Highlights of CVPR'26: Tsinghua University has developed a mirror illusion art generator that supports 3D printing.

【Introduction】The AutoMIA system developed by Tsinghua University research team can convert any two images into 3D artworks with mirror illusion effects, realizing visual presentation of different patterns on the front and mirrored sides, and supports 3D printing.

First, look at the two sets of photos below: in each set, the left side shows the front pattern of the object in front of the mirror, and the right side shows its reflection in the mirror. The odd thing is that the front view and the mirror view show almost two completely unrelated things — the first set shows "CVPR" on the front, but it becomes "2026" in the mirror; the second set shows "mountain" on the front, but it turns into "water" in the mirror.

Their outlines, colors and even meanings do not match, as if it were a kind of "illusion".

The underlying reason is that it comes from a specially designed 3D object, which uses the "symmetry completion effect" of the human visual system to create illusions, and this type of phenomenon is called mirror illusion art.

By carefully designing the 3D geometry and surface colors, the researchers made a real object present completely different visual effects when viewed directly and through a mirror.

The following are the two corresponding "decryption" videos.

Traditionally, the design of such artworks requires ingenious concepts from professional artists, as well as a large amount of complex manual design experience and mathematical calculations. The threshold for ordinary people to design such artworks is very high.

So is there a way to automatically help people complete such art designs, so that ordinary people can also turn their own inspirations into artworks?

Recently, a research team from Tsinghua University and other institutions proposed AutoMIA, an AI design method that automatically generates mirror illusion art. Users only need to specify any two images, corresponding to the patterns in front of the mirror and in the mirror respectively, and AutoMIA can automatically complete the design of 3D artworks, and supports direct 3D printing to produce the artworks.

Paper link: http://arxiv.org/abs/2607.02015

Code link: https://github.com/zxp555/AutoMIA

The paper has been accepted by CVPR 2026 and was rated as a Highlight Paper of the conference; at the same time, due to its high computational efficiency, this work also won the Efficient CVPR award.

Research Background

Optical illusion art has long been an important issue in the cross-domain of art, psychology and computer vision. Common 2D illusions include hybrid images, visually distorted images, etc.; 3D illusions include Shadow Art, Multi-View Wire Art, etc.

These methods usually rely on specific perspectives, projections or shadow relationships to make the same object show different appearances under different viewing conditions.

Mirror Illusion Art is a more special type of 3D illusion: when a real object is placed in front of a mirror, the observer sees one pattern from the front and another pattern in the mirror.

Previously, related designs often required a lot of manual derivation and manual adjustment, and mostly focused on shape design, making it difficult to process color patterns at the same time. If methods like Shadow Art are directly adopted, it is easy to produce surface noise, background floating blocks or internal fractures, which affect visual quality and physical production.

Therefore, the authors raised a direct question: can AI automatically design such mirror illusion objects? Given two target images, can the system optimize the 3D shape and color simultaneously to make the same object present two specified appearances from two perspectives?

Research Methodology

AutoMIA models this problem as a two-view inverse design task. The system represents the 3D object with voxels, and each voxel contains density and color information; then through differentiable volume rendering, the object is rendered from two specified perspectives, and the rendering results are compared with the two target images.

The optimization objective includes both shape similarity and color similarity, so that the generated object not only meets the contour requirements, but also can generate the colors of the target images.

In the actual optimization process, the authors found four key problems: surface noise, background noise, internal fracture, and the imbalance between shape and color optimization.

To this end, the paper proposes four core mechanisms.

The first is PAC, namely Projection-Aligned Connected Component Pruning. The system projects the connected components in the current voxel to two perspectives one by one, and scores them according to their overlap with the target mask and background overflow, retaining the structures that truly contribute to the target pattern, and removing floating or invalid noise parts.

The second is PWA, namely Position-based Weighted Adaptive Noise Suppression. It adjusts the background penalty weight according to the distance from the pixel to the target area, so that the wrong projections far away from the target pattern are more strongly penalized, thereby reducing distant background noise.

The third is IVP, namely Inner Voxel Protection. Since the supervised images mainly constrain the surface of the object, the inner voxels are easily incorrectly weakened during optimization, leading to structural fracture. IVP identifies the inner voxels and imposes a lower density limit to help maintain the integrity of the object, which is especially important for subsequent 3D printing.

The last is SCD, namely Shape-Color Decoupled Optimization. AutoMIA divides the training process into three stages: shape optimization, joint shape and color optimization, and color refinement, preventing colors from interfering with the geometric structure prematurely and reducing conflicts between colors from different perspectives.

Experimental Results

The paper verified AutoMIA with multiple types of 2D images as inputs, including English letters, numbers, Chinese characters, geometric patterns, emojis, cartoon patterns and logos.

In the experiments, the authors compared AutoMIA with methods such as Shadow Art (SA) and Shadow Art Revisited (SAR). Since the baseline methods mainly support shape optimization, the comparison focuses on indicators such as smoothness, noise level and shape consistency.

The results show that AutoMIA has achieved obvious advantages in reconstruction quality. According to the paper report, AutoMIA reaches a smoothness (SL) of 0.989, reduces the noise level (NL) to 0.049, and achieves a shape consistency (SS) of 0.931; in contrast, SA and SAR have obvious gaps in shape consistency and noise control.

At the same time, AutoMIA has an average design time of only 76 seconds on a single RTX 3090, with an average video memory usage of about 2.6GB, reflecting high computational efficiency.

From the visualization results, SA and SAR are prone to problems such as floating noise, uneven surfaces or internal fractures, making it difficult for the object to form a stable and complete physical structure.

The results generated by AutoMIA can not only better match the outlines of the two target images, but also reconstruct the colored surfaces of the input images, making both the front view and the mirror view have clear visual semantics.

This is particularly critical for mirror illusion art: if the shape is correct but the colors are chaotic, or the colors are close but the structure is broken, it will eventually be difficult to form a convincing illusion effect.

Ablation experiments further show that PAC, PWA, IVP and SCD all contribute to the final results. Removing any module will lead to a decline in reconstruction quality. For example, removing PAC will significantly increase noise and reduce shape consistency; removing SCD will affect shape, color and smoothing effects.

This shows that AutoMIA does not simply rely on voxel optimization, but is systematically designed for the problems that arise in mirror illusion design.

The following is an example from the digital world. As can be seen from the figure, a mirror illusion object with good effect is rendered in the Blender digital simulation software.

In addition to digital results, the paper also presents multiple physical production cases. The authors converted the optimized voxel model into a 3D-printable format and produced real objects through 3D printing.

In the physical photos, the front and the mirror view of the object show different patterns, indicating that this method can successfully move from digital design to the physical world.

Users can make various types of artworks based on this method.

Conclusion and Outlook

The researchers proposed AutoMIA, an AI design system for automatically generating Mirror Illusion Art.

It can start from two 2D images, jointly optimize the shape and color of 3D objects, and improve the quality of results through mechanisms such as PAC, PWA, IVP and SCD, making the generated objects smoother, more complete, and more suitable for actual production.

This work demonstrates the new possibilities of AI in computational design and artistic creation: it can not only generate images, but also design physically realizable 3D illusion objects. An object has only one form in reality, but can present another visual identity in the mirror. This "inconsistency between the front view and the mirror view" effect also provides new research directions for visual perception, inverse graphics and digital manufacturing.

Author Introduction

The authors of the paper are, in order: ZHU Xiaopei, a Shuimu Scholar at Tsinghua University, with co-supervisor Professor ZHU Jun; LI Zeyuan (co-first author), an undergraduate at Huazhong University of Science and Technology, and Professor ZHU Jun and Associate Professor HU Xiaolin (corresponding authors) from Tsinghua University.

References: http://arxiv.org/abs/2607.02015

This article is from the WeChat official account "AI Era", author: AI Era; editor: LRST, authorized by 36Kr for release.