Researchers from Tsinghua University and other institutions have presented AutoMIA, an AI design system built to automatically generate Mirror Illusion Art from two 2D images. In the setup described by the paper, a user specifies one image for the object seen in front of a mirror and another for the image seen in the reflection, and the system produces a 3D design that can also be fabricated through 3D printing.
The paper is available at http://arxiv.org/abs/2607.02015, and the code has been released at https://github.com/zxp555/AutoMIA. The work has been accepted to CVPR 2026 as a Highlight Paper. It also received the Efficient CVPR award, which the article linked to its computational efficiency.
One object, two different visual identities
The examples shown in the paper focus on a specific visual effect: the pattern visible on the object itself differs sharply from what appears in its mirror reflection. In one case, the front view reads “CVPR,” while the mirror image reads “2026.” In another, the front shows the Chinese character concept for “mountain,” while the reflected image shows “water.”
The authors attribute the effect to a specially designed 3D object and a feature of human vision described in the article as a symmetry completion effect. Under this setup, direct observation and mirrored observation produce two different perceptual results. This class of work is referred to as Mirror Illusion Art.
Designing such pieces has usually required specialist artistic intuition, repeated manual adjustment, and substantial mathematical work. AutoMIA is positioned as a way to automate that process and lower the barrier for turning a user’s visual idea into a manufacturable object.
Why the team built it
The paper places the work within the broader field of optical illusion art, which spans art, psychology, and computer vision. Common 2D illusions include hybrid images and visual distortion images. On the 3D side, prior forms include Shadow Art and Multi-View Wire Art.
Those approaches generally rely on viewpoint, projection, or shadow relationships so that one object can appear differently under different viewing conditions. Mirror Illusion Art is a narrower problem: a real object is placed in front of a mirror, one pattern is seen directly, and another appears in the reflection.
According to the article, earlier design methods often depended on manual derivation and hand tuning, and many focused mostly on shape rather than full-color patterns. Borrowing directly from Shadow Art-style methods could also introduce surface noise, floating background artifacts, and internal fractures, all of which hurt visual quality and make physical fabrication harder.
That led to the paper’s central question: given two target images, can AI automatically optimize both 3D shape and color so the same object presents two specified appearances from two views?
How AutoMIA formulates the task
AutoMIA treats the problem as a dual-view inverse design task. The 3D object is represented with voxels, and each voxel contains density and color information. The system then uses differentiable volumetric rendering to render the object from two specified views and compares the results with the two target images.
The optimization objective includes both shape similarity and color similarity. That means the generated object is trained to match the target contours while also reproducing the intended color patterns.
During optimization, the researchers identified four recurring problems: surface noise, background noise, internal breakage, and an imbalance between optimizing geometry and optimizing color. The paper introduces four mechanisms to address them.
PAC: projection-aligned connected-component pruning
PAC projects each connected component in the current voxel structure into the two target views, scores it by overlap with the target mask and by background spillover, then keeps structures that contribute to the intended pattern and removes floating or ineffective noise.
PWA: position-based weighted adaptive denoising
PWA adjusts the background penalty by the distance between a pixel and the target region. Wrong projections farther from the target receive a stronger penalty, which helps suppress remote background noise.
IVP: internal voxel preservation
Because supervision from the target images mainly constrains the object surface, internal voxels can be weakened during optimization and cause structural breakage. IVP identifies internal voxels and applies a density floor so the object remains intact, which is especially important for later 3D printing.
SCD: shape-color decoupled optimization
AutoMIA splits training into three stages: shape optimization, joint shape-and-color optimization, and color refinement. The idea is to prevent color from interfering with geometry too early and to reduce color conflicts across the two views.
Experimental results against prior methods
The paper evaluates AutoMIA with several classes of 2D inputs, including English letters, numbers, Chinese characters, geometric patterns, emoji, cartoon graphics, and logos.
For comparison, the team tested AutoMIA against Shadow Art, or SA, and Shadow Art Revisited, or SAR. Since those baseline methods mainly support shape optimization, the article said the comparison centered on smoothness, noise level, and shape consistency.
In the reported results, AutoMIA achieved a smoothness score of 0.989, a noise level of 0.049, and a shape similarity score of 0.931. SA and SAR, by comparison, showed clear gaps in both shape consistency and noise control.
On efficiency, the system posted an average design time of 76 seconds on a single NVIDIA RTX 3090, with average memory use of about 2.6 GB. The article said this high efficiency was one reason the work received the Efficient CVPR award.
The visual comparisons in the paper show SA and SAR producing more floating noise, rough surfaces, and internal fractures, which makes it harder to form a stable and complete physical object. AutoMIA, in contrast, better matches the contours of both target images and reconstructs colored surfaces that preserve clear visual meaning in both the direct view and the mirror view.
That point matters for this category of illusion art. If the shape is correct but color is chaotic, or if color is close but the structure is broken, the final illusion becomes much less convincing.
Ablation studies, simulation, and printed objects
The ablation study reported in the paper shows that PAC, PWA, IVP, and SCD all contribute to the final result. Removing any one of them reduces reconstruction quality. The article gives two examples: removing PAC significantly increases noise and lowers shape similarity, while removing SCD harms shape, color, and smoothness.
The team used those results to argue that AutoMIA is not just plain voxel optimization, but a system designed around the specific failure modes of mirror-illusion generation.
The paper also includes a set of examples in a digital environment. According to the article, the mirror-illusion objects were rendered with strong results in Blender-based simulation.
Beyond simulation, the authors presented several physical fabrication cases. They converted the optimized voxel models into 3D-printable formats and produced real objects with a 3D printer. In the photographs, the front view of the object and its mirrored view display different patterns, showing that the method can move from digital design into the physical world.
The article adds that users can apply the method to create various forms of artwork.
Authors and conclusion
The researchers describe AutoMIA as an AI system for automatically generating Mirror Illusion Art. Starting from two 2D images, it jointly optimizes 3D shape and color, and uses PAC, PWA, IVP, and SCD to improve quality so the generated object is smoother, more complete, and better suited to real-world fabrication.
The listed authors are Zhu Xiaopei, a Shuimu Scholar at Tsinghua University whose collaborating advisor is Professor Zhu Jun; Li Zeyuan, a co-first author and undergraduate at Huazhong University of Science and Technology; and corresponding authors Professor Zhu Jun and Associate Professor Hu Xiaolin of Tsinghua University.
As summarized in the article, the project points to another use of AI in computational design and artistic creation. It does not stop at image generation. It designs 3D illusion objects that can be physically realized, with one form in reality and another visual identity in the mirror. The paper frames that result as opening new directions for visual perception, inverse graphics, and digital fabrication.
This article cites the WeChat official account “Xinzhiyuan” as the original source. The author is listed as Xinzhiyuan and the editor as LRST.