PhysMirror: Physics-Aware Mirror Object Generation

Xuan-Bach Mai ^†

University of Science & VNU-HCM

Duy-Phuc Nguyen ^†

University of Science & VNU-HCM

Quoc-Van Le

University of Science & VNU-HCM

Tam V. Nguyen ^*

University of Dayton

Thanh-Toan Do

Monash University

Huu Le

VinFast & VinUniversity

Duong-Van Nguyen

VinFast & VinUniversity

Minh-Triet Tran

University of Science & VNU-HCM

Trung-Nghia Le ^‡,*

University of Science & VNU-HCM

^†Equal contribution, ^‡Corresponding author, ^*Co-supervisors

Paper (Coming Soon)

Code

Examples of Text-to-Image Generation with PhysMirror

Abstract

Synthesizing physically accurate mirror reflections remains a fundamental challenge for modern text-to-image diffusion models, which are increasingly critical for generating synthetic training data for embodied AI and robotic perception. These models typically struggle with strict geometric constraints, leading to hallucinations that degrade the utility of the synthetic data. To address this, we introduce a novel, end-to-end physics-aware generation framework namely PhysMirror that natively enforces projective geometry through explicit 3D spatial priors. Our method automatically lifts prompted objects into 3D meshes and constructs a lightweight, mathematically exact mirror scene within a simulated environment. By rendering this explicit 3D scene, we extract precise 2D conditioning elements, such as depth maps and segmentation maps, that serve as robust guiding signals for downstream diffusion models, guiding them to generate images with physically correct mirror reflections. Moreover, we introduce Mirror Consistency Score (MCS), a reference-free, fully automated metric that quantifies physical correctness using dense feature matching and vanishing point convergence. Experimental results on our newly constructed MirrOB dataset demonstrate that our approach outperforms state-of-the-art baselines in reflection accuracy and physical realism, while maintaining strong text-to-image semantic alignment, providing a reliable pipeline for embodied AI data generation.

Method Overview

We propose PhysMirror, a modular pipeline that transitions from a text prompt to a physically grounded 3D scene, then projects back to 2D spatial conditioning maps to guide photorealistic image generation. As illustrated in the figure below, the pipeline consists of four stages: (1) lifting prompted objects into 3D meshes, (2) composing a physically accurate mirror scene, (3) rendering conditioning maps from a carefully chosen viewpoint, and (4) guiding a text-to-image generative model with these spatial priors.

Overview of the proposed PhysMirror pipeline. Given a text prompt, primary objects are parsed and lifted into 3D meshes via a text-to-3D model (Stage 1). These meshes are assembled into a physically accurate 3D mirror scene (Stage 2). A virtual camera renders precise spatial maps such as depth and segmentation (Stage 3). Finally, conditioning maps guide a text-to-image model to synthesize the final image (Stage 4). — Overview of the proposed PhysMirror. Given a text prompt, primary objects are parsed and lifted into 3D meshes via a text-to-3D model (Stage 1). These meshes are then assembled into a physically accurate 3D mirror scene (Stage 2). Next, a virtual camera is positioned to render precise spatial maps, such as depth and segmentation (Stage 3). Finally, these conditioning maps guide a text-to-image generative model to synthesize the final photorealistic image with geometrically correct mirror reflections (Stage 4).

Mirror Consistency Score

To quantitatively evaluate the geometric and physical correctness of mirror reflections without requiring human annotations, we propose the Mirror Consistency Score (MCS). Our metric leverages zero-shot segmentation and dense part-level feature matching to establish corresponding keypoints between a real object and its reflection. Based on the principles of projective geometry, the lines connecting these matched keypoint pairs must converge at a single vanishing point in a 2D image. MCS evaluates the tightness of this intersection cluster to quantify reflection accuracy and physical realism.

Visualization Results

Our physics-aware method consistently generates geometrically correct reflections compared to baseline models, maintaining structural coherence and physical correctness.

Qualitative comparison of synthesized mirror reflections across varying scene complexities. — **Qualitative comparisons.** (a) Unconditioned baselines (SDXL and FLUX.1-dev) frequently fail to produce physically accurate reflections, whereas our physics-aware method consistently generates geometrically correct results. (b) Ablation study comparing alternative spatial conditioning strategies, demonstrating how segmentation priors (Seg2Any) and custom fine-tuning (SynMirror LoRA) can introduce visual artifacts or perspective errors compared to our primary method.

Quantitative Results

We evaluate the unconditioned baseline models against our proposed physics-aware pipeline across varying levels of scene complexity (1, 2, and 3 objects).

Method	MCS 1 Obj ↑	MCS 2 Obj ↑	MCS 3 Obj ↑	MCS Overall ↑	CLIP Score ↑	CLIP-IQA ↑	MANIQA ↑	MUSIQ ↑
SDXL	0.480	0.580	0.548	0.527	33.284	0.6499	0.5114	73.4996
FLUX.1-dev	0.476	0.689	0.670	0.589	33.361	0.6051	0.4733	70.5174
PhysMirror (Ours)	0.727	0.779	0.744	0.746	28.914	0.6589	0.4929	70.7800

BibTeX

@inproceedings{physmirror2026,
  author = {Mai, Xuan-Bach and Nguyen, Duy-Phuc and Le, Quoc-Van and Nguyen, Tam V. and Do, Thanh-Toan and Le, Huu and Nguyen, Duong-Van and Tran, Minh-Triet and Le, Trung-Nghia},
  title = {PhysMirror: Physics-Aware Mirror Object Generation},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year = {2026},
}

Acknowledgment

This research is funded by Vietnam National University - Ho Chi Minh City (VNU-HCM).