Ankit Dhiman 1,2* · Manan Shah 1* · R Venkatesh Babu 1
* Equal Contribution
1 Vision and AI Lab, IISc Bangalore
2 Samsung R & D Institute India - Bangalore

MirrorVerse builds upon our prior work Reflecting Reality, pushing the frontier of mirror reflection generation by adding diversity in the synthetic dataset creation pipeline and leveraging curriculum learning for generalizing to real-world scenes.
We introduce SynMirrorV2, a large-scale synthetic dataset containing 207K samples with full scene geometry, including depth maps, normal maps, and segmentation masks. SynMirrorV2 has high-fidelity training samples featuring variable object poses, occlusions, and multi-object setups.
- 📦 SynMirrorV2 Dataset: 207K synthetic samples with diverse object configurations and camera poses.
- 🧩 Curriculum Learning Strategy: a curriculum learning strategy that progressively adapts to complex scenarios, enabling state-of-the-art model to generalize better to real-world reflections.
- 🖼️ Multi-object Reflection Generation: First approach to effectively handle complex multi-object mirror scenes.
- 📊 Robust Benchmarks: Demonstrates strong quantitative and qualitative gains over previous SOTA.
- [14/07/2025] 🔥
Release the SynMirrorV2 Dataset - Release 🔥
checkpoints trained on SynMirrorV2Link - [07/06/2025] 🔥
Release codebase for creating synthetic datasetLink - [] Add interactive notebook demo for inference
The following table summarizes the key checkpoints mentioned in the project, along with their links and descriptions.
Checkpoint Name | Link | Description |
---|---|---|
MirrorFusion-v2 | Google Drive | This checkpoint is trained on single and multiple objects from SynMirrorV2. |
MirrorFusion-v2-MSD | Google Drive | This checkpoint is finetuned on real-world MSD dataset. |
@inproceedings{dhiman2025mirrorverse,
title={MirrorVerse: Pushing Diffusion Models to Realistically Reflect the World},
author={Dhiman, Ankit and Shah, Manan and Babu, R Venkatesh},
booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
pages={11239--11249},
year={2025}
}
This work builds on the foundation of Reflecting Reality. We also thank the developers of BlenderProc, diffusers, and SAM for their amazing tools and libraries.