base_model:
- mair-lab/thinking-sft-simple
EARL - RL Fine-tuned (S + C) thinking (8B)
Model Name: mair-lab/earl-thinking-sft-simple.rl-simple-n-complex
Model Size: 8B parameters
Base Checkpoint: mair-lab/sft-think-simple
Training Method: Supervised Fine-Tuning (SFT think (S)) → Reinforcement Learning (RL) on Simple + Complex Edits
Datasets: Simple Edit (S), Complex Edit (C)
This model is part of the EARL benchmark study:
📄 EARL: The Promise of RL for Autoregressive Image Editing
Model Summary
This model starts from the reasoning-based SFT checkpoint (sft-think-simple) and is optimized using reinforcement learning across both simple and complex edit instructions. While it incorporates chain-of-thought supervision, it still trails the non-reasoning RL model in overall benchmark performance.
➡️ Inference instructions: GitHub Repo
Benchmark Results
| Model Description | OmniEdit | EmuEdit | AURORA | MB | VisMin | I2EBench | AVG |
|---|---|---|---|---|---|---|---|
| SFT think (S) | 4.34 | 3.76 | 2.88 | 3.36 | 3.46 | 3.21 | 3.50 |
| EARL SFT think (S) → RL (S+C) | 4.65 | 3.78 | 3.23 | 3.67 | 3.39 | 3.36 | 3.68 |
📉 Note: The RL version improves modestly over the SFT think (S) baseline but does not match the performance of the non-reasoning SFT (S) RL model, indicating current limitations of reasoning-guided supervision in image editing.
Use Cases
- Research on visual reasoning in RL
- Multi-instruction and compositional image editing
- Comparative analysis between reasoning and non-reasoning RL approaches