and generation. In: International Con- ference on Machine Learning. Proceedings of Machine Learning Research, vol. 162, pp. 12888–12900. PMLR (2022)

    work page 2022

  23. [24]

    In: Proceedings of the 33rd ACM International Conference on Multimedia

    Li, N., Li, Z., Tang, Z., Yu, Y., Zou, L., Li, C.: Bridging the gap: Consistent image outpainting via training-free noise optimization. In: Proceedings of the 33rd ACM International Conference on Multimedia. pp. 9969–9977 (2025).https://doi.org/ 10.1145/3746027.3755278

    work page doi:10.1145/3746027.3755278 2025

  24. [25]

    In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision

    Li, Y., Jiang, L., Yang, M.H.: Controllable and progressive image extrapolation. In: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. pp. 2139–2148 (2021).https://doi.org/10.1109/WACV48630.2021.00219

    work page doi:10.1109/wacv48630.2021.00219 2021

  25. [26]

    VideoMAE V2: Scaling Video Masked Autoencoders with Dual Masking

    Li, Y., Liu, H., Wu, Q., Mu, F., Yang, J., Gao, J., Li, C., Lee, Y.J.: Gligen: Open-set grounded text-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22511–22521 (2023). https://doi.org/10.1109/CVPR52729.2023.02156

    work page doi:10.1109/cvpr52729.2023.02156 2023

  26. [27]

    In: Advances in Neural Information Processing Systems

    Liang, J., Wu, C., Hu, X., Gan, Z., Wang, J., Wang, L., Liu, Z., Fang, Y., Duan, N.:NUWA-Infinity:Autoregressiveoverautoregressivegenerationforinfinitevisual synthesis. In: Advances in Neural Information Processing Systems. vol. 35 (2022)

    work page 2022

  27. [28]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision 18 J

    Lin, H., Pagnucco, M., Song, Y.: Edge guided progressively generative image outpainting. In: Proceedings of the IEEE/CVF Conference on Computer Vision 18 J. Kim et al. and Pattern Recognition Workshops. Lecture Notes in Computer Science, vol. 13683, pp. 806–815. Springer (2021).https://doi.org/10.1109/CVPRW53098. 2021.00090

    work page doi:10.1109/cvprw53098 2021

  28. [29]

    Lin,S.,Liu,B.,Li,J.,Yang,X.:Commondiffusionnoiseschedulesandsamplesteps areflawed.In:ProceedingsoftheIEEE/CVFWinterConferenceonApplicationsof Computer Vision. pp. 5392–5399 (2024).https://doi.org/10.1109/WACV57701. 2024.00532

    work page doi:10.1109/wacv57701 2024

  29. [30]

    StyleCrafter: Enhancing Stylized Text-to-Video Generation with Style Adapter

    Liu, G., Xia, M., Zhang, Y., Chen, H., Xing, J., Wang, X., Yang, Y., Shan, Y.: StyleCrafter: Enhancing stylized text-to-video generation with style adapter. arXiv preprintarXiv:2312.00330(2023).https://doi.org/10.48550/arXiv.2312.00330

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2312.00330 2023

  30. [31]

    In: Advances in Neural Infor- mation Processing Systems

    Liu, K., Zhu, Z., Li, C., Liu, H., Zeng, H., Hou, J.: Prefpaint: Aligning image inpainting diffusion model with human preference. In: Advances in Neural Infor- mation Processing Systems. vol. 37, pp. 30554–30589 (2024)

    work page 2024

  31. [32]

    In: European Conference on Computer Vision

    Liu, S., Zeng, Z., Ren, T., Li, F., Zhang, H., Yang, J., Jiang, Q., Li, C., Yang, J., Su, H., et al.: Grounding DINO: Marrying DINO with grounded pre-training for open-set object detection. In: European Conference on Computer Vision. Lecture Notes in Computer Science, vol. 15105, pp. 38–55. Springer (2024).https://doi. org/10.1007/978-3-031-72970-6_3

    work page doi:10.1007/978-3-031-72970-6_3 2024

  32. [33]

    In: International Conference on Learning Representations

    Loshchilov, I., Hutter, F.: Decoupled weight decay regularization. In: International Conference on Learning Representations. OpenReview.net (2019)

    work page 2019

  33. [34]

    Boosting Image Outpainting with Semantic Layout Prediction

    Ma, Y., Ma, J., Zhou, M., Chen, Q., Ge, T., Jiang, Y., Lin, T.: Boosting im- age outpainting with semantic layout prediction. arXiv preprint arXiv:2110.09267 (2021)

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  34. [35]

    In: International Conference on Machine Learning

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., et al.: Learning transferable visual models from natural language supervision. In: International Conference on Machine Learning. Proceedings of Machine Learning Research, vol. 139, pp. 8748–8763. PMLR (2021)

    work page 2021

  35. [36]

    In: 2022 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR)

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 10674–10685 (2022). https://doi.org/10.1109/CVPR52688.2022.01042

    work page doi:10.1109/cvpr52688.2022.01042 2022

  36. [37]

    Painting Outside the Box: Image Outpainting with GANs

    Sabini, M., Rusak, G.: Painting outside the box: Image outpainting with gans. arXiv preprint arXiv:1808.08483 (2018)

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  37. [38]

    In: Advances in Neural Information Processing Systems

    Schuhmann, C., Beaumont, R., Vencu, R., Gordon, C., Wightman, R., Cherti, M., Coombes, T., Katta, A., Mullis, C., Wortsman, M., et al.: LAION-5B: An open large-scale dataset for training next generation image-text models. In: Advances in Neural Information Processing Systems. vol. 35, pp. 25278–25294 (2022)

    work page 2022

  38. [39]

    DriveArena: A Closed-loop Generative Simulation Platform for Autonomous Driving

    Song, D.Y., Yu, J.J., Cho, D.: Progressive artwork outpainting via latent diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 15405–15415 (2025).https://doi.org/10.1109/ICCV51701.2025. 01429

    work page doi:10.1109/iccv51701.2025 2025

  39. [40]

    OpenReview.net (2021)

    Song,J.,Meng,C.,Ermon,S.:Denoisingdiffusionimplicitmodels.In:International Conference on Learning Representations. OpenReview.net (2021)

    work page 2021

  40. [41]

    In: IEEE Conf

    Sun, L., Wu, R., Ma, Z., Liu, S., Yi, Q., Zhang, L.: Pixel-level and semantic- level adjustable super-resolution: A Dual-LoRA approach. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 2333– 2343 (2025).https://doi.org/10.1109/CVPR52734.2025.00223

    work page doi:10.1109/cvpr52734.2025.00223 2025

  41. [42]

    IEEE Transactions on Image Pro- cessing28(1), 394–409 (2019).https://doi.org/10.1109/TIP.2018.2866698 Blueprint-Guided Artwork Outpainting 19

    Tan, W.R., Chan, C.S., Aguirre, H.E., Tanaka, K.: Improved ArtGAN for condi- tional synthesis of natural image and artwork. IEEE Transactions on Image Pro- cessing28(1), 394–409 (2019).https://doi.org/10.1109/TIP.2018.2866698 Blueprint-Guided Artwork Outpainting 19

    work page doi:10.1109/tip.2018.2866698 2019

  42. [43]

    In: Advances in Neural In- formation Processing Systems

    Tancik, M., Srinivasan, P., Mildenhall, B., Fridovich-Keil, S., Raghavan, N., Sing- hal, U., Ramamoorthi, R., Barron, J., Ng, R.: Fourier features let networks learn high frequency functions in low dimensional domains. In: Advances in Neural In- formation Processing Systems. vol. 33, pp. 7537–7547 (2020)

    work page 2020

  43. [44]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Teterwak, P., Sarna, A., Krishnan, D., Maschinot, A., Belanger, D., Liu, C., Free- man, W.T.: Boundless: Generative adversarial networks for image extension. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 10520–10529 (2019).https://doi.org/10.1109/ICCV.2019.01062

    work page doi:10.1109/iccv.2019.01062 2019

  44. [45]

    DriveArena: A Closed-loop Generative Simulation Platform for Autonomous Driving

    Tsai, S.R., Chang, W.C., Lee, J.Y., Su, C.H., Liu, Y.L.: Lightsout: Diffusion-based outpainting for enhanced lens flare removal. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 6353–6363 (2025).https:// doi.org/10.1109/ICCV51701.2025.00599

    work page doi:10.1109/iccv51701.2025.00599 2025

  45. [46]

    Cross-dimension affinity distillation for 3d em neuron segmentation,

    Wang, X., Darrell, T., Rambhatla, S.S., Girdhar, R., Misra, I.: Instancediffusion: Instance-level control for image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6232–6242 (2024). https://doi.org/10.1109/CVPR52733.2024.00596

    work page doi:10.1109/cvpr52733.2024.00596 2024

  46. [47]

    Wang, Y., Wei, Y., Qian, X., Zhu, L., Yang, Y.: Sketch-guided scenery image outpainting.IEEETransactionsonImageProcessing30,2643–2655(2021).https: //doi.org/10.1109/TIP.2021.3054477

    work page doi:10.1109/tip.2021.3054477 2021

  47. [48]

    IEEE Transactions on Image Processing33, 1375–1388 (2024).https://doi.org/10.1109/TIP.2024.3357290

    Wang, Y., Wei, Y., Qian, X., Zhu, L., Yang, Y.: ReGO: Reference-guided out- painting for scenery image. IEEE Transactions on Image Processing33, 1375–1388 (2024).https://doi.org/10.1109/TIP.2024.3357290

    work page doi:10.1109/tip.2024.3357290 2024

  48. [49]

    In: European Conference on Computer Vision

    Wu, T., Si, C., Jiang, Y., Huang, Z., Liu, Z.: FreeInit: Bridging initialization gap in video diffusion models. In: European Conference on Computer Vision. Lecture Notes in Computer Science, vol. 15061, pp. 378–394. Springer (2024).https:// doi.org/10.1007/978-3-031-72646-0_22

    work page doi:10.1007/978-3-031-72646-0_22 2024

  49. [50]

    In: Proceedings of the Asian Conference on Computer Vision

    Xu, Y., Pagnucco, M., Song, Y.: DHG-GAN: Diverse image outpainting via de- coupled high frequency semantics. In: Proceedings of the Asian Conference on Computer Vision. Lecture Notes in Computer Science, vol. 13847, pp. 168–184. Springer (2022).https://doi.org/10.1007/978-3-031-26293-7_11

    work page doi:10.1007/978-3-031-26293-7_11 2022

  50. [51]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Xu, Y., Gu, T., Chen, W., Chen, A.: OOTDiffusion: Outfitting fusion based latent diffusion for controllable virtual try-on. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 39, pp. 8996–9004 (2025).https://doi.org/10.1609/ AAAI.V39I9.32973

    work page 2025

  51. [52]

    Cross-dimension affinity distillation for 3d em neuron segmentation,

    Xu, Z., Zhang, J., Liew, J.H., Yan, H., Liu, J.W., Zhang, C., Feng, J., Shou, M.Z.: MagicAnimate: Temporally consistent human image animation using dif- fusion model. In: Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition. pp. 1481–1490 (2024).https://doi.org/10.1109/ CVPR52733.2024.00147

    work page arXiv 2024

  52. [53]

    In: Proceedings of the AAAI Conference on Artificial Intelligence

    Yan, K., Ji, L., Wu, C., Liang, J., Zhou, M., Duan, N., Ma, S.: HORIZON: High-resolution semantically controlled panorama synthesis. In: Proceedings of the AAAI Conference on Artificial Intelligence. vol. 38, pp. 6431–6439 (2024). https://doi.org/10.1609/AAAI.V38I6.28463

    work page doi:10.1609/aaai.v38i6.28463 2024

  53. [54]

    In: 2022 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR)

    Yang, C.A., Tan, C.Y., Fan, W.C., Yang, C.F., Wu, M.L., Wang, Y.C.F.: Scene graph expansion for semantics-guided image outpainting. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 15596– 15605 (2022).https://doi.org/10.1109/CVPR52688.2022.01517

    work page doi:10.1109/cvpr52688.2022.01517 2022

  54. [55]

    In: Proceedings of the Asian Conference on Computer Vision

    Yang, J., Wang, H., Zhu, Z., Liu, C., Wu, M.W., Sun, M.: VIP: Versatile image outpainting empowered by multimodal large language model. In: Proceedings of the Asian Conference on Computer Vision. Lecture Notes in Computer Science, 20 J. Kim et al. vol. 15476, pp. 59–76. Springer (2024).https://doi.org/10.1007/978-981-96- 0917-8_4

    work page doi:10.1007/978-981-96- 2024

  55. [56]

    In: Proceedings of the IEEE/CVF International Con- ference on Computer Vision

    Yang, Z., Dong, J., Liu, P., Yang, Y., Yan, S.: Very long natural scenery image prediction by outpainting. In: Proceedings of the IEEE/CVF International Con- ference on Computer Vision. pp. 10560–10569 (2019).https://doi.org/10.1109/ ICCV.2019.01066

    work page arXiv 2019

  56. [57]

    In: European Conference on Computer Vision

    Yao, K., Gao, P., Yang, X., Sun, J., Zhang, R., Huang, K.: Outpainting by queries. In: European Conference on Computer Vision. pp. 153–169. Springer (2022). https://doi.org/10.1007/978-3-031-20050-2_10

    work page doi:10.1007/978-3-031-20050-2_10 2022

  57. [58]

    IP-Adapter: Text Compatible Image Prompt Adapter for Text-to-Image Diffusion Models

    Ye, H., Zhang, J., Liu, S., Han, X., Yang, W.: IP-Adapter: Text compati- ble image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721 (2023).https://doi.org/10.48550/arXiv.2308.06721

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2308.06721 2023

  58. [59]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision

    Zhang, L., Rao, A., Agrawala, M.: Adding conditional control to text-to-image diffusion models. In: Proceedings of the IEEE/CVF International Conference on Computer Vision. pp. 3813–3824 (2023).https://doi.org/10.1109/ICCV51070. 2023.00355

    work page doi:10.1109/iccv51070 2023

  59. [60]

    In: International Conference on Learning Representations

    Zhang, S., Huang, J., Zhou, Q., Wang, Z., Wang, F., Luo, J., Yan, J.: Continuous- multiple image outpainting in one-step via positional query and a diffusion-based approach. In: International Conference on Learning Representations. OpenRe- view.net (2024)

    work page 2024

  60. [61]

    IEEE Transactions on Pattern Analysis and Machine Intelligence47(10), 9217–9231 (2025).https:// doi.org/10.1109/TPAMI.2025.3586824

    Zhang, S., Zhou, Q., Wang, Z., Li, H., Yan, J.: Easyoutpainter: One step image outpainting with both continuous multiple and resolution. IEEE Transactions on Pattern Analysis and Machine Intelligence47(10), 9217–9231 (2025).https:// doi.org/10.1109/TPAMI.2025.3586824

    work page doi:10.1109/tpami.2025.3586824 2025

  61. [62]

    VideoMAE V2: Scaling Video Masked Autoencoders with Dual Masking

    Zheng, G., Zhou, X., Li, X., Qi, Z., Shan, Y., Li, X.: Layoutdiffusion: Controllable diffusion model for layout-to-image generation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 22490–22499 (2023). https://doi.org/10.1109/CVPR52729.2023.02154

    work page doi:10.1109/cvpr52729.2023.02154 2023

  62. [63]

    IEEE Transactions on Pattern Analysis and Machine Intelligence47(3), 1714–1728 (2025).https://doi.org/10.1109/ TPAMI.2024.3510752

    Zhou, D., Li, Y., Ma, F., Yang, Z., Yang, Y.: Migc++: Advanced multi-instance generation controller for image synthesis. IEEE Transactions on Pattern Analysis and Machine Intelligence47(3), 1714–1728 (2025).https://doi.org/10.1109/ TPAMI.2024.3510752

    work page arXiv 2025

  63. [64]

    Cross-dimension affinity distillation for 3d em neuron segmentation,

    Zhou, D., Li, Y., Ma, F., Zhang, X., Yang, Y.: Migc: Multi-instance genera- tion controller for text-to-image synthesis. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. pp. 6818–6828 (2024). https://doi.org/10.1109/CVPR52733.2024.00651

    work page doi:10.1109/cvpr52733.2024.00651 2024

  64. [65]

    In: European Conference on Computer Vision

    Zhu, S., Chen, J.L., Dai, Z., Dong, Z., Xu, Y., Cao, X., Yao, Y., Zhu, H., Zhu, S.: Champ: Controllable and consistent human image animation with 3D para- metric guidance. In: European Conference on Computer Vision. Lecture Notes in Computer Science, vol. 15113, pp. 145–162. Springer (2024).https://doi.org/ 10.1007/978-3-031-73001-6_9

    work page doi:10.1007/978-3-031-73001-6_9 2024

  65. [66]

    High-Resolution Artwork Outpainting with Global Blueprint Guidance and Layout Control

    Zhuang, J., Zeng, Y., Liu, W., Yuan, C., Chen, K.: A task is worth one word: Learn- ing with task prompts for high-quality versatile image inpainting. In: European Conference on Computer Vision. Lecture Notes in Computer Science, vol. 15116, pp. 195–211. Springer (2024).https://doi.org/10.1007/978-3-031-73636-0_12 Supplementary Material for “High-Resoluti...

    work page doi:10.1007/978-3-031-73636-0_12 2024

This paper was first reviewed by glm-5.2 on July 8, 2026.