README.md

---
language:
- en
base_model:
- BAAI/Emu3-Stage1
---
# EARL - SFT think (S) (8B)

**Model Size:** 8B parameters  
**Base Model:** [BAAI/Emu3-Stage1](https://huggingface.co/BAAI/Emu3-Stage1)  
**Dataset:** Simple Edit  
**Training Objective:** Supervised Fine-Tuning (SFT) with Chain-of-Thought reasoning  

This model is introduced in our paper: [EARL: The Promise of RL for Autoregressive Image Editing](https://arxiv.org/abs/2508.01119).

## Overview

EARL - SFT think (S) is a fine-tuned 8B vision-language model designed for autoregressive image editing. It extends the base Emu3 model with **chain-of-thought supervision**, enabling step-by-step reasoning to tackle complex editing tasks. Training leverages the Simple Edit dataset, focusing on editable instructions grounded in visual understanding.

🔗 **Inference script and usage:** [GitHub Repository](https://github.com/saba96/EARL?tab=readme-ov-file)

## Benchmark Results

| Model             | OmniEdit | EmuEdit | AURORA | MB   | VisMin | I2EBench | **AVG** |
|------------------|----------|---------|--------|------|--------|----------|---------|
| **SFT (S)**       | 5.73     | 3.66    | 3.58   | 3.19 | 3.57   | 3.59     | **3.88** |
| **SFT think (S)** | 4.34     | 3.76    | 2.88   | 3.36 | 3.46   | 3.21     | **3.50** |

> ⚠️ Despite integrating reasoning capabilities, the **SFT think** variant underperforms slightly compared to the standard **SFT** model in average benchmark scores.

## Intended Use

This model is suited for research and development in image editing tasks that benefit from interpretable reasoning, such as instructional or multi-step visual modifications.