Pith. sign in

REVIEW 4 major objections 5 minor 29 references

Palette Aligned Image Diffusion

T0 review · 4 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Palette-Adapter conditions text-to-image diffusion on a color palette by encoding it as a sparse histogram, with entropy and distance dials for adherence and a negative-histogram mechanism to suppress unwanted hues.

desk verdict A useful palette-conditioning adapter with a genuine new twist, but the 'consistently outperforms' headline outruns Table 1 and the block-selection choice needs an ablation. read the letter →

arxiv 2509.02000 v1 pith:VBVA3NCA submitted 2025-09-02 cs.CV

Elad Aharoni , Noy Porat , Dani Lischinski , Ariel Shamir This is my paper
classification cs.CV
keywords palette-conditionedgenerationtext-to-imagediffusioncolorpalettessparsehistogramconditioningclassifier-freeguidancenegativeentropyanddistancecontrolscolor-balanceddatasetcuration
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 a lightweight adapter can make a frozen text-to-image diffusion model obey a user's color palette—a short list of 3–10 colors—while keeping the image sharp, coherent, and true to the prompt. The key move is to treat the palette as a sparse color histogram and feed it through a single cross-attention block in the U-Net, augmented with two scalar control parameters: histogram entropy, which governs how many distinct shades appear, and the distance between the palette and the image's full histogram, which controls how far colors may deviate from the palette. A third mechanism, a negative histogram, lets users explicitly suppress unwanted hues during classifier-free guidance. The authors support the claim with quantitative metrics, qualitative comparisons, and a user study showing they outperform prior palette-conditioning methods on the balance of adherence and quality.

What carries the argument

The load-bearing object is the histogram token set — a non-spatial encoding of a color condition as four tokens in a 34×12×10 HSV histogram plus one token for scalar features. This token set is injected at a single, isolated cross-attention block of the U-Net (up_blocks.0.attentions.1, the second attention module of the first upsampling block). Two scalar controls — histogram entropy (the Shannon entropy of the target distribution) and palette-to-histogram distance (a Quadratic-Chi Histogram Distance using CIEDE2000 color differences) — modulate adherence; a negative-histogram variant feeds a user-specified list of unwanted colors as the CFG null condition.

What would settle it

Generate images for 1,000 prompt–palette pairs sampled uniformly from a broad distribution (e.g., COCO captions and random palettes), and compare the average EMD to the target palette for the adapter conditioned at up_blocks.0.attentions.1 versus the same adapter conditioned at the mid-block or a concatenation of the top-3 blocks from the paper's screening table. If any alternative site yields a statistically lower EMD while preserving CLIP score and MusiQ quality on identical seeds, the single-critical-block claim is false. Separately, fixing the entropy and distance scalars to constants and

Watch

Extended reading notes

Core claim

The central claim: a color palette can be treated as a sparse histogram and used to condition a frozen text-to-image diffusion model through a single intermediate cross-attention block, with no spatial maps or reference images. The palette is projected into four tokens in a 34×12×10 HSV histogram, fed through decoupled cross-attention layers; a fifth token carries the histogram's entropy and its distance from the image's full histogram. These two scalars constitute a continuous dial between quantized, strictly palette-bound output and naturalistic, color-rich output. A negative-histogram condition, replacing the null in classifier-free guidance, lets users push the image away from unwanted h

Load-bearing premise

The method stakes its color control on one specific internal attention layer of the diffusion model, chosen as the most color-critical from a screen of just 20 prompts paired with 20 palettes; if that layer isn't the most effective for the wide variety of real-world prompts and palettes, color adherence will suffer.

Editorial extensions

If this is right

  • A user can go from a 3–10 color palette and a text prompt to images that perceptually match those colors, with a single slider controlling how strictly the image stays inside the palette.
  • The entropy dial alone can push the same prompt and seed from flat, posterized art to detailed, gradient-rich renders, enabling style exploration without changing the prompt.
  • The negative-histogram mechanism lets a user explicitly suppress unwanted hues (for example, yellow bananas or an off-brand color) during generation rather than relying on chance or multiple samples.
  • Because the conditioning occupies a single U-Net cross-attention block and the base model stays frozen, the adapter adds only a small parameter overhead and can be dropped into existing SDXL-based pipelines.
  • In the reported quantitative and user-study comparisons, the adapter lands on the best Pareto point among palette-conditioning methods for the palette-adherence vs. image-quality tradeoff.

Reading between the lines

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

  • The sparse-histogram framing suggests a general recipe for conditioning a frozen diffusion model on any small discrete set: encode the set as a few tokens, add interpretable scalars (entropy, distance), and train with full-condition augmentation. This could transfer to material palettes, lighting themes, or even musical key/scale constraints in audio diffusion.
  • The negative-histogram trick is a concrete instance of a broader principle: in classifier-free guidance, extrapolate from a semantic anti-condition rather than an empty one. The same recipe could suppress whole object classes or styles if the negative condition is supplied through a text or token encoder.
  • The two scalar controls are additive to the adapter, so in principle a user could change entropy or distance values at inference on an already-trained adapter without retraining, turning the adapter into a zero-cost style dial for any palette.
  • A natural stress test is to scale up the block-selection screen from 20 prompts × 20 palettes to a few thousand diverse pairs; the authors' choice of up_blocks.0.attentions.1 is a claim about that screen, and a larger screen could reveal multiple effective blocks or a different winner.
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 / 5 minor

Summary. The paper introduces Palette-Adapter, a method for conditioning SDXL on user-specified color palettes. Palettes are encoded as sparse HSV histograms (34×12×10 bins), and two scalar controls—histogram entropy and palette-to-histogram distance—are injected as an additional token to modulate adherence and color variation. A negative palette condition is introduced for CFG-style suppression of unwanted hues. Training is multi-stage: first all U-Net attention blocks are trained, then a single block (up_blocks.0.attentions.1) is selected based on average EMD over a 20-prompt × 20-palette screen, and finally that block is fine-tuned. Training uses a curated dataset of 2.4M images to balance rare colors. The method is evaluated on 1000 COCO prompts with quantitative metrics (EMD, MUSIQ, CLIP) and a 40-participant user study, reporting favorable adherence–quality trade-offs against T2I-Adapter, FLUX IC Palette, SW-Guidance, and baseline SDXL.

Significance. If the results hold, Palette-Adapter offers a practical and lightweight interface for color control in text-to-image generation, with interpretable scalar controls and a negative-palette mechanism that addresses a real failure mode of CFG. The paper includes a user study, quantitative comparison, and a clear architectural recipe. However, the strength of the central claim is currently undercut by (i) an overstatement in the abstract relative to Table 1, (ii) an unablated architectural commitment to a single attention block selected from a small screen, and (iii) a missing ablation for the curated dataset. These are not fatal to the method's promise, but they need to be addressed before the claims of consistent, generalizable state-of-the-art behavior can be accepted.

major comments (4)
  1. [Abstract and Table 1] The abstract states the method 'consistently outperforms existing approaches in achieving both strong palette adherence and high image quality.' This is contradicted by Table 1: SW-Guidance achieves the lowest EMD (0.063 vs. 0.1414), and baseline SDXL achieves the highest Koniq (72.118 vs. 70.642), AVA (5.647 vs. 5.619), and CLIP (0.316 vs. 0.311). The narrative in Section 4.2 correctly frames the result as a balance or Pareto improvement, but the abstract and intro overstate the outcome. Please revise the central claim to reflect that the method trades off against these baselines and excels on the harmonic-mean/Pareto criterion.
  2. [Supplementary §1.2] The entire color-conditioning pathway is committed to a single U-Net attention block, up_blocks.0.attentions.1, chosen from a screen of 20 prompts × 20 palettes using average EMD. This selection is made before the final fine-tuning stage, and there is no ablation after fine-tuning comparing the final single-block adapter against (a) the all-block adapter, (b) top-2/top-3 blocks, or (c) other individual blocks. The 400-sample screen can hide interactions between block efficacy and palette sparsity, saturation, or prompt color associations (e.g., the prototypical-color failure in §3.3). Because the paper claims broad generalization across prompts and palettes, this missing ablation is load-bearing: if up_blocks.0.attentions.1 is suboptimal for a substantial subset of inputs, the method will under- or mis-apply color conditioning. Please add a post-fine-tuning ablation or, failing that, res
  3. [Section 3.5 and Supplementary §1.1] The curated dataset (2.4M images with 400K rare-color images added to LAION-Art) is listed as a core contribution, and the paper argues it 'improves generalization' and 'ensures better coverage of the color space.' However, no ablation is reported comparing the final model trained on the curated set versus the same architecture trained only on LAION-Art. Without this comparison, the benefit of the curation is not demonstrated. Please provide an ablation measuring palette-adherence or color-coverage metrics on rare-color palettes, or explicitly soften the claim to a design choice rather than a validated improvement.
  4. [Section 3.2] Reproducibility is hindered by undisclosed parameters. The 'clipped and sharpened variant of CIEDE2000' is referenced but the clip bounds and sharpening exponent are not given. Similarly, the negative-palette weights in Figure 6 are cited (0.07 and 0.3) but no guidance is provided for setting them, and the entropy/distance scalar encoding (how these scalars are normalized or scaled into the token) is not specified. These values are needed to reproduce the method's behavior. Please report them or release code.
minor comments (5)
  1. [Section 4.2] The 'Reference*' row in Table 1 is computed on PicMonkey palettes, not on the COCO prompts/palettes used for all other rows. The note explains this, but the comparison is apples-to-oranges; consider reporting it only as an absolute reference point, not as a row in the comparative table.
  2. [Section 3.3, Eq. (1)] The equation for the combined noise prediction uses weight w for the positive condition and (1-w) for the negative condition. It would be clearer to define the recommended range of w (e.g., w=0.5 as default) and how it interacts with the standard CFG scale. The current description is ambiguous about whether this is a convex combination or an extrapolation.
  3. [Supplementary Table 4] The dropout probabilities table has a slight inconsistency: text embeddings show 80% for both 'Histogram' and 'Palette' columns, which presumably should be 80% for text and 5% for the 'None' case; consider labeling columns more clearly (e.g., 'Color Condition Type' vs. 'Text' rows) to avoid confusion.
  4. [Figures 4 and 5] The relative-entropy and relative-distance values used in these figures are not stated in the captions. Adding the numeric values (or a colorbar) would make the control mechanism easier to interpret.
  5. [General] There are minor typos, e.g., 'blanced distribution' in Section 2.4 and '2TI' instead of 'T2I' in Figure 7 and Table 2 captions. Please proofread.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the method is evaluated against external benchmarks, and the only self-citations are minor and non-load-bearing.

full rationale

The Palette-Adapter's central claim is an empirical capability claim. The conditioning pathway is trained on external data (LAION-Art plus curated samples), and the headline quantitative results (Table 1) and user study (Table 2) are evaluated against external benchmarks: COCO 2017 validation set and PicMonkey designer-extracted palettes. The only metric-driven architectural selection is the choice of up_blocks.0.attentions.1 as the color-critical block, made via a 20-prompt × 20-palette EMD screen in Supplementary §1.2. This is model selection on a small validation sample, not fitting of the reported test numbers; the final model is subsequently fine-tuned, and its EMD, MusiQ, CLIP, and user-study scores are measured, not optimized. The scalar entropy, distance, and negative-palette controls are user-set at inference and are not fitted parameters renamed as predictions. The two self-citations ([Cohen et al. 2024], [Frenkel et al. 2024]) are cited as inspiration for balanced/multistage conditioning, but the critical-block result is independently tabulated in Table 3, and the headline comparisons do not depend on any theorem or result imported from those papers. The paper's own limitations section acknowledges text/color trade-offs and overlapping entropy/distance scales; these are correctness risks rather than circular reductions. No equation or construction in the paper defines the reported palette-alignment metric in terms of the training objective, so no prediction reduces to its input by construction.

Assumptions & free parameters 4 free parameters · 6 assumptions · 0 invented entities

The central claim rests on hand-chosen numeric design choices (histogram binning, CIEDE2000 sharpening parameters, curation threshold) and on domain assumptions about the base model and evaluation metrics. No invented entities such as new forces or particles are introduced; the negative histogram is a conditioning mechanism, not a new entity.

free parameters (4)
  • HSV histogram binning (34, 12, 10)
    Hand-chosen to balance perceptual resolution and tractability; the central condition representation depends on this discretization, and alternatives are not validated.
  • CIEDE2000 clip and sharpening parameters
    The paper uses a 'clipped and sharpened variant' (Sec 3.2) but does not give the clip threshold or sharpening exponent; these affect the distance condition and the EMD metric used in evaluation.
  • Threshold for 'low valued histogram bins' in dataset curation
    The curation samples 400K images containing colors from low-valued bins (Sec 3.5), but the threshold for 'low' is not defined, so the dataset rebalancing is not reproducible.
  • Negative palette weights in examples = 0.07 (sofa), 0.3 (apple)
    Manual weights set per example to control attenuation of unwanted colors; these are demonstration controls, not a fitted global parameter.
assumptions (6)
  • standard math Shannon entropy (Eq 2) is the correct measure of color distribution variety.
    Used to define the entropy control in Sec 3.4; standard definition, no derivation needed.
  • domain assumption A frozen SDXL with IP-Adapter-style decoupled cross-attention can carry palette information through a single intermediate U-Net block.
    The architecture (Sec 3.2) assumes the identified block, up_blocks.0.attentions.1, is sufficient; this is an empirical assumption supported only by a 20-prompt, 20-palette screen in the supplementary.
  • domain assumption The curated dataset (2M LAION-Art + 400K rare-color LAION-2B-en images) provides sufficiently balanced color coverage for generalization.
    Sec 3.5 argues LAION-Art is color-biased and the augmentation addresses it, but no ablation isolates the effect of curation on the final metrics.
  • domain assumption EMD with a sharpened CIEDE2000 bin distance, MUSIQ, and CLIP are valid proxies for palette adherence, perceptual quality, and prompt adherence respectively.
    The quantitative evaluation (Sec 4.2) treats these as ground truth; they are standard metrics but their calibration for palette conditioning is assumed.
  • domain assumption CFG with a manually supplied negative palette (Eq 1) cleanly suppresses unwanted hues without harming semantic coherence.
    The negative-palette mechanism (Sec 3.3) is an extension of standard CFG; the paper demonstrates examples but does not prove this holds broadly across prompts.
  • domain assumption The entropy and distance scalar controls learned during training remain meaningful at inference when set by users.
    The relative-entropy ablation (supplementary Table 5) is performed on an entropy-only adapter, not the full multi-condition model, so the transfer to the full model is assumed.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Palette Aligned Image Diffusion." pith.science (2026). https://pith.science/paper/VBVA3NCA

@misc{pith2026250902000,
  author       = {Pith},
  title        = {Pith review of: Palette Aligned Image Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VBVA3NCA}},
  note         = {Machine review of arXiv:2509.02000}
}
read the original abstract

We introduce the Palette-Adapter, a novel method for conditioning text-to-image diffusion models on a user-specified color palette. While palettes are a compact and intuitive tool widely used in creative workflows, they introduce significant ambiguity and instability when used for conditioning image generation. Our approach addresses this challenge by interpreting palettes as sparse histograms and introducing two scalar control parameters: histogram entropy and palette-to-histogram distance, which allow flexible control over the degree of palette adherence and color variation. We further introduce a negative histogram mechanism that allows users to suppress specific undesired hues, improving adherence to the intended palette under the standard classifier-free guidance mechanism. To ensure broad generalization across the color space, we train on a carefully curated dataset with balanced coverage of rare and common colors. Our method enables stable, semantically coherent generation across a wide range of palettes and prompts. We evaluate our method qualitatively, quantitatively, and through a user study, and show that it consistently outperforms existing approaches in achieving both strong palette adherence and high image quality.

Figures

Figures reproduced from arXiv: 2509.02000 by the authors.

Figure 1
Figure 1. Palette Aligned Image Diffusion: Our method enables precise control over color distribution, generating images, [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Post-hoc recoloring of a grayscale image vs. palette [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Training the Palette-Adapter: During training, we [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (12 more)
Figure 6
Figure 6. Figure 6: Images generated based on a color palette condition [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 5
Figure 5. Figure 5: Distance conditioned images, generated with low [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 7
Figure 7. Figure 7: Additional comparison results are found in Section 2 of [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 10
Figure 10. Figure 10: Artwork images generated based on a color condi [PITH_FULL_IMAGE:figures/full_fig_p009_10.png]
Figure 8
Figure 8. Figure 8: The colors in random images generated by SDXL 1.0 [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: A collection of images generated using similar [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 13
Figure 13. Figure 13: Post-hoc recoloring of a grayscale image using [PITH_FULL_IMAGE:figures/full_fig_p010_13.png]
Figure 15
Figure 15. Figure 15: Palette based recoloring comparison between our [PITH_FULL_IMAGE:figures/full_fig_p010_15.png]
Figure 14
Figure 14. Figure 14: Palette alignment with different styles: Generated [PITH_FULL_IMAGE:figures/full_fig_p010_14.png]
Figure 17
Figure 17. Figure 17: User study results: Our method strikes an optimal [PITH_FULL_IMAGE:figures/full_fig_p012_17.png]
Figure 18
Figure 18. Figure 18: User study: A collection of images taken from the [PITH_FULL_IMAGE:figures/full_fig_p012_18.png]
Figure 19
Figure 19. Figure 19: Comparison with other methods: Images generated using different palettes and prompts but same seed. Images are [PITH_FULL_IMAGE:figures/full_fig_p013_19.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 18 canonical work pages

  1. [3]

    negative

    This analysis reveals that the up_blocks.0.attentions.1 block, corresponding to the second attention module in the first U-Net upsampling block, exhibits the most significant influence on color fidelity, as indicated by achieving the lowest average EMD. 1.3 Conditioning Dropout During training we mix between full histograms and palettes and described in t...

    work page

  2. [4]

    2024; Frenkel et al

    To identify the layers most critical for color control, we adopt a method inspired by [Cohen et al. 2024; Frenkel et al. 2024]. First, we train the full adapter, which consists of all 11 attention blocks (totaling 70 layers). Subsequently, each attention block is evaluated in isolation: with only one block active and all others off, we perform inference u...

    work page 2024

  3. [5]

    Conditional Balance: Improving Multi-Conditioning Trade-Offs in Image Generation

    Conditional Balance: Improving Multi-Conditioning Trade-Offs in Image Generation. arXiv:2412.19853 [cs.CV] https://arxiv.org/abs/2412.19853 Coolors.co

    work page Pith review arXiv

  4. [6]

    [Ac- cessed 24-10-2024]

    The Super Fast Color Palettes Generator! https://coolors.co/. [Ac- cessed 24-10-2024]. Yarden Frenkel, Yael Vinker, Ariel Shamir, and Daniel Cohen-Or

    work page 2024

  5. [7]

    Implicit Style-Content Separation using B-LoRA

    Implicit Style- Content Separation using B-LoRA. arXiv:2403.14572 [cs.CV] https://arxiv.org/abs/ 2403.14572 GeorgeQi

    work page Pith review arXiv

  6. [11]

    arXiv abs/2302.08453 (2023)

    T2I-Adapter: Learning Adapters to Dig out More Controllable Ability for Text-to-Image Diffusion Models. arXiv abs/2302.08453 (2023). Ofir Pele and Michael Werman

    arXiv 2023

  7. [14]

    arXiv abs/2305.06402 (2023)

    Analyzing Bias in Diffusion-based Face Generation Models. arXiv abs/2305.06402 (2023). PicMonkey

    arXiv 2023

  8. [15]

    https://www.picmonkey.com/blog/color-combinations- graphic-design

    100 Color Combinations for Designs | Examples & Inspiration | Pic- Monkey — picmonkey.com. https://www.picmonkey.com/blog/color-combinations- graphic-design. [Accessed 12-11-2024]. François Pitié, Anil C. Kokaram, and Rozenn Dahyot

    work page 2024

Show all 29 references
  1. [17]

    arXiv abs/2307.01952 (2023)

    SDXL: Improving Latent Diffusion Models for High-Resolution Image Synthesis. arXiv abs/2307.01952 (2023). Pylette

    2023 arXiv

  2. [18]

    https://qtiptip.github.io/Pylette/

    A Python color extraction library. https://qtiptip.github.io/Pylette/. [Accessed 09-11-2024]. Aditya Ramesh, Prafulla Dhariwal, Alex Nichol, Casey Chu, and Mark Chen

    2024

  3. [19]

    arXiv abs/2204.06125 (2022)

    Hierarchical Text-Conditional Image Generation with CLIP Latents. arXiv abs/2204.06125 (2022). Erik Reinhard, Michael Ashikhmin, Bruce Gooch, and Peter Shirley

    2022 arXiv

  4. [21]

    arXiv abs/2112.10752 (2022)

    High-Resolution Image Synthesis with Latent Diffusion Models. arXiv abs/2112.10752 (2022). Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily L Denton, Kamyar Ghasemipour, Raphael Gontijo Lopes, Burcu Karagol Ayan, Tim Salimans, et al

    2022 arXiv

  5. [22]

    Advances in Neural Information Processing Systems 35 (2022), 36479– 36494

    Photorealistic text-to-image diffusion models with deep language understanding. Advances in Neural Information Processing Systems 35 (2022), 36479– 36494. Claude E. Shannon

    2022

  6. [25]

    http://en

    Color scheme — Wikipedia, The Free Encyclopedia. http://en. wikipedia.org/w/index.php?title=Color%20scheme&oldid=1247518638. [Accessed 24-10-2024]. Feng Yang, Junjie Ke, Peyman Milanfar, Qifei Wang, and Yilin Wang

    2024

  7. [26]

    arXiv abs/2308.06721 (2023)

    IP-Adapter: Text Compatible Image Prompt Adapter for Text-to-Image Diffusion Models. arXiv abs/2308.06721 (2023). Lvmin Zhang, Anyi Rao, and Maneesh Agrawala

    2023 arXiv

  8. [27]

    A dress”. (b) Random images for “A waiting room with a sofa and a table

    Adding Conditional Control to Text-to-Image Diffusion Models. arXiv abs/2302.05543 (2023). Palette Aligned Image Diffusion (a) Images generated with different seeds for the prompt “A dress”. (b) Random images for “A waiting room with a sofa and a table. ” (c) Color distributio...

    2023 arXiv

  9. [1948]

    Bell System Technical Journal 27, 3 (1948), 379–423

    A Mathematical Theory of Communication. Bell System Technical Journal 27, 3 (1948), 379–423. doi:10.1002/j.1538-7305.1948.tb01338.x Gaurav Sharma, Wencheng Wu, and Edul N Dalal

    1948

  10. [2001]

    IEEE Computer Graphics and Applications 21 (10 2001), 34–41

    Color Transfer between Images. IEEE Computer Graphics and Applications 21 (10 2001), 34–41. doi:10.1109/38.946629 Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer

    2001 doi

  11. [2005]

    Color Research & Application 30, 1 (2005), 21–30

    The CIEDE2000 color- difference formula: Implementation notes, supplementary test data, and math- ematical observations. Color Research & Application 30, 1 (2005), 21–30. Sanzo Wada

    2005

  12. [2007]

    Computer Vision and Image Understanding 107, 1 (2007), 123–137

    Automated colour grading using colour distribution transfer. Computer Vision and Image Understanding 107, 1 (2007), 123–137. doi:10.1016/j.cviu.2006.11.011 Special issue on color image processing. Dustin Podell, Zion English, Kyle Lacey, Andreas Blattmann, Tim Dockhorn, Jonas ...

    2007 doi

  13. [2009]

    Fast and robust Earth Mover’s Distances. In Proc. ICCV. 460–467. doi:10.1109/ICCV.2009.5459199 Ofir Pele and Michael Werman

    2009

  14. [2015]

    ACM TOG (Proc

    Palette-based Photo Recoloring. ACM TOG (Proc. SIGGRAPH) 34, 4 (July 2015). Rakesh Kumar Chaudhary

    2015

  15. [2017]

    doi:10.1145/3092919.3092926 Huiwen Chang, Ohad Fried, Yiming Liu, Stephen DiVerdi, and Adam Finkelstein

  16. [2019]

    International Journal of Business Marketing and Management 4, 1 (2019), 4–9

    Influence of colour on visual arts. International Journal of Business Marketing and Management 4, 1 (2019), 4–9. Nadav Z. Cohen, Oron Nir, and Ariel Shamir

    2019

  17. [2022]

    arXiv abs/2207.12598 (2022)

    Classifier-Free Diffusion Guidance. arXiv abs/2207.12598 (2022). Lianghua Huang, Di Chen, Yu Liu, Yujun Shen, Deli Zhao, and Jingren Zhou

    2022 arXiv

  18. [2023]

    arXiv abs/2302.09778 (2023)

    Composer: Creative and Controllable Image Synthesis with Composable Conditions. arXiv abs/2302.09778 (2023). LAION. 2023a. LAION 2B English Aesthetic Dataset. https://huggingface.co/datasets/ laion/laion2B-en-aesthetic LAION. 2023b. laion/laion-art · Datasets at Hugging Face. ...

    2023 arXiv

  19. [2024]

    https://color.adobe.com/create/color-wheel

    Adobe Color. https://color.adobe.com/create/color-wheel. [Ac- cessed 24-10-2024]. Elad Aharoni, Yakov Shambik, and Dani Lischinski

    2024

  20. [2025]

    arXiv:2503.19034 [cs.CV] https: //arxiv.org/abs/2503.19034 Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie

    Color Conditional Generation with Sliced Wasserstein Guidance. arXiv:2503.19034 [cs.CV] https: //arxiv.org/abs/2503.19034 Chong Mou, Xintao Wang, Liangbin Xie, Yanze Wu, Jian Zhang, Zhongang Qi, Ying Shan, and Xiaohu Qie

    arXiv

  21. [6312]

    doi:10.1007/978-3-642-15552-9_54 Malsha V

    749–762. doi:10.1007/978-3-642-15552-9_54 Malsha V. Perera and Vishal M. Patel

    doi

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.