ziqiangoodgood/MIRAGE

MIRAGE: Benchmarking and Aligning Multi-Instance Image Editing

Ziqian Liu and Stephan Alaniz


Abstract: Instruction-guided image editing has seen remarkable progress with models like FLUX.2 and Qwen-Image-Edit, yet they still struggle with complex scenarios involving multiple similar instances, each requiring individual edits. We observe that state-of-the-art models suffer from severe over-editing and spatial misalignment when faced with multiple identical instances and composite instructions. To address this, we introduce a comprehensive benchmark specifically designed to evaluate fine-grained consistency in multi-instance and multi-instruction settings. We further propose Multi-Instance Regional Alignment via Guided Editing (MIRAGE), a training-free framework for precise, localized editing. By leveraging a vision-language model to decompose complex instructions into region-specific subsets, MIRAGE employs a multi-branch parallel denoising strategy that injects target-region latents into the global representation while preserving background integrity through a reference trajectory. Extensive evaluations on MIRAGE-Bench and RefEdit-Bench demonstrate that our framework significantly outperforms existing methods in achieving precise instance-level modifications while maintaining strong background consistency.

The MIRAGE-Bench contains 100 samples, each consisting of an image, a composite editing instruction formed by combining five sub-instructions, and the corresponding ground-truth mask. This benchmark is designed to evaluate image editing models in more complex referring-expression scenarios.

It can be loaded directly using:

from datasets import load_dataset
dataset = load_dataset("ziqiangoodgood/MIRAGE", split="test")

Citation

If you use this code or the benchmark in your research, please cite our paper:

@misc{liu2026miragebenchmarkingaligningmultiinstance,
      title={MIRAGE: Benchmarking and Aligning Multi-Instance Image Editing}, 
      author={Ziqian Liu and Stephan Alaniz},
      year={2026},
      eprint={2604.05180},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2604.05180}, 
}