Training Stable Diffusion 3.5 Medium with DiffusionOPD on OCR Tasks
I ran the experiment of post-training SD3.5 Medium with DiffusionOPD on OCR tasks, as documented in the docs. I post some results and figures here for my reference.
Table of Contents
1. Prepare a SD3.5 Medium Teacher Model
The first step is to prepare a teacher. Here, I follow the documentation to obtain such teacher by training SD3.5 Medium on OCR tasks with FlowGRPO. The final result (reward, loss)
Figure 1: Reward curve. Reward is generated by Qwen-2.5-VL-3B-Instruct.
From the image, we reach \(0.9113\) at step \(160\), \(0.94\) at step \(380\).
2. Run DiffusionOPD
The final result:
Figure 2: Reward curve. Reward is algo generated by Qwen-2.5-VL-3B-Instruct
From the image, we reach \(0.94\) at very early stage, step \(40\), indicating the potential of DiffusionOPD.