acceptodds
Under review as a conference paper at ICLR 2027

Moore, Escher, Penrose: A Conformal Golden Braid

Abstract

"I don't think I have ever done anything as peculiar in my life. Among other things, it shows a young man looking with interest at a print on the wall of an exhibition that features himself. How can this be? Perhaps I am not far removed from Einstein's curved universe." This is how M.C. Escher described his 1956 litograph *Print Gallery*. This work of art has drawn mathematicians as much as artists. The image is a conformal power map of an ordinary picture. An ordinary generative model fails when prompted for such abnormality. This paper generates this form of art with off-the-shelf pretrained diffusion models. Applying the conformal transformation inside the sampling loop is not enough on its own: the model either "fixes” the twist as if it were damage, or drifts out of it. We therefore use the Moore–Penrose pseudo-inverse of the non-invertible map. The pair is idempotent and projects any input onto the image of . Even so, the denoiser would only ever see an out-of-distribution twisted picture. To solve this we interweave the diffusion steps with both directions of the transform: steps between and advance the underlying straight image, and steps between and realistically stitch the details of the transformed one. The Penrose identity guarantees that the result is a transformed image. We extend the construction to other conformal transformations, and obtain a tool for artistic, deliberately unrealistic image generation.

Primary area:
generative models
Keywords:
Diffusion, Conditioning, Escher, Art, Conformal Map

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_vudNOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a training-free approach to generating Escher-style recursive scenes using a frozen text-to-image model. It constructs a matched forward transformation and a reflexive generalized inverse for scale-recursive geometry, and alternates denoising between untwisted and transformed representations. The sampling procedure additionally uses recursive warm-up and in-loop super-resolution. The paper validates the approach through qualitative galleries, cumulative component ablations, a checkerboard-based geometric round-trip experiment, and qualitative transfer to a pixel-space backbone.

Critical strengths

A creative and technically meaningful geometric construction. The matched extraction and extension operators provide an explicit way to restore recursive structure without repeatedly compounding the forward warp. This is a useful contribution to controlling a frozen generative model under the particular geometry considered.

Visually engaging results with evidence of applicability beyond a single setting. The examples cover different scene types and transformations, and the PixelDiT results provide preliminary evidence that the procedure is not restricted to the main latent-space backbone. These demonstrations make the proposed application compelling.

Critical weaknesses
  1. The final outputs lack quantitative evaluation of recursive content consistency and geometric consistency. This is my primary concern. First, a nested composition does not necessarily mean that the inner copies faithfully reproduce the outer content. The paper does not quantify how consistently subjects, objects, spatial layouts, and textures are preserved across copies, or how this consistency changes across recursion levels. Second, it does not quantify how closely the final generated images satisfy the prescribed scale-and-rotation correspondences. The checkerboard experiment in Appendix D evaluates the round-trip properties of the geometric operators, rather than the final generated images, and therefore cannot substitute for such an evaluation. Consequently, despite the visually appealing examples, recursive content preservation and geometric consistency in the final outputs remain insufficiently supported by systematic empirical evidence.
  2. The independent benefit and reliability of braided sampling are insufficiently established. Figure 8 presents cumulative component comparisons for only two prompt–seed instances. Introducing source-space denoising also introduces the inverse transformation, additional re-noising, and a later forward transformation. Consequently, the comparison does not isolate whether the improvement comes from denoising in the source representation or from the additional geometric interventions. A comparison against transformed-space-only denoising with projection at matched intervention times is missing. The continuous operator identities explain the geometric construction, but do not resolve this empirical question. Although the additional galleries broaden the feasibility demonstration, the paper does not report aggregate performance across prompts and seeds that would establish the magnitude and reliability of the benefit. Together, these limitations leave the central sampling contribution insufficiently substantiated. The current evidence demonstrates promising examples, but does not adequately establish how well the complete method preserves its intended constraints or how much its alternating design contributes.
Questions for authors
  1. Final-output consistency: Have you quantitatively evaluated whether the inner copies faithfully reproduce the outer content, and how this consistency changes across resolvable recursion levels? Additionally, how closely do the final generated images satisfy the prescribed scale-and-rotation correspondences? Measurements across multiple prompts and seeds would help establish both recursive content consistency and final geometric consistency.

  2. Benefit of source-space denoising: Is there a controlled comparison between braided sampling and denoising exclusively in the transformed representation, using PE=TT†P_E=TT^\dagger to restore the constraint at matched intervention times? Matching the denoising budget, re-noising schedule, and super-resolution policy would help determine whether source-space denoising provides an independent benefit.

Reviewer_o4VTOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces a training-free approach for generating Escher-inspired, self-referential images using pretrained diffusion models. The method combines conformal geometric transformations with a matched generalized inverse to enforce recursive image structures. Through a braided diffusion sampling strategy, denoising alternates between an untwisted, scale-repeating representation and its geometrically transformed counterpart, allowing the scene and its recursive structure to develop together. The approach produces visually compelling results across several transformation families without requiring additional model training.

Critical strengths
  • The results presented in the paper are pure art. The generated images are visually stunning, creative, and inspiring. They demonstrate a delightful intersection of mathematics, generative modeling, and artistic expression.
  • The paper itself is benefits from artistic writing. The playful exposition, Escher references, and self-referential humor make it a genuinely enjoyable read. The title is fantastic, and the homage to Gödel, Escher, Bach: An Eternal Golden Braid is particularly fitting. I would love to see more of this creativity and personality in scientific writing.
  • An excellent introduction. The introduction is engaging, well-motivated, and beautifully connects Escher's artistic exploration to the proposed mathematical construction. It captures the reader's interest while providing a compelling motivation for the work.
  • Related work is also well explored and categorized.
Critical weaknesses

Section 2:

  • The mathematical exposition is occasionally convoluted and could benefit from a more intuitive presentation. The notation is dense, and the motivation behind certain definitions is not always immediately apparent. For example, $L=\log\lambda$ is introduced in L149, without initially explaining its geometric significance. More importantly, the intuition behind the complex power mapping could be developed before introducing its formal construction. The exposition sometimes focuses on mathematical details before establishing the bigger picture.
  • Additional geometric visualizations would greatly improve readability. In particular, detailed coordinate-grid illustrations showing how the logarithmic coordinates and conformal power mappings transform the image would be extremely valuable. Figures similar to those in de Smit & Lenstra (2003), or adaptations of the constructions in the Appendix, could help readers build intuition without having to mentally reconstruct the transformations.
  • The discussion of Moore-Penrose generalized inverses could be better motivated. It is not immediately clear why the reader needs the full mathematical background, particularly the final two Penrose identities for the non-linear mapping, given that the proposed construction relies on the reflexive identities. A more direct explanation of the desired properties and their role in the method would make this section easier to follow.
  • The Droste and Escher representations should be defined more explicitly. These are central concepts throughout the manuscript, particularly in Figure 3, but their definitions and differences are somewhat buried deep in the text. A clear, intuitive introduction to both representations, ideally accompanied by a dedicated illustration, would considerably improve readability.
  • Add the original reference for the SPNN acronym, since it first appears on the related work but it should be referenced again.

Section 3:

  • Some of the text could be moved to the Appendix. I would favor that to explain more clearly the paper idea in broader, bigger-picture, more accessible terms. Drop the math, or explain it more intuitively.
Questions for authors
  • Not clear to me if the time-travelling is applied per block and how many times. Did you apply time-travelling going back from the last denoising steps all the way to the beginning? Did you ablate on the time-travelling parameters?
  • Can you show more of the paper limitations? Cases that it wouldn't work, too extreme deformations, etc? I think its important to include these on the supplemental material.
  • For the final manuscript version, can you include tests with better more SOTA diffusion networks? Krea2 (this one is particularly good and fast), Z-image, Qwen, etc. Flux 1.0 is such an outdated model. Also it works a validation that the method works across different models.
  • Sampling with 128 steps seems overkill. This should be ablated for improving the performance of the method.
Reviewer_hUZ4OpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes to create artistic (Escher Rotating picture) images during diffusion sampling .

Critical strengths

The proposed paper is very creative, and reminds me a shock when I first met the paper 'Visual Anagrams' (CVPR 2024). This paper has similar motivation of Visual Anagrams, but another intriguing creative idea. I was impressed by the authors idea and their paper presentation. The visual outputs are impressive and their idea and analysis is interesting.

Critical weaknesses

I don't have fine any critical weakness on this paper.

One possible concern is that for the papers similar to this idea (Visual Anagram, Image as Sound, Factorized Diffusion ..) they commonly suffer from its low success rate. Is there any analysis on how frequently the failure cases happening?

Questions for authors

See weakness

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision summary and new Krea 2 results

Revision summary and new Krea 2 results

We thank all three reviewers for their thoughtful feedback and constructive discussion. Their suggestions have substantially strengthened the paper. Our latest revision additionally includes Krea 2 Raw results in the new Figure 14, extending Appendix G.2, “Transfer to More Models,” beyond the existing PixelDiT experiments.

For convenience, the main additions across the revision are:

  • Systematic final-image evaluation: all 120 outputs from 30 prompts, two seeds, and two twist settings, assessing recursive correspondence, independently recovered geometry, and AI-assisted visual and semantic criteria (new Section 4.4 and new Table 2). The new Appendix F.1 and new Tables 8–12 provide the full protocols, detailed results, and sensitivity checks.
  • Braiding versus projection: 62 paired settings across 35 scenes, showing lower average first-level disagreement with braiding across the reported first-level measures (new Appendix F.2, new Table 13, and new Figure 11).
  • Sampling cost and limitations: comparisons at 128, 64, and 32 steps, with numerical measurements and visual examples (new Appendix F.3, new Tables 14–15, and new Figure 12), plus expanded failure examples (new Appendix G.1 and new Figure 13).
  • Clearer exposition: revised Sections 2–3, explicit sampling details, and geometric illustrations in the new Figures 2 and 17.
  • Transfer across models: the new Krea 2 examples complement the FLUX and PixelDiT results, demonstrating the method on three frozen backbones without additional model training.

We appreciate the reviewers’ recognition of these improvements and hope the additional evidence is helpful in the remaining discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.