[Overview] The AutoMIA system developed by the Tsinghua University team can transform any two images into a 3D artwork with a mirror illusion effect, enabling distinct visual patterns on each side and supporting 3D printing.
Look at the two sets of photos below: In each set, the left image shows the front pattern of an object in front of a mirror, and the right image shows its reflection. Strangely, the front and its reflection appear to be almost entirely unrelated— in the first set, the front reads "CVPR," while the reflection shows "2026"; in the second set, the front shows "mountain," while the reflection shows "water."
Their outlines, colors, and even meanings don’t match—it’s like an “illusion.”
The reason is that it comes from a specially designed 3D object that exploits the human visual system’s “symmetry completion effect” to create an illusion; this type of phenomenon is known as mirror illusion art.
Researchers designed the three-dimensional geometry and surface colors of a real object to create distinctly different visual effects when observed directly versus in a mirror.
Here are two sets of corresponding "decrypt" videos.
Traditionally, designing such artworks requires skilled artists to apply creative insight, extensive manual design experience, and complex mathematical calculations, making the design process highly inaccessible to the average person.
Is there a way to automate the process of helping people create such artwork, allowing ordinary individuals to turn their inspiration into art?
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—one representing the pattern in front of a mirror and the other representing its reflection—and AutoMIA will automatically design a 3D artwork, ready for direct 3D printing.
Paper link: http://arxiv.org/abs/2607.02015
Code link: https://github.com/zxp555/AutoMIA
The paper has been accepted by CVPR 2026 and designated as a Highlight Paper; additionally, due to its high computational efficiency, it has received the Efficient CVPR Award.
Research Background
Optical illusion art has long been a significant topic at the intersection of art, psychology, and computer vision. Common 2D illusions include blended images and visual distortion images; 3D illusions include Shadow Art and Multi-View Wire Art.
These methods typically rely on specific viewpoints, projections, or shadow relationships to make the same object appear differently under varying observation conditions.
Mirror Illusion Art is a more specialized form of 3D illusion: when a real object is placed in front of a mirror, the observer sees one pattern from the front and a different pattern in the reflection.
Previously, related designs often required extensive manual derivation and hand-tuning, with a focus primarily on shape design, making it difficult to simultaneously handle color patterns. Directly adopting methods such as Shadow Art could easily result in surface noise, background blobs, or internal fractures, compromising visual quality and physical fabrication.
Therefore, the author poses a direct question: Can AI automatically design such mirrored illusion objects? Given two target images, can the system simultaneously optimize the 3D shape and color so that the same object appears as two specified appearances from two different viewpoints?
Research methodology
AutoMIA models this problem as a two-view inverse design task. The system represents 3D objects using voxels, each containing density and color information; it then differentiably renders the object from two specified viewpoints and compares the rendered results with two target images.
The optimization goal includes both shape similarity and color similarity, ensuring that the generated object meets the contour requirements while also matching the color of the target image.
In practical optimization, the author identified four key issues: surface noise, background noise, internal fractures, and imbalance in shape and color optimization.
To this end, the paper proposes four core mechanisms.
First is PAC, or Projection-Based Alignment Connected Component Pruning. The system projects each connected component in the current voxel into two viewpoints, scores them based on their overlap with the target mask and background leakage, and retains only structures that genuinely contribute to the target pattern, removing floating or invalid noise components.
Second is PWA, or Position-Weighted Adaptive Noise Suppression. It adjusts the background penalty weight based on the distance of each pixel from the target region, applying stronger penalties to erroneous projections far from the target pattern, thereby reducing background noise at a distance.
Third is IVP, or Internal Voxel Protection. Since supervised images primarily constrain object surfaces, internal voxels are prone to being incorrectly weakened during optimization, leading to structural fractures. IVP identifies internal voxels and applies a minimum density threshold to help preserve object integrity, which is especially critical for subsequent 3D printing.
Finally, there is SCD, or Shape-Color Decoupling Optimization. AutoMIA divides the training process into three stages: shape optimization, joint shape-color optimization, and color refinement, preventing color from interfering with geometric structure too early and reducing conflicts between colors across different viewpoints.
Experimental results
AutoMIA was validated using multiple classes of 2D images as input, including English letters, numbers, Chinese characters, geometric patterns, emojis, cartoon illustrations, and logos.
In the experiment, the authors compared AutoMIA with methods such as Shadow Art (SA) and Shadow Art Revisited (SAR). Since the baseline methods primarily support shape optimization, the comparison focused on metrics including smoothness, noise level, and shape consistency.
The results show that AutoMIA achieves a clear advantage in reconstruction quality. According to the paper, AutoMIA achieves a Smoothness Level (SL) of 0.989, reduces Noise Level (NL) to 0.049, and attains a Shape Similarity (SS) of 0.931; in comparison, SA and SAR exhibit significant gaps in shape consistency and noise control.
Meanwhile, AutoMIA achieves an average design time of just 76 seconds per design on a single RTX 3090, with an average GPU memory usage of approximately 2.6 GB, demonstrating high computational efficiency.
From the visualization results, SA and SAR tend to produce floating noise, uneven surfaces, or internal fractures, making it difficult to form a stable and complete solid structure.
The results generated by AutoMIA not only better match the contours of the two target images but also reconstruct the colored surfaces of the input images, ensuring clear visual semantics for both frontal and mirror perspectives.
This is especially critical for mirror illusion art: if the shape is correct but the colors are chaotic, or if the colors are close but the structure is fragmented, the final result will struggle to create a convincing illusion.
Ablation studies further demonstrate that PAC, PWA, IVP, and SCD all contribute to the final results; removing any module leads to a decline in reconstruction quality. For instance, removing PAC significantly increases noise and reduces shape consistency, while removing SCD affects shape, color, and smoothness.
This indicates that AutoMIA is not merely reliant on voxel optimization, but rather features a systematic design addressing issues specific to mirror hallucinations.
Below are examples from the digital world; as shown in the figure, well-rendered mirrored illusion objects have been created using the Blender digital simulation software.
In addition to numerical results, the paper presents several physical fabrication case studies. The authors converted the optimized voxel models into 3D-printable formats and physically produced real objects using 3D printing.
In the physical photograph, the object's front and its reflection display different patterns, demonstrating that this method successfully bridges digital designs with the physical world.
Users can use this method to create various types of artwork.
Conclusion and Outlook
Researchers have developed AutoMIA, an AI design system for automatically generating Mirror Illusion Art.
It can jointly optimize the 3D object's shape and color from two 2D images, and enhance result quality through mechanisms such as PAC, PWA, IVP, and SCD, producing smoother, more complete outputs better suited for real-world fabrication.
This work demonstrates new possibilities for AI in computational design and artistic creation: it can not only generate images but also design 3D illusory objects that are physically realizable. An object that exists in reality with only one form yet presents a different visual identity in a mirror—this “discrepancy between front view and reflection”—opens new research directions in visual perception, inverse graphics, and digital fabrication.
Author Bio
The authors of the paper are, in order: Zhu Xiaopei, a Tsinghua University Shuimu Scholar, with Professor Zhu Jun as her mentor; Li Zeyuan (co-first author), an undergraduate student at Huazhong University of Science and Technology; and Professor Zhu Jun and Associate Professor Hu Xiaolin (corresponding author) from Tsinghua University.
Reference: http://arxiv.org/abs/2607.02015
This article is from the WeChat public account "AI New Era," authored by AI New Era; edited by LRST.