acceptodds
Under review as a conference paper at ICLR 2027

Native Generation of Tileable Pixel-Art Textures from a Material Name and a Target Colour

Abstract

Game and mod textures are authored natively at tiny canvases: a 16×16 tile is not a downscaled photograph but a grid in which every pixel is a design decision, the palette holds at most a handful of slots, and the tile must stay continuous across its own wrap boundary because it is used tiled. We study the task of painting a solid-colour region of an image with a tileable pixel-art texture given only a material name and the region's colour, and we generate it natively at the target resolution rather than rendering large and downsampling. Our model decodes palette-index and luminance-rank tokens in parallel, keeps tileability by construction through a toroidal relative attention bias that adds no parameters, and conditions on palettes retrieved from a memory of artist palettes. Against the strongest render-and-downsample pipeline, annotators prefer our 16×16 tiles in 73.6% of decided pairs; on the region-colouring task the de-biased judge prefers them 60.3% (family interval excludes 0.5), and that judge's 57.3% on the name-to-texture comparison is a lower bound on the same human reading. The tiles also win the majority of judge comparisons against a same-data pixel diffusion model, score distillation, and artist-tile retrieval. Measured end to end, a tile takes 0.51 s, 28× faster than the strongest render pipeline and four orders of magnitude faster than score distillation. Getting these comparisons to mean anything required repairing the rulers: we show that render-and-downsample baselines collect a structural gain purely from a larger canvas, that a widely used tiling score awards a perfect value to a flat fill, and that a pairwise judge's resolvable-rate floor is far below chance.

Primary area:
generative models
Keywords:
pixel art generation, tileable texture synthesis, masked generative transformer, retrieval-augmented generation, color-conditioned image generation, relative position bias, evaluation of generative models

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_BgseOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents TRD, a fast model that draws seamless 16×16 pixel-art textures from a material name and a colour. It borrows real artists' palettes and generates the tile directly instead of shrinking a large image. It takes about 0.5 s per tile, 28× faster than the strongest render-and-shrink baseline. Humans preferred its tiles 73.6% of the time in a small study, but the AI judge's edge was borderline, and the advantage disappears at larger sizes.

Critical strengths

The paper defines a concrete, practical task for modders/indie developer community: given a material name and a target colour, paint a region with a seamless 16×16 pixel-art texture. The proposed TRD draws the tile directly at its final size, so it is much more efficient than the alternatives. it's about 28x faster than render-and-strink pipeline and way much faster than SD-piXL.

The evaluation is careful. It reports noise floors, checks the VLM judge on identical pairs, uses intervals that group related materials, tests only on artists unseen in training, and shows unselected samples.

The experimental results show that at 16×16, TRD gets the best image-quality score among the generators (KID 8.8, against 10.0 for PixelDiff and 28.9 for Render), the opus5 VLM judge consistently prefers it over PixelDiff, Render on the colour task, and SD-piXL, and ablations confirm that both palette retrieval and the wrap-around position signal matter.

Critical weaknesses

All the data comes from one source, with only 4000+ training tiles from limited packs. That is a little bit small and narrow to show the method generalizes to other games, art styles or pixel-art domains.

The human evaluation is small as well. Only 4 annotators and 91 decided pairs, covering only TRD+rr vs render, with no human study for the colour task or the other baselines.

Most of TRD's quality comes from retrieving artist palettes (removing retrieval raises KID from 9.7 to 35.1), and simply retrieving and recolouring an artist tile ties TRD on the colour task (53.4%, not significant), so the generative model adds limited value over retrieval.

The modeling side novelty is limited. The method mostly combines existing parts (MaskGIT decoding, a circular relative position bias and palette retrieval), and a variant that wraps at the wrong period scores just as well (KID 7.3 vs 8.4), suggesting the wrap-around design helps seams but not overall quality.

Questions for authors

Will you run a fresh human study with the fixed instructions, as your own notes proposed, and report it in place of the amended round?

Is there a human study for the colour task and for PixelDiff and retrieve, not just render?

How well does it generalise to other pack ecosystems or other artists' styles outside Luanti?

Reviewer_z8skOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents Toroidal Rank Decoding (TRD), a method for generating 16×16 tileable pixel-art textures from only a material name and a region colour. It generates the tile natively, instead of rendering a large image and downsampling it. The authors train a Masked Generative Image Transformer (MaskGIT) that decodes palette-index and luminance-rank tokens in parallel. The model ensures tileability by construction through a toroidal relative attention bias. It is also conditioned on palettes retrieved from a memory of artist palettes. The results suggest that this approach could outperform other baselines (e.g., a render-and-downsample method) on particular task at a much lower cost.

Critical strengths
  1. The proposed framework is well designed and fits to the artist workflow. The method lets the user choose a colour and a material name to generate pixel-art textures. Furthermore, it decouples colour from structure, so the colour can be changed without generating the tile again. It also generates the tile natively, which the authors show works better than first generating a texture at a higher resolution and then downsampling it to 16×16 pixel art.
  2. The method is fast and lightweight. Compared to diffusion method (i.e.,SD-πXL), which can take hours to generate one pixel-art image, TRD takes only less than a second. It also has significantly fewer number of parameters (only 33M).
Critical weaknesses
  1. The writing and figure formatting could be improved, particularly Fig.2 and some formatting breaks (e.g., L238), which make the paper look unpolished.
  2. The output quality is still questionable, given that the generated tiles need to follow both the colour and the material name given by the user. For example, in Fig.1(b), the generated textures don’t look plausible, as they lack the essence of the chosen material. Gold and emerald ore lose both the stone base and the gems, and log birch comes out as brown wooden planks. The tiles look like generic textures with a retrieved or user colour applied.
  3. Related work and baselines are not well explored for this task. Only one published prior method (SD-πXL) is compared, and it is still absent from the main quantitative and qualitative comparisons (e.g., Table 2 and Figs.7-8). The headline baselines, Render, Retrieve and PixelDiff (Sec.6.1), are all built by the authors. The authors should consider comparing with state-of-the-art tileable generation methods (e.g., ControlMat [1], Tiled Diffusion [2], etc.).
Questions for authors
  1. About baselines: Could the authors add recent published tileable generation methods (e.g., [1], [2]) to the main comparisons (Table 2 and Figs.7–8)? Some of them are only mentioned once, in Sec.2 (L105). Could the authors also consider adding a full quantitative and qualitative evaluation of SD-πXL?
  2. About fair assessment: The authors provide retrieved artist palettes only to TRD (L263–265), and this gives the largest gain: KID drops from 35.1 to 9.7, as shown in Table 4. How would other baselines (e.g., PixelDiff and Render) perform under the same condition (the full palette)?
  3. About potential data leakage: The authors mention that 448 out of 450 test materials already appear in training set, which drawn by other artists (L166). Since palettes are fetched by material name, the model can basically look up a training palette for almost every test material, so the test doesn’t really show whether TRD can pick right colours for a new material. How would TRD perform on materials it has never seen before, for example, on a material-disjoint split? Could the authors also report per-pack results across all 12 test packs (L164)?

[1] Vecchio et al. “ControlMat: A Controlled Generative Approach to Material Capture.” TOG 2024.

[2] Madar et al. “Tiled Diffusion.” CVPR 2025.

Reviewer_ADTtOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies native generation of small, tileable static textures conditioned on a text prompt and a region colour (platte). Different from baselines that follow a rendering large and downsampling protocal, this paper designs a model that directly generates in the native tile space. TRD adapts masked token generation to a representation combining palette and luminance rank, introduces toroidal relative attention, and retrieves artist palettes for conditioning. Evaluation uses a comprehensive corpus with generative and retrieval baselines, ablations, and human/VLM preferences, and achieves state-of-the-art performance in both quality and speed.

Critical strengths

The paper provides useful designs for native tile generation and palette retrieval, and proposes evaluation protocols for this task.

  • The problem is well-defined and the model design is reasonable. Directly generating a small palette-indexed grid is well suited to this application. In addition, the paper introduces a comprehensive evaluation protocol, including empirical noise floors, material-family confidence intervals, and human/VLM preferences.
  • The experiments support generating tiles in the native space. The paper illustrates the limitations of rendering large images and then downsampling them through concrete examples. Figure 3 shows how direct downsampling loses structure and produces blurry textures, while Figure 8 shows similar limitations even when a pixel-art LoRA is used before downsampling. TRD produces more coherent results in several of these examples.
  • A plausible way to handle periodic boundaries in MaskGIT architecture. The toroidal attention bias lets the model treat opposite edges of a tile as neighbours, bringing the intuition behind circular padding in a U-Net into a MaskGIT architecture. The ablations provide useful evidence that it helps in this setting.
Critical weaknesses
  • The qualitative results are not consistently convincing. In Figure 8, TRD+rr can indeed generate tileable textures with some detail. However, some examples show clear problems with material readability. For example, the results for "gold ore" and "emerald ore" lack clearly recognisable ore inclusions, while "log birch" lacks the characteristic light bark and dark markings of birch.
  • Insufficient human evaluation. The main user study includes only four retained participants, and most materials are judged by just one person. Human evaluation is especially important here, since the method is intended to help artists create pixel-art textures. With so few participants and little overlap in what they evaluate, it is hard to know how much the results reflect individual taste. A larger study with target users and several ratings per material would make the claimed benefits more convincing.
  • The benefit over retrieval is not yet clear. Retrieve performs well on the distribution metrics in Table 2, and TRD shows no clear preference advantage over retrieval plus recolouring on the main colour-conditioned task. Figure 8 also shows that some generated tiles lose the distinctive features that make the artist references easy to recognise, such as birch bark markings and diamond inclusions. This leaves a practical question: at 16×16, and under the evaluated task setting, when would an artist benefit from generating a new tile rather than retrieving and recolouring an existing one?
  • Paper writing. I could follow the content of the paper, but I recommend that the authors polish the writing, especially the abstract and introduction, to make the main claims more precise and easier to assess against the experimental evidence.
Questions for authors

See weaknesses.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.