Pith. sign in

REVIEW 4 major objections 4 minor 59 references

Reviving Ancient Paintings via Poem: A Colorization Framework for Aligning Cultural Semantics

T0 review · 4 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read By grounding diffusion colorization in a classical color palette learned from poems, a new framework restores faded ancient paintings with historically faithful colors while preserving their ink structure.

desk verdict Genuinely novel poem-to-palette projection and structure-aware attention, but the unpaired FID/color-WD reference distribution is undefined, leaving the 'historical authenticity' claim ungrounded. read the letter →

arxiv 2607.17638 v1 pith:R5U7ZBRG submitted 2026-07-20 cs.CV

Junming Gao , Biao Zhu , Xiaosong Wang , Tan Tang This is my paper
classification cs.CV
keywords ancientpaintingrestorationpoem-guidedcolorizationdiffusionmodelscolorpalettelearningstructure-awareattentiondigitalheritageculturalsemantics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that classical Chinese poems can serve as a reliable cultural prior for restoring the faded colors of ancient paintings, resolving an ambiguity that purely visual methods cannot. To do this, it introduces a Poetic Painting Projector (P3) that turns poem embeddings into a constrained classical color palette, and a Structure-Aware Semantic Attention (SASA) module that injects those colors only within ink and brushstroke boundaries. The framework is trained on a hybrid dataset of synthetic degradations plus expert-restored artifacts, and it reports state-of-the-art reconstruction metrics, a 54% reduction in color-distribution error, and 65.8% expert preference in a blind study. If correct, this shows that cultural semantics can be operationalized as a geometric constraint on color generation.

What carries the argument

The central mechanism is the pair P3 + SASA. P3 projects poem token embeddings from a frozen text encoder onto a learnable dictionary of 16 classical color bases via cosine attention (a convex combination), pretrained through a poem-to-palette proxy task and then frozen; a native structure latent and a low-frequency color anchor are decoupled to separate ink topology from chromatic traces. SASA computes a structure-aware gate and a boundary-normal field from the structure latent, projects out the cross-boundary component of each cross-attention output, modulates the residual by the gate, and re-injects it; a structure-weighted denoising loss reinforces the same gating during training.

What would settle it

Compare PoemColor's outputs to real conservation-grade before/after pairs of faded paintings; if the predicted colors deviate substantially from spectroscopically measured surviving pigments or from known restorations, while synthetic test pairs remain accurate, the synthetic-degradation premise is falsified.

Watch

Extended reading notes

Core claim

PoemColor establishes that poem-guided diffusion colorization can recover historically plausible colors for ancient Chinese paintings by aligning poetic semantics with a classical color manifold. The core discovery is that the semantic gap between poems and colors can be closed by a pretrained poem-to-palette projection, and that the resulting color condition can be propagated safely by suppressing cross-boundary semantic components and gating injection strength according to painting structure. Across paired expert-restored data and unpaired museum artifacts, the framework dominates all eleven baselines on every reported metric, with the largest gains in color-distribution fidelity, while a

Load-bearing premise

The central load-bearing premise is that the synthetic degradation used for most of the training data faithfully mimics how ancient pigments actually fade; if it does not, the reported gains on real museum artifacts reflect the synthetic distribution rather than historical accuracy.

Editorial extensions

If this is right

  • Poem guidance narrows the ambiguous color hypothesis space for faded paintings, reducing modern semantic bias and over-saturation that plain text conditioning produces.
  • A frozen poem-to-palette prior provides a stable classical color anchor, so diffusion training does not corrupt the learned color routing.
  • Boundary-normal projection and spatial gating keep poem-induced colors inside ink-consistent regions, preventing leakage across weak contours.
  • The hybrid synthetic-plus-expert dataset provides scalable supervision despite the scarcity of real paired restorations.
  • The structure-aware mechanism generalizes across artistic media, as shown by zero-fine-tuning results on Western oil paintings.

Reading between the lines

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

  • If the poetic prior is truly what drives the gains, then minimally editing a poem (e.g., changing seasonal words) should induce predictable, independent palette shifts; this controllability could be tested in a dedicated user study that the paper does not report.
  • The unpaired-museum FID and color-error improvements may reflect the synthetic degradation distribution more than historical accuracy; a validation on real conservation before/after pairs is the decisive test.
  • The expert-restored fine-tuning set's size is never reported, so the method's real-data leverage is unclear; ablating its size would separate synthetic-supervision gains from genuine expert-reference gains.
  • The same architecture could be transferred to other heritage domains with textual traditions (murals, ceramics, tapestries) by building a palette-proxy task from their surviving color vocabularies.
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

4 major / 4 minor

Summary. The paper proposes PoemColor, a ControlNet-based poem-guided colorization framework for ancient Chinese paintings. It introduces a Poetic Painting Projector (P3) trained with a poem-to-palette proxy task to map CLIP text embeddings into a learned classical color basis, and a Structure-Aware Semantic Attention (SASA) module that uses boundary-normal projection and a spatial gate to constrain color propagation along ink structures. A hybrid dataset combines synthetic degradation of well-preserved paintings with a small expert-restored real subset. Experiments report state-of-the-art results on paired and unpaired test sets, ablations of P3/SASA/weighted loss, and an expert user study showing 65.8% preference. The paper's central claim is that poem-aligned semantic projection plus structure-aware propagation materially improves historically faithful colorization.

Significance. If the claims hold, PoemColor is a valuable contribution to the narrow but culturally significant task of ancient painting restoration: it demonstrates that poem-derived semantic priors can be projected into a constrained color space and injected with explicit structural regularization, and the hybrid data strategy addresses a real data-scarcity bottleneck. Strengths include a held-out 30% of expert-colorized artifacts for paired testing, ablations that move in expected directions, a quantified user study with a chi-square test (65.8%, χ²=60.50), and a clear two-stage training recipe with a frozen poem-to-palette projector. However, the strongest quantitative evidence for real-world "historical authenticity" currently rests on an under-specified unpaired evaluation and on the realism of a synthetic degradation model inherited from prior same-group work. These gaps are load-bearing for the central claim and need to be closed before the paper can be accepted.

major comments (4)
  1. [§4.1/Table 1] The unpaired FID (19.84) and ΔColorfulness(WD) (1.84) are the only quantitative evidence for real-world historical authenticity, but the reference distribution is never defined. FID is a distance between two image sets; without knowing the 'real' set, 19.84 is uninterpretable. If the reference is the same faded museum artifacts used as inputs, an identity model would score perfectly and the metric rewards non-restoration. If it is a separate corpus of well-preserved or expert-restored paintings, its source, size, preprocessing, and alignment protocol must be specified. This must be fixed for the central claim to be evaluable.
  2. [§4.1 and §4.2] The expert-colored subset is the only real paired ground truth, yet its size is never reported. The paper states '70% used for fine-tuning and the remaining 30% reserved exclusively for quantitative testing' without giving counts. Given that the expert subset appears small (the paper calls it 'high-quality but limited'), the paired results in Table 1 may be based on very few images. Report the number of paintings, the number of restorers, and the partition sizes, and provide per-metric confidence intervals or at least standard errors on Tables 1–2.
  3. [§3.5 and §4.1] The 'classical color manifold' used by P3 is defined by K-Means palettes extracted from the authors' own training paintings (Eq. 16), and the synthetic degradation strategy is described as 'inspired by PRevivor' and 'validated in PRevivor'—same-group prior work. This makes the evaluation potentially circular: the model is trained and evaluated on colors and degradations derived from the same data distribution. A concrete test would be to compare the learned dictionary against an independent art-historical palette corpus, or to evaluate on a held-out set of real faded paintings with expert-provided color references and report per-case accuracy. As written, the historical-fidelity claim is only as strong as the synthetic degradation's realism, which is asserted rather than independently validated.
  4. [§4.3/Table 2] The ablation table reports FID and Color WD without stating on which evaluation set they are computed—paired expert references, unpaired museum artifacts, or a synthetic test set. Since Table 1 distinguishes 'Paired' and 'Unpaired' settings with different metrics, the ablation values (e.g., 19.84 vs 1.84) cannot be interpreted unless the reference distribution is specified. State explicitly which set and which reference these numbers refer to; otherwise the 'complementary' conclusion for projection + gate is not supported.
minor comments (4)
  1. [§4.2] The visual comparison paragraph references 'sample d and e' and 'sample b/g' but Figure 2 uses a multi-panel layout; please label panels clearly and refer to them by consistent identifiers (e.g., 'Figure 2a'–'2m').
  2. [§4.4] The user study reports 8 experts, 34 paintings, and 272 votes. It would be helpful to report the exact question wording and the rating scale, and to clarify whether experts were shown paired or unpaired inputs. Also, the statement 'across all evaluation dimensions' is not supported by Figure 3, which appears to aggregate preferences rather than showing per-dimension breakdowns.
  3. [§5] The 'Generalization Discussion' on the 50-Artworks dataset would be stronger with quantitative results (e.g., FID/color WD before and after adaptation) and a clear description of how the method was applied to Western oil paintings without fine-tuning. As written, the claim of 'universality' rests on a single visual figure.
  4. [Throughout] Some references are malformed or incomplete: 'chang et al., 2023' appears in lowercase in §4.1, and the L-CAD entry in the references uses inconsistent capitalization. Please proofread the bibliography and citation style.

Circularity Check

2 steps flagged · score 4.0 of 10

Headline gains rest on a self-cited synthetic degradation and a self-defined palette manifold; paired expert metrics provide partial independent support.

  1. self citation load bearing [§4.1 Datasets]
    "Therefore, following the common practice in old photo restoration (Wan et al., 2020), we generate a large-scale synthetic dataset by applying a stochastic degradation strategy inspired by PRevivor (Tang et al., 2025) to well-preserved paintings. This synthetic construction provides scalable supervision for learning fading patterns, and its effectiveness has been validated in PRevivor as well as by our experimental results for training poem-guided colorization."

    PRevivor (Tang et al., 2025) is authored by Tan Tang and Junming Gao, two of the present paper's authors. The synthetic degradation is the source of nearly all training pairs (the expert-colored subset is used only for fine-tuning and its size is never reported). The paper's only external-sounding validation of the degradation's realism is the self-citation 'validated in PRevivor'; the appended 'as well as by our experimental results' refers to experiments on the same hybrid dataset. Thus the central training infrastructure's realism rests on a same-group citation rather than on an independently established degradation model.

  2. self definitional [§3.5 Eq. (16)-(17) and §1]
    "For each ancient painting, we extract Kp = 8 dominant HSV colors by K-Means and represent the sorted palette as: c= [h1, s1, v1, . . . , hKp , sKp , vKp ]∈R 3Kp ... P3 does not directly inject unconstrained text embeddings; instead, it projects poetic context into a historically constrained palette subspace."

    The 'historically constrained palette subspace' / 'classical color manifold' is not an externally given historical ground truth; it is the K-Means palette summary of the very paintings used to train P3. Because P3 is pretrained to regress these palettes from poems (Lpalette) and then frozen as the 'poetic color prior', the claim that poem guidance is aligned to 'classical' colors is true by construction relative to the training set. Any success at 'historical authenticity' measured by color-distribution distance is therefore a measure of fit to the training set's own palette statistics, not to an independent historical reference.

full rationale

The main method (P3 + SASA) is not circular by itself: P3 is trained with a genuine palette-regression loss on poem-painting pairs, and SASA is a deterministic geometric gate; neither reduces to the other. The paired evaluation against expert-colorized artifacts (PSNR 17.44, SSIM 0.681, LPIPS 0.334) is an independent check, as is the blind expert study (65.8% preference, chi-square 60.50), and the paper compares against 11 external baselines. However, two load-bearing supports are not fully independent. First, the synthetic degradation that provides the bulk of training supervision is 'inspired by PRevivor (Tang et al., 2025)' - a same-group paper - and its effectiveness is asserted by citing that same paper ('validated in PRevivor as well as by our experimental results'), with the expert subset used only for fine-tuning and its size unreported. Second, the 'classical color manifold' into which P3 projects poems is defined by K-Means palettes of the paper's own training paintings (§3.5), so the 'historically constrained' subspace is a self-defined target rather than an external ground truth. The unpaired FID/ΔColorfulness(WD) metrics also lack a stated reference distribution; if the reference is the training palette distribution, the headline color-error reduction partly measures fit to the P3 pretraining target. The Limitations section concedes failure in 'double-blind' and 'over-cleaning' scenarios but does not repair the reference-set gap. These issues raise the circularity score to 4 but not higher, because the central derivation and the expert-ground-truth experiments retain independent content.

Assumptions & free parameters 5 free parameters · 6 assumptions · 0 invented entities

The paper's additions are architectural: a frozen poem-to-palette dictionary (P3), boundary-normal projection and gate (SASA), and a training dataset. The 'classical color' prior is defined by K-Means on the authors' own collection, and the degradation model comes from the authors' prior PRevivor; no independent historical or pigment evidence anchors the color manifold. All other quantities are standard (CLIP, VAE, LDM, ControlNet).

free parameters (5)
  • Number of classical color bases Kb = 16 = 16
    Hand-chosen size of the learnable color dictionary D (§3.3); caps the expressive range of the 'classical color manifold' that P3 projects poems into. No search or justification reported.
  • Palette size for proxy task Kp = 8 = 8
    Hand-chosen number of dominant HSV colors extracted by K-Means for poem-to-palette pretraining (§3.5, Eq. 16); defines the color supervision signal.
  • Palette loss weights wh, lambda_ortho, lambda_usage = not reported
    Weights in Lpalette and LP3 (Eqs. 17-18) are introduced without values; their balance determines how much the P3 prior trusts hue vs saturation/value and how hard the dictionary is pushed to stay orthogonal and balanced.
  • SASA multi-scale dilation set {1,2,4,8} = 1,2,4,8
    Hand-chosen dilation scales for the structure-guided gate synthesis (§3.4, Eq. 10); no ablation on the scale set is reported.
  • SASA injection layers = unspecified
    'Applied to the input blocks and middle block of ControlNet' (§3.4) — the exact layer selection is a design choice that controls where poem-conditioned colors are gated; no layer ablation is reported.
assumptions (6)
  • domain assumption Synthetic fading (stochastic degradation 'inspired by PRevivor') faithfully models real pigment aging
    §4.1 — most training pairs are synthetic; if the degradation model mismatches real fading (yellowing, partial pigment loss), the unpaired real-world gains are artifacts of the synthetic distribution.
  • domain assumption K-Means dominant colors of the authors' painting collection define a historically valid 'classical color system'
    §3.5, Eq. 16 — the poem-to-palette proxy task treats extracted HSV palettes as ground-truth classical colors; no art-historical source is cited for this equivalence.
  • domain assumption Expert hand-colorized artifacts are valid historical ground truth
    §4.1 — restorers collaborated with the authors; inter-expert agreement is not reported; the test metrics and the 'historical authenticity' claim rest on this subjective reference.
  • domain assumption CLIP text embeddings of VLM-grounded poem descriptions preserve the poem's color semantics
    §3.3 — QwenVL2.5 distills the poem into a grounded description P, and a frozen CLIP text encoder encodes it; errors in distillation propagate directly into the color prior.
  • ad hoc to paper Removing the feature component along a learned lift of the 2D Sobel boundary normal suppresses cross-boundary color leakage
    §3.4, Eqs. 13-14 — the 2D normal is lifted to feature space by a learned linear projection; the geometric interpretation ('suppresses components that propagate across boundary-normal directions') is a design assumption, not an established result.
  • standard math Stable Diffusion v2.1 plus ControlNet provide a suitable frozen generative backbone
    §3.1 — the framework inherits LDM and ControlNet machinery as standard background; the claim depends on this backbone being a valid prior for ancient-painting colorization.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reviving Ancient Paintings via Poem: A Colorization Framework for Aligning Cultural Semantics." pith.science (2026). https://pith.science/paper/R5U7ZBRG

@misc{pith2026260717638,
  author       = {Pith},
  title        = {Pith review of: Reviving Ancient Paintings via Poem: A Colorization Framework for Aligning Cultural Semantics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R5U7ZBRG}},
  note         = {Machine review of arXiv:2607.17638}
}
read the original abstract

The irreversible fading of ancient paintings disrupts the "congruence between poems and paintings", a core aesthetic principle where visual imagery harmonizes with literary inscriptions. Although diffusion models provide strong generative priors, restoring historically faithful colors remains difficult: visual restoration is inherently ambiguous, while direct text guidance often causes modern semantic bias, over-saturation, and cross-boundary color leakage. To address this, we propose PoemColor, a poem-guided ancient painting colorization framework. Our method aligns poetic cultural semantics with painting restoration through two key designs. First, the Poetic Painting Projector (P3) converts implicit poetic context into a classical color-aware condition via poem-to-palette pretraining, reducing the ambiguity of poem-to-color mapping. Second, Structure-Aware Semantic Attention (SASA) regulates how poetic color semantics are injected into the diffusion backbone by jointly controlling their propagation direction and regional injection strength. In addition, we construct a hybrid restoration dataset that integrates synthetic degradation with expert-restored artifacts, providing both scalable supervision and real classical color references. Extensive experiments demonstrate that our framework significantly outperforms state-of-the-art methods, delivering controllable colorization that revives both historical authenticity and poetic semantics.

Figures

Figures reproduced from arXiv: 2607.17638 by the authors.

Figure 1
Figure 1. Overall architecture of our poem-guided painting colorization framework. (a) Given a faded painting Ifaded and poem L, the framework extracts the structure latent znative, low-frequency color anchor canchor, and poetic condition epoetic, then injects them into the ControlNet restoration backbone. (b) P3 projects CLIP text tokens onto a learnable classical color dictionary and fuses the projected color representation… view at source ↗
Figure 2
Figure 2. Qualitative comparison with state-of-the-art methods. We compare PoemColor against 11 baselines categorized by architecture: GAN/CNN-based , Transformer-based , and Diffusion-based methods. The left panel (a-d) displays reconstruction results on the Paired dataset with Ground Truth (GT), while the right panel (e-m) demonstrates colorization on Unpaired artifacts. tures are insufficient for poem-guided restoration. R… view at source ↗
Figure 3
Figure 3. Distribution of expert preference scores. Our approach consistently stabilizes in the 65%–70% range, significantly sur￾passing top-tier baselines like L-CAD and ControlColor. 8 [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Visual demonstration of cross-domain applicability on Western oil paintings. The PoemColor successfully modulates color while preserving the original artwork’s distinct structural strokes. Limitations and Future Work While our decoupled de￾sign ensures graceful degrada…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 1 linked inside Pith

  1. [1]

    Using hybrid knowledge engineering and image processing in color virtual restoration of ancient murals , author =

  2. [2]

    New Generation Computing , volume =

    Chen, Lieu-Hen and Tsai, Meng-Feng and Hsu, Chien-Hui and Chen, Yu-Sheng , title =. New Generation Computing , volume =

  3. [3]

    2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=

    Finecontrolnet: Fine-level text control for image generation with spatially aligned text control injection , author=. 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) , pages=. 2025 , organization=

    2025

  4. [4]

    Haslam, Devin and Ibrahim, Soad and Elmesalami, Ayman , title =. Recent

  5. [5]

    ACM Transactions on Graphics (TOG) , volume=

    Unicolor: A unified framework for multi-modal colorization with transformer , author=. ACM Transactions on Graphics (TOG) , volume=. 2022 , publisher=

    2022

  6. [6]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Image-to-image translation with conditional adversarial networks , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  7. [7]

    Proceedings of the 29th ACM International Conference on Multimedia , pages=

    Focusing on persons: Colorizing old images learning from modern historical movies , author=. Proceedings of the 29th ACM International Conference on Multimedia , pages=

  8. [8]

    European Conference on Computer Vision , pages=

    ColorFormer: Image colorization via color memory assisted hybrid-attention transformer , author=. European Conference on Computer Vision , pages=. 2022 , organization=

    2022

Show all 59 references
  1. [9]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Ddcolor: Towards photo-realistic image colorization via dual decoders , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  2. [10]

    European Conference on Computer Vision , pages=

    Bigcolor: Colorization using a generative color prior for natural images , author=. European Conference on Computer Vision , pages=. 2022 , month=

    2022

  3. [11]

    European conference on computer vision , pages=

    Diffbir: Toward blind image restoration with generative diffusion prior , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  4. [12]

    Liang et al

    Control Color: Multimodal Diffusion-Based Interactive Image Colorization: Z. Liang et al. , author=. International Journal of Computer Vision , volume=. 2025 , publisher=

    2025

  5. [13]

    2024 , eprint=

    DINOv2: Learning Robust Visual Features without Supervision , author=. 2024 , eprint=

    2024

  6. [14]

    2025 , eprint=

    ControlNeXt: Powerful and Efficient Control for Image and Video Generation , author=. 2025 , eprint=

    2025

  7. [15]

    2021 , url =

    Color degradation mapping of rock art paintings using microfading spectrometry , journal =. 2021 , url =

    2021

  8. [16]

    2022 , url =

    Salmona, Antoine and Bouza, Lucía and Delon, Julie , journal =. 2022 , url =

    2022

  9. [17]

    2022 , eprint=

    Denoising Diffusion Implicit Models , author=. 2022 , eprint=

    2022

  10. [18]

    ACM SIGGRAPH 2022 conference proceedings , pages=

    Palette: Image-to-image diffusion models , author=. ACM SIGGRAPH 2022 conference proceedings , pages=

    2022

  11. [19]

    2025 , eprint=

    PRevivor: Reviving Ancient Chinese Paintings using Prior-Guided Color Transformers , author=. 2025 , eprint=

    2025

  12. [20]

    Machine learning and knowledge extraction , volume=

    A comprehensive review of yolo architectures in computer vision: From yolov1 to yolov8 and yolo-nas , author=. Machine learning and knowledge extraction , volume=. 2023 , publisher=

    2023

  13. [21]

    proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Bringing old photos back to life , author=. proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  14. [22]

    L-CAD: Language-based Colorization with Any-level Descriptions using Diffusion Priors , volume =

    chang, zheng and Weng, Shuchen and Zhang, Peixuan and Li, Yu and Li, Si and Shi, Boxin , booktitle =. L-CAD: Language-based Colorization with Any-level Descriptions using Diffusion Priors , volume =

  15. [23]

    European conference on computer vision , pages=

    CT 2: Colorization transformer via color tokens , author=. European conference on computer vision , pages=. 2022 , month=

    2022

  16. [24]

    Proceedings of the IEEE international conference on computer vision , pages=

    Unpaired image-to-image translation using cycle-consistent adversarial networks , author=. Proceedings of the IEEE international conference on computer vision , pages=

  17. [25]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Adding conditional control to text-to-image diffusion models , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  18. [26]

    SIGGRAPH Asia 2023 Conference Papers , pages=

    Diffusing colors: Image colorization with text guided diffusion , author=. SIGGRAPH Asia 2023 Conference Papers , pages=

    2023

  19. [27]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Attngan: Fine-grained text to image generation with attentional generative adversarial networks , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  20. [28]

    Proceedings of the IEEE international conference on computer vision , pages=

    Stackgan: Text to photo-realistic image synthesis with stacked generative adversarial networks , author=. Proceedings of the IEEE international conference on computer vision , pages=

  21. [29]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    High-resolution image synthesis with latent diffusion models , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  22. [30]

    arXiv preprint arXiv:2204.06125 , volume=

    Hierarchical text-conditional image generation with clip latents , author=. arXiv preprint arXiv:2204.06125 , volume=

    arXiv

  23. [31]

    2021 , eprint=

    Paint4Poem: A Dataset for Artistic Visualization of Classical Chinese Poems , author=. 2021 , eprint=

    2021

  24. [32]

    npj Heritage Science , volume=

    Poe2CLP: phrase-level attention and cross-modal semantic alignment for poem generate chinese landscape paintings , author=. npj Heritage Science , volume=. 2025 , publisher=

    2025

  25. [33]

    Proceedings of the AAAI conference on artificial intelligence , volume=

    T2i-adapter: Learning adapters to dig out more controllable ability for text-to-image diffusion models , author=. Proceedings of the AAAI conference on artificial intelligence , volume=

  26. [34]

    Proceedings of the 31st International Conference on Computational Linguistics , pages=

    Poetry in pixels: Prompt tuning for poem image generation via diffusion models , author=. Proceedings of the 31st International Conference on Computational Linguistics , pages=

  27. [35]

    2023 , eprint=

    IP-Adapter: Text Compatible Image Prompt Adapter for Text-to-Image Diffusion Models , author=. 2023 , eprint=

    2023

  28. [36]

    , booktitle =

    Zhao, Shihao and Chen, Dongdong and Chen, Yen-Chun and Bao, Jianmin and Hao, Shaozhe and Yuan, Lu and Wong, Kwan-Yee K. , booktitle =. Uni-ControlNet: All-in-One Control to Text-to-Image Diffusion Models , volume =

  29. [37]

    2023 , eprint=

    Composer: Creative and Controllable Image Synthesis with Composable Conditions , author=. 2023 , eprint=

    2023

  30. [38]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Ominicontrol: Minimal and universal control for diffusion transformer , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

  31. [39]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

    Attention calibration for disentangled text-to-image personalization , author=. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , pages=

  32. [40]

    Heritage Science , volume=

    FHS-adapter: fine-grained hierarchical semantic adapter for Chinese landscape paintings generation , author=. Heritage Science , volume=. 2024 , publisher=

    2024

  33. [41]

    Learning to Generate Poetic Chinese Landscape Painting with Calligraphy , booktitle =

    Yuan, Shaozu and Dai, Aijun and Yan, Zhiling and Liu, Ruixue and Chen, Meng and Chen, Baoyang and Qiu, Zhijie and He, Xiaodong , editor =. Learning to Generate Poetic Chinese Landscape Painting with Calligraphy , booktitle =

  34. [42]

    2025 , eprint=

    Qwen2.5-VL Technical Report , author=. 2025 , eprint=

    2025

  35. [43]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Coco-stuff: Thing and stuff classes in context , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  36. [44]

    2009 IEEE conference on computer vision and pattern recognition , pages=

    Imagenet: A large-scale hierarchical image database , author=. 2009 IEEE conference on computer vision and pattern recognition , pages=. 2009 , organization=

    2009

  37. [45]

    European conference on computer vision , pages=

    Pixel-aware stable diffusion for realistic image super-resolution and personalized stylization , author=. European conference on computer vision , pages=. 2024 , organization=

    2024

  38. [46]

    Best Artworks of All Time , year =

  39. [47]

    IEEE transactions on image processing , volume=

    Image quality assessment: from error visibility to structural similarity , author=. IEEE transactions on image processing , volume=. 2004 , publisher=

    2004

  40. [48]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    The unreasonable effectiveness of deep features as a perceptual metric , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  41. [49]

    Human vision and electronic imaging VIII , volume=

    Measuring colorfulness in natural images , author=. Human vision and electronic imaging VIII , volume=. 2003 , month=

    2003

  42. [50]

    GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium , volume =

    Heusel, Martin and Ramsauer, Hubert and Unterthiner, Thomas and Nessler, Bernhard and Hochreiter, Sepp , booktitle =. GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium , volume =

  43. [51]

    Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains , volume =

    Tancik, Matthew and Srinivasan, Pratul and Mildenhall, Ben and Fridovich-Keil, Sara and Raghavan, Nithin and Singhal, Utkarsh and Ramamoorthi, Ravi and Barron, Jonathan and Ng, Ren , booktitle =. Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional D...

  44. [52]

    Proceedings of ACM Symposium on User Interface Software and Technology , pages=

    PColorizor: Re-coloring Ancient Chinese Paintings with Ideorealm-congruent Poems , author=. Proceedings of ACM Symposium on User Interface Software and Technology , pages=

  45. [53]

    International conference on machine learning , pages=

    Learning transferable visual models from natural language supervision , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  46. [54]

    Localization, and Dense Features , volume=

    SigLIP 2: Multilingual Vision-Language Encoders with Improved Semantic Understanding , author=. Localization, and Dense Features , volume=

  47. [55]

    arXiv preprint arXiv:2511.14099 , year=

    FAPE-IR: Frequency-Aware Planning and Execution Framework for All-in-One Image Restoration , author=. arXiv preprint arXiv:2511.14099 , year=

  48. [56]

    Humanities and Social Sciences Communications , year=

    From traditional craft to digital restoration: an intelligent rebirth of ancient Chinese painting restoration technique , author=. Humanities and Social Sciences Communications , year=

  49. [57]

    Heritage Science , volume=

    Sgrgan: sketch-guided restoration for traditional Chinese landscape paintings , author=. Heritage Science , volume=. 2024 , publisher=

    2024

  50. [58]

    npj Heritage Science , volume=

    Digital restoration of ancient Jiangnan murals via proxy learning and structural guidance , author=. npj Heritage Science , volume=. 2026 , publisher=

    2026

  51. [59]

    arXiv preprint arXiv:2602.22983 , year=

    Obscure but effective: Classical chinese jailbreak prompt optimization via bio-inspired search , author=. arXiv preprint arXiv:2602.22983 , year=

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.