README.md

---
license: apache-2.0
task_categories:
- image-to-image
language:
- en
tags:
- image-editing
size_categories:
- n<1K
---
<div align="center">

# ProductConsistency Benchmark

</div>

<div align="center">

[![model](https://img.shields.io/badge/HFModel-Kalaido--qwenedit--lora-red?logo=huggingface)](https://huggingface.co/FractalAIResearch/Kalaido-qwenedit-lora)
[![model](https://img.shields.io/badge/HFModel-Kalaido--qwen--2512--lora-red?logo=huggingface)](https://huggingface.co/FractalAIResearch/Kalaido-qwen-2512-lora)

</div>

<div align="center">

[![report](https://img.shields.io/badge/📄_Technical_Report-ProductConsistency-blue?logo=arxiv)](https://arxiv.org/abs/2606.19103)
[![Dataset License](https://img.shields.io/badge/license-Apache_2.0-green.svg)](LICENSE)

</div>

<div align="center">

<img src="assets/distribution.png" alt="ProductConsistency Benchmark Distribution" width="70%">

</div>

---

## Dataset

The ProductConsistency Benchmark is a high-quality dataset for instruction-based, product-centric image editing in English. Every record is one evaluation tuple: an input product image plus a single edit instruction describing an advertisement-style scene change.

- 870 samples formed from 174 synthetic product images (8 product categories) with 5 distinct edit instructions per image.
- Images emphasize legible on-pack text and consistent branding; instructions are written to stress identity preservation under realistic marketing-style edits.



Benchmark images are drawn with uniform coverage over categories and over the range of rendered text lengths, using the same synthetic product generation and OCR-based text verification pipeline as the broader ProductConsistency release. For each selected product image, a language model generates five edit instructions that are unique to that image. To ensure dataset quality, a final human verification step confirms that on-product text in each benchmark image matches the intended ground truth before instructions are paired and released.

---

## Evaluation on this benchmark

| Model | CER ↓ | Seg CLIP-I ↑ | Seg DINO-I ↑ | Product consistency ↑ | Aesthetics ↑ | Text fidelity ↑ | Overall ↑ |
|-------|------:|-------------:|-------------:|------------------------:|-------------:|-----------------:|----------:|
| HiDream-E1-1 | 3.8774 | 0.8390 | 0.7240 | 6.5828 | 7.4134 | 3.4477 | 5.8146 |
| OmniGen2 | 1.7094 | 0.8858 | 0.7790 | 7.6739 | 7.8613 | 4.9908 | 6.8422 |
| BAGEL | 1.6810 | 0.8767 | 0.7515 | 7.7260 | 7.8088 | 6.2203 | 7.2520 |
| Step1x-edit-v1p2 | 1.1909 | 0.8626 | 0.7157 | 7.5812 | 7.4636 | 7.0414 | 7.3621 |
| RePlan-Flux | 0.2914 | 0.9174 | 0.8085 | 8.4118 | 6.8085 | 8.8727 | 8.0311 |
| RePlan-Qwen | 0.5164 | 0.9010 | 0.7419 | 7.6963 | 6.3391 | 7.6542 | 7.2298 |
| Nano Banana | 1.1868 | 0.8860 | 0.7020 | 8.8256 | 8.3552 | 8.0839 | 8.4167 |
| Qwen-Image-Lightning | 0.6073 | 0.8920 | 0.7680 | 8.5188 | 8.1941 | 8.1977 | 8.2738 |
| Edit-R1-Qwen | 0.4430 | 0.9046 | 0.7597 | 8.5834 | 8.3314 | 8.1542 | 8.3565 |
| Edit-R1-Flux | 0.1550 | 0.9195 | 0.7966 | 8.7015 | 8.0226 | 9.0142 | 8.5798 |
| GPT-Image-1 High | 0.3315 | 0.9080 | 0.7800 | 9.0134 | 8.5598 | 8.4077 | 8.6300 |
| Qwen-Image-Edit-2511 | 1.0682 | 0.8728 | 0.7080 | 8.4578 | 8.2467 | 7.5958 | 8.1003 |
| Qwen-Image-Edit-2511 + SFT + Cyclic Reward | 0.2080 | 0.9245 | 0.7990 | 8.8866 | 8.3373 | 8.8923 | 8.7055 |
| Flux.1-Kontext-dev | 0.1490 | 0.9210 | 0.8110 | 8.7111 | 7.9506 | 8.9096 | 8.5240 |
| Flux.1-Kontext-dev + SFT + Cyclic Reward | 0.1204 | 0.9224 | 0.8115 | 8.7996 | 7.9901 | 9.0740 | 8.6216 |

---

## Usage



```python
from datasets import load_dataset

ds = load_dataset("FractalAIResearch/ProductConsistencyBenchmark")
row = ds[0]
img, instr = row["image"], row["edit_instruction"]
```

---

## License

This dataset is released under the [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0)

---

## 📖 Citation

```bibtex
@misc{khanna2026productconsistencyimprovingproductidentity,
  title={ProductConsistency: Improving Product Identity Preservation in Instruction-Based Image Editing via SFT and RL},
  author={Mukund Khanna and Raj Singh Yadav and Kunal Singh},
  year={2026},
  eprint={2606.19103},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2606.19103},
}
```