Pith. sign in

REVIEW 4 major objections 9 minor 36 references

TexSketch turns plain line drawings into textured colored sketches by procedure, not by copying anime datasets, so colorization models can train without manual labels or fixed style bias.

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 19:44 UTC pith:AZO3XY37

load-bearing objection Solid NPR systems paper with a clean pipeline and a real user study; the synthetic-supervision claim is sold harder than the evidence supports. the 4 major comments →

arxiv 2607.24261 v1 pith:AZO3XY37 submitted 2026-07-27 cs.GR

TexSketch: Bringing Texture-Aware Colorization to Sketches

Taraash Mittal , Gaurav Rai , Ojaswa Sharma This is my paper
classification cs.GR
keywords sketch colorizationprocedural stylizationsynthetic dataset generationgeometric descriptorsnon-photorealistic renderingOpen Shading Languagevision-language color assignmenttexture-aware colorization
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.

Sketch colorization models need large paired sketch–color datasets, but existing ones are small, costly to annotate, and locked into anime cel-shading, so trained models fail on watercolor, pencil, and other styles. TexSketch replaces hand-colored corpora with a fully automatic pipeline: it segments a sketch into regions, assigns semantic colors with a vision-language model, builds dense geometric fields from each region’s boundary, and drives programmable shaders that paint the fill. Style is defined by shader parameters rather than inherited from a static art collection, so one sketch can yield many different textured pairs at scale and without artists. In a human study, raters preferred these outputs to two recent learning-based colorizers on coherence, fidelity, and artifact level. The claim is that this controllable synthetic stream can supply the missing supervision for texture-aware, cross-domain sketch colorization.

Core claim

A fully automatic pipeline—segment-anything region extraction, context-aware HSL color assignment by a vision-language model, boundary-propagated geometric descriptor fields, and Open Shading Language procedural stylization—produces colored sketches that human raters judge more coherent, faithful, and clean than leading reference-based colorizers, while giving explicit control over non-photorealistic textures that anime-centric datasets cannot supply.

What carries the argument

Boundary-derived geometric descriptor fields (distance-to-edge, curvature, shadow proxy, direction field) propagated inward by jump flooding and fed into programmable OSL shaders that modulate a VLM-predicted base color—decoupling semantic color from texture so artistic style can be programmed rather than learned from a fixed corpus.

Load-bearing premise

That human preference for the rendered images is enough to prove they will work as effective synthetic training data; the paper never trains or evaluates any downstream colorization model on its pairs.

What would settle it

Train a standard sketch colorization model on TexSketch pairs spanning watercolor, cel, and crayon styles, then test whether cross-domain accuracy on non-anime line art rises relative to the same model trained only on existing anime-centric paired datasets.

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

If this is right

  • Paired sketch–color datasets can be generated at scale without manual annotation or artist supervision.
  • One line drawing can yield many high-quality training pairs in distinct artistic mediums by changing shader parameters.
  • Colorization research can escape anime-centric style bias by training on procedurally textured supervision.
  • The same controllable stream can serve as a benchmark source for texture-aware sketch colorization.

Where Pith is reading between the lines

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

  • Real downstream gains will depend on whether SAM segmentation errors and VLM color mistakes systematically poison the learned mapping more than the texture diversity helps.
  • The geometric-descriptor-plus-shader stack could also supply paired data for related NPR tasks such as sketch-to-painting transfer or line-art cleanup.
  • Learning shader parameters from a few artist exemplars would address the paper’s own limit that hand-written shaders cannot capture artist-specific intent.

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 / 9 minor

Summary. The paper presents TexSketch, a fully automatic procedural pipeline that turns an input line sketch into a colored, stylized rendering: SAM segments the sketch into regions, a Qwen-family VLM assigns each region an HSL base color conditioned on previously colored regions, boundary-derived geometric descriptors (distance-to-edge, curvature, a |n·L| shadow proxy, and a direction field) are propagated into region interiors via Jump Flooding with exponential falloff, and OSL shaders composite strokes, stippling, Gabor noise, and quantized shading into the final image. Shader parameter variation yields distinct styles (cel, watercolor, crayon) from one sketch. The motivation is dataset generation: existing paired sketch-color corpora are anime-centric, and the paper argues procedural generation provides scalable, style-diverse synthetic supervision for sketch colorization. Evaluation is a 40-participant, 15-sample MOS study over five criteria in which TexSketch outscores ColorizeDiffusionXL and SketchDeco, plus qualitative style-control results. No downstream colorization model is trained or evaluated on TexSketch data.

Significance. If the supervision claim were substantiated, the work would be significant: paired sketch-color data for non-anime styles is a genuine bottleneck, and a controllable procedural generator is a plausible route around it. The paper's concrete strengths are a fully automatic, well-specified pipeline (SAM segmentation, VLM-guided HSL assignment with a context history, JFA-propagated geometric descriptors, OSL stylization with explicit style knobs for cel and watercolor), constructive closed-form descriptor definitions, and a 40-participant human study showing preference over two named baselines. The style-controllability demonstration (Fig. 5) is a real, if qualitative, contribution. However, the paper's significance currently rests on an untested proxy: human preference for the renderings is offered as evidence of downstream training utility, and that link is neither established here nor safe to assume.

major comments (4)
  1. [Abstract / §VII (Conclusion)] Abstract, §I contributions, and §VII: the paper's stated purpose is to supply 'synthetic supervision for sketch colorization' that mitigates anime-centric dataset bias, yet no experiment trains or evaluates any colorization model on TexSketch-generated pairs. The only quantitative evidence is a perceptual MOS study (Table I), which supports plausibility and preference, not usefulness as training signal. These are distinct claims: procedural output can be perceptually plausible while encoding systematic regularities (the exponential falloff of Eq. 1, the |n·L| shadow proxy of Eq. 7, the quantized lightness of Eq. 14) that a downstream model absorbs as shortcuts. Either add a minimal downstream experiment (e.g., fine-tune or train an existing colorizer on TexSketch pairs and report performance on a non-anime target domain such as watercolor or colored pencil), or scope the abstract and con
  2. [§IV, Table I] Table I reports MOS from 40 participants x 15 samples with no significance tests and no per-criterion confidence intervals in the table itself (Fig. 4's caption mentions 95% CIs but the manuscript text does not state whether differences are significant). Two of the five margins over ColorizeDiffusionXL are small — color consistency 3.66 vs 3.59 and aesthetics 3.55 vs 3.45 — and with 15 stimuli the effective sample for stimulus-level effects is limited. The claim 'highest mean opinion score on all five evaluation criteria' (§IV) should be backed by an appropriate test (e.g., paired tests over participants with stimulus-level aggregation, or mixed-effects models) and CIs reported in Table I. The sketch-fidelity (3.70 vs 3.37) and semantic-correctness (3.47 vs 3.17) gaps look robust; the others may not be.
  3. [§IV (baselines and protocol)] The evaluation compares TexSketch (a dataset-generation pipeline) against two reference-based colorization methods (ColorizeDiffusionXL [25], SketchDeco [23]) on perceptual quality of individual outputs. The manuscript does not state what reference images or palettes the baselines were given, how the 15 test sketches were selected from ImageNetSketch and TU-Berlin, or whether baselines were run by the authors or taken from official implementations. This matters because (a) reference-based methods are sensitive to reference choice, so an unspecified protocol makes the comparison hard to interpret or reproduce, and (b) the comparison does not isolate the paper's actual contribution, which is data generation. A more on-target comparison would contrast TexSketch data against an alternative scalable source (e.g., recolored anime data, or simple flat-fill procedural baselines) in the downstrea
  4. [§IV / §V (scalability and diversity evidence)] The scalability claim ('scalable dataset generation', Abstract; 'efficient generation of stylistically diverse paired sketch–color data', §VII) is asserted but no generation statistics are given: how many images were synthesized, wall-clock time per sketch, failure rate of SAM segmentation on freehand TU-Berlin sketches (§VI admits segmentation is a hard dependency), or any diversity metric over the generated corpus. Since the motivating bottleneck is dataset scale and diversity, at least aggregate throughput figures, a segmentation failure-rate estimate on TU-Berlin, and a quantitative diversity measure (e.g., color histogram coverage or embedding-space spread across style knobs in §V) would substantiate the central motivation.
minor comments (9)
  1. [§VI] Typo in §VI: 'Lastly, Finally, although the procedural stylization pipeline...' — remove one of the two discourse markers.
  2. [§III-B / Appendix] Reference [6] cites the 2023 Qwen-VL paper, but the actual model used (version, checkpoint, temperature, and how deterministic the HSL assignments are across runs) is not stated. Given that semantic color assignment (§III-B) is a core pipeline stage, reproducibility requires the exact model and decoding settings; the Appendix prompts are a good start.
  3. [§IV / Table I] The 'Artifact level' criterion in §IV and Table I is ambiguous: higher scores are said to indicate better quality, but the criterion name suggests more artifacts. Please clarify the anchor wording shown to participants (presumably 'fewer artifacts = higher score').
  4. [§IV (user study)] User-study protocol details are thin: participant background (artists vs. general population), presentation order randomization across the three methods, and whether ratings were collected per-image in isolation or side-by-side. A sentence or two per item would suffice.
  5. [§III-C.3, Eq. 7] Eq. 7 uses the absolute Lambertian |n·L|, which discards the sign and thus cannot distinguish lit from shadow-facing boundaries; both hemispheres receive identical values. A brief justification (or note that directionality is intentionally ignored) would help, since this is the only shading cue.
  6. [§III-D.6, Eq. 17] Eq. 17 composes lightness multiplicatively over three remapped fields (G, M_shadow, M_curvature); the product of several sub-unity fields can drive lightness toward zero. Stating the remap ranges used in practice would clarify the dynamic range of the output.
  7. [§III-C–D] Sensitivity to the free hyperparameters (σ in Eqs. 1/8, γ in Eqs. 3–4, τ_p in Eq. 12, tonal levels L in Eq. 14) is not analyzed; a small ablation or at least the fixed values used would aid reproducibility.
  8. [§II] The related-work section would benefit from situating the work against synthetic-data/domain-randomization literature for training (e.g., procedural dataset generation in other vision tasks), since the paper's motivation is precisely the transferability of synthetic supervision.
  9. [References] Reference [15] lists 'Anonymous, D. community, and G. Branwen' — please format consistently with venue guidelines; also several venue strings are inconsistent ('Ieee Access', capitalization of conference names).

Circularity Check

0 steps flagged

No significant circularity: TexSketch is a constructive procedural pipeline evaluated by external human ratings, not a fitted-or-self-defined prediction loop.

full rationale

The paper’s load-bearing chain is engineering construction, not a closed derivation. Region masks come from SAM; base HSL colors from a VLM conditioned on sketch context and prior-region history; geometric fields (distance-to-edge, curvature, shadow proxy, direction) are computed from boundary geometry via JFA/PCA and explicit formulas (Eqs. 1–9); appearance is produced by hand-authored OSL shaders (Gabor noise, quantized lightness, stroke primitives) whose parameters are user-chosen style knobs, not fits to the reported MOS. Evaluation is a separate human preference study (n=40, 15 samples) against external baselines, not a quantity recovered from the same inputs used to define the method. There is no self-definitional identity (X defined as Y then “predicted” as Y), no parameter fitted to a data subset and re-reported as prediction, no load-bearing uniqueness theorem imported from overlapping authors, and no renaming of a known empirical law as a first-principles result. The untested claim that the outputs are effective synthetic supervision for downstream colorizers is an evidence gap, not circularity under the stated criteria. Honest finding: score 0; steps empty.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 2 invented entities

The central claim rests on off-the-shelf foundation models behaving well on sketches, on hand-chosen geometric proxies standing in for artist shading, and on a library of manually authored shader parameters—not on a closed-form theorem. Free parameters are the numerous falloff, density, and quantization knobs; axioms are domain assumptions about SAM/VLM/NPR; the invented entity is the integrated TexSketch pipeline itself.

free parameters (4)
  • shadow distance falloff σ = 60.0
    Exponential attenuation of boundary shadow into the interior; set to 60.0 in §III-C.3 without a fit procedure or sensitivity study.
  • distance-field contrast γ = 0.5
    Power-law on normalized distance masks (§III-C.1); fixed at 0.5 to emphasize near-boundary gradients.
  • primitive spawn density τ_p = 0.8
    Hash threshold controlling stochastic stroke/stipple density in the OSL stage (§III-D.3).
  • tonal quantization levels L and shader style knobs
    Number of lightness bands, warp magnitudes, Gabor weights/frequencies, stroke length ℓ, noise scales, and per-style overrides (cel vs watercolor) are author-chosen to achieve target looks; not derived from data.
axioms (4)
  • domain assumption SAM automatic mask generation, after area/border filtering and smoothing, yields semantically coherent closed regions on freehand and ImageNet-style sketches.
    Invoked as the first pipeline stage (§III-A); failures are admitted to propagate in §VI.
  • domain assumption A vision-language model conditioned on the red-highlighted region, global sketch, and color history produces perceptually and semantically plausible HSL base colors in structured JSON.
    §III-B and Appendix prompts; no quantitative color-accuracy metric is reported.
  • domain assumption Boundary-derived fields (distance, curvature via ring-intersection+PCA, Lambertian shadow proxy, tangent direction), JFA-propagated with exponential falloff, are sufficient geometric cues to drive convincing hand-drawn shading and stroke alignment.
    Entire §III-C motivates descriptors as what 'an artist implicitly follows'; used as hard inputs to all shaders.
  • ad hoc to paper Human mean-opinion scores on 15 sketches are an adequate proxy for usefulness as training supervision for sketch colorization.
    Results §IV equate higher MOS with 'effective for synthetic supervision' without a train-on-synthetic / test-on-real experiment.
invented entities (2)
  • TexSketch end-to-end procedural colored-sketch generator no independent evidence
    purpose: Unify segmentation, VLM color planning, geometric descriptor fields, and OSL stylization into a controllable dataset factory.
    Named contribution in Abstract and §I; existence is demonstrated by the pipeline and figures, not by an external physical prediction.
  • Interior/global dual distance mask and curvature/shadow/direction descriptor suite as shader guidance no independent evidence
    purpose: Supply dense per-pixel cues for NPR primitives inside each SAM region.
    Specific combination and propagation recipe in §III-C is paper-defined; individual ideas (distance fields, Gabor, JFA) are prior.

pith-pipeline@v1.2.0-grok45-kimik3 · 15838 in / 3453 out tokens · 80184 ms · 2026-07-31T19:44:07.670111+00:00 · methodology

0 comments
read the original abstract

Reference-based sketch colorization methods rely on large paired datasets that preserve both the structural and stylistic characteristics of hand-drawn artwork. However, existing datasets are limited in scale, expensive to annotate, and bound to fixed, often inconsistent artistic style biases that propagate to downstream models and limit cross-domain generalization. We present TexSketch, a controllable procedural framework for generating colored-sketch datasets with programmable artistic styles via geometric analysis and shader-driven stylization. Our fully automatic pipeline integrates region extraction, semantic color prediction, and shader-based rendering. By defining artistic appearance procedurally rather than inheriting it from a static corpus, TexSketch enables scalable dataset generation without manual annotation or artist supervision. Human studies demonstrate that TexSketch generates perceptually plausible colored sketches with high stylistic diversity, providing a controllable, scalable source of synthetic supervision for sketch colorization.

Figures

Figures reproduced from arXiv: 2607.24261 by Gaurav Rai, Ojaswa Sharma, Taraash Mittal.

Figure 1
Figure 1. Figure 1: Overview of the proposed TexSketch pipeline. The input sketch is first partitioned into semantically coherent regions, [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Visual results of the proposed TexSketch framework on sketches from ImageNet [29] and TU-Berlin [7] datasets. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: Mean opinion score per evaluation criterion for each [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of our method with ColorizeD [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Examples demonstrating the style controllability of [PITH_FULL_IMAGE:figures/full_fig_p008_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

36 extracted references · 3 linked inside Pith

  1. [1]

    Animecolor: Reference-based animation colorization with diffusion transformers,

    Y . Zhang, L. Wang, H. Wang, D. Wu, Z. Lin, F. Wang, and L. Song, “Animecolor: Reference-based animation colorization with diffusion transformers,” inProceedings of the 33rd ACM International Conference on Multimedia, pp. 6682–6690, 2025

    2025

  2. [2]

    Colorizediffusion: Improving reference-based sketch colorization with latent diffusion model,

    D. Yan, L. Yuan, E. Wu, Y . Nishioka, I. Fujishiro, and S. Saito, “Colorizediffusion: Improving reference-based sketch colorization with latent diffusion model,” in2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pp. 5092–5102, IEEE, 2025

    2025

  3. [3]

    Mangadit: Reference- guided line art colorization with hierarchical attention in diffusion transformers,

    Q. Qiu, J. Mao, K. Masui, and X. Wang, “Mangadit: Reference- guided line art colorization with hierarchical attention in diffusion transformers,”arXiv preprint arXiv:2508.09709, 2025

    Pith/arXiv arXiv 2025

  4. [4]

    Open shading language,

    L. Gritz, C. Stein, C. Kulla, and A. Conty, “Open shading language,” in ACM SIGGRAPH 2010 Talks, pp. 1–1, 2010

    2010

  5. [5]

    Sam 2: Segment anything in images and videos,

    N. Ravi, V . Gabeur, Y .-T. Hu, R. Hu, C. Ryali, T. Ma, H. Khedr, R. Rädle, C. Rolland, L. Gustafson,et al., “Sam 2: Segment anything in images and videos,” inInternational Conference on Learning Repre- sentations, vol. 2025, pp. 28085–28128, 2025

    2025

  6. [6]

    Qwen-vl: A versatile vision-language model for un- derstanding, localization, text reading, and beyond,

    J. Bai, S. Bai, S. Yang, S. Wang, S. Tan, P. Wang, J. Lin, C. Zhou, and J. Zhou, “Qwen-vl: A versatile vision-language model for un- derstanding, localization, text reading, and beyond,”arXiv preprint arXiv:2308.12966, 2023

    Pith/arXiv arXiv 2023

  7. [7]

    How do humans sketch objects?,

    M. Eitz, J. Hays, and M. Alexa, “How do humans sketch objects?,” ACM Transactions on graphics (TOG), vol. 31, no. 4, pp. 1–10, 2012

    2012

  8. [8]

    A neural representation of sketch drawings,

    D. Ha and D. Eck, “A neural representation of sketch drawings,”arXiv preprint arXiv:1704.03477, 2017

    Pith/arXiv arXiv 2017

  9. [9]

    Sketchsegnet+: An end-to-end learning of rnn for multi-class sketch semantic segmentation,

    Y . Qi and Z.-H. Tan, “Sketchsegnet+: An end-to-end learning of rnn for multi-class sketch semantic segmentation,”Ieee Access, vol. 7, pp. 102717–102726, 2019

    2019

  10. [10]

    Object category un- derstanding via eye fixations on freehand sketches,

    R. K. Sarvadevabhatla, S. Suresh, and R. V . Babu, “Object category un- derstanding via eye fixations on freehand sketches,”IEEE Transactions on Image Processing, vol. 26, no. 5, pp. 2508–2518, 2017

    2017

  11. [11]

    The sketchy database: learning to retrieve badly drawn bunnies,

    P. Sangkloy, N. Burnell, C. Ham, and J. Hays, “The sketchy database: learning to retrieve badly drawn bunnies,”ACM Transactions on Graph- ics (TOG), vol. 35, no. 4, pp. 1–12, 2016

    2016

  12. [12]

    Sketchyscene: Richly-annotated scene sketches,

    C. Zou, Q. Yu, R. Du, H. Mo, Y .-Z. Song, T. Xiang, C. Gao, B. Chen, and H. Zhang, “Sketchyscene: Richly-annotated scene sketches,” in Proceedings of the european conference on computer vision (ECCV), pp. 421–436, 2018

    2018

  13. [13]

    Texturegan: Controlling deep image synthesis with texture patches,

    W. Xian, P. Sangkloy, V . Agrawal, A. Raj, J. Lu, C. Fang, F. Yu, and J. Hays, “Texturegan: Controlling deep image synthesis with texture patches,” inProceedings of the IEEE conference on computer vision and pattern recognition, pp. 8456–8465, 2018

    2018

  14. [14]

    Scribbler: Controlling deep image synthesis with sketch and color,

    P. Sangkloy, J. Lu, C. Fang, F. Yu, and J. Hays, “Scribbler: Controlling deep image synthesis with sketch and color,” inProceedings of the IEEE conference on computer vision and pattern recognition, pp. 5400–5409, 2017

    2017

  15. [15]

    Danbooru2021: A large- scale crowdsourced & tagged anime illustration dataset

    Anonymous, D. community, and G. Branwen, “Danbooru2021: A large- scale crowdsourced & tagged anime illustration dataset.” https://gwern. net/danbooru2021, January 2022. Accessed: 2026-07-17

    2022

  16. [16]

    Two-stage sketch colorization,

    L. Zhang, C. Li, T.-T. Wong, Y . Ji, and C. Liu, “Two-stage sketch colorization,”ACM Transactions on Graphics (TOG), vol. 37, no. 6, pp. 1–14, 2018

    2018

  17. [17]

    Delaunay painting: Per- ceptual image colouring from raster contours with gaps,

    A. D. Parakkat, P. Memari, and M.-P. Cani, “Delaunay painting: Per- ceptual image colouring from raster contours with gaps,” inComputer Graphics Forum, vol. 41, pp. 166–181, Wiley Online Library, 2022

    2022

  18. [18]

    Winding number features for vector sketch colorization,

    D. Scrivener, E. Coldren, and E. Chien, “Winding number features for vector sketch colorization,” inComputer Graphics Forum, vol. 43, p. e15141, Wiley Online Library, 2024

    2024

  19. [19]

    Kisscolor: Kinetic and intuitive stroke stretching for vector drawing colorization,

    Y . Dong, H. Xin, Z. Dou, R. Xu, Y . Liu, S. Chen, S. Xin, C. Tu, T. Ko- mura, and W. Wang, “Kisscolor: Kinetic and intuitive stroke stretching for vector drawing colorization,”ACM Transactions on Graphics (TOG), vol. 44, no. 6, pp. 1–13, 2025

    2025

  20. [20]

    Ciallo: Gpu- accelerated rendering of vector brush strokes,

    S. Ciao, Z. Guan, Q. Liu, L.-Y . Wei, and Z. Wang, “Ciallo: Gpu- accelerated rendering of vector brush strokes,” inACM SIGGRAPH 2024 Conference Papers, pp. 1–11, 2024

    2024

  21. [21]

    Reference-based sketch image colorization using augmented-self reference and dense semantic correspondence,

    J. Lee, E. Kim, Y . Lee, D. Kim, J. Chang, and J. Choo, “Reference-based sketch image colorization using augmented-self reference and dense semantic correspondence,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 5801–5810, 2020

    2020

  22. [22]

    Magiccolor: Multi-instance sketch colorization,

    Y . Zhang, Y . Ma, B. Wang, Q. Chen, and Z. Wang, “Magiccolor: Multi-instance sketch colorization,” inProceedings of the IEEE/CVF International Conference on Computer Vision, pp. 15205–15217, 2025

    2025

  23. [23]

    Sketchdeco: Training-free latent composition for precise sketch colourisation,

    C. Utintu and Y .-Z. Song, “Sketchdeco: Training-free latent composition for precise sketch colourisation,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 484–494, 2026

    2026

  24. [24]

    Image referenced sketch colorization based on animation creation 8 workflow,

    D. Yan, X. Wang, Z. Li, S. Saito, Y . Iwasawa, Y . Matsuo, and J. Guo, “Image referenced sketch colorization based on animation creation 8 workflow,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 23391–23400, 2025

    2025

  25. [25]

    Towards high-resolution and disentangled reference-based sketch colorization,

    D. Yan, X. Wang, R. Wang, Z. Li, J. Yu, Y . Iwasawa, Y . Matsuo, and J. Guo, “Towards high-resolution and disentangled reference-based sketch colorization,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 11472–11481, 2026

    2026

  26. [26]

    Jump flooding in gpu with applications to voronoi diagram and distance transform,

    G. Rong and T.-S. Tan, “Jump flooding in gpu with applications to voronoi diagram and distance transform,” inProceedings of the 2006 symposium on Interactive 3D graphics and games, pp. 109–116, 2006

    2006

  27. [27]

    Principal component analysis: A natural approach to data exploration,

    F. L. Gewers, G. R. Ferreira, H. F. D. Arruda, F. N. Silva, C. H. Comin, D. R. Amancio, and L. D. F. Costa, “Principal component analysis: A natural approach to data exploration,”ACM Comput. Surv., vol. 54, May 2021

    2021

  28. [28]

    Gabor noise by example,

    B. Galerne, A. Lagae, S. Lefebvre, and G. Drettakis, “Gabor noise by example,”ACM Transactions on Graphics (SIGGRAPH Conference Proceedings), vol. 31, August 2012

    2012

  29. [29]

    Learning robust global representations by penalizing local predictive power,

    H. Wang, S. Ge, Z. Lipton, and E. P. Xing, “Learning robust global representations by penalizing local predictive power,”Advances in neural information processing systems, vol. 32, 2019. APPENDIX The prompts given to the vision-language model [6] are as follows. Listing 1: System prompt You are an expert anime, illustration, and abstract art colorist. You...

    2019

  30. [30]

    Look ONLY at the area covered by the bright red mask

  31. [31]

    hair", "skin

    Identify the object. - If concrete: Identify the exact semantic object (e.g., "hair", "skin", "jacket"). - If abstract/surreal: Identify its compositional role (e.g., "foreground geometric shape", "flowing background ribbon", "negative space")

  32. [32]

    - For abstract items: Establish a cohesive but highly varied conceptual color palette (e.g., neon cyberpunk, pastel vaporwave, vibrant primary colors)

    Determine the color strategy: - For concrete items: Use natural/logical colors (e.g., human skin tones, matching hair). - For abstract items: Establish a cohesive but highly varied conceptual color palette (e.g., neon cyberpunk, pastel vaporwave, vibrant primary colors)

  33. [33]

    identified_object

    Review the history to maintain harmony, but strictly enforce contrast. If the image is abstract, YOU MUST NOT color distinct overlapping shapes with the same hue. Force yourself to pick a contrasting color from your palette. Output format must be strictly raw JSON: { "identified_object": "<what physical object or abstract shape is inside the red mask>", "...

  34. [34]

    What is the object inside the red highlight? write it down in ' identified_object'

  35. [35]

    If it is skin or hair, use human-natural tones

    Choose an appropriate HSL color. If it is skin or hair, use human-natural tones. If it matches a previous segment's item type, make it match or harmonize

  36. [36]

    Return the JSON structure. 9