On stage 1 training #67
Description
Hi, nice job there, I was trying to replicate some of the results of your paper, and I have just a couple of questions:
What data was it used to train stage 1? I see that the Qwen2vl_dataset is somehow always returning an input image and a generated image. In which cases did you find https://github.com/PKU-YuanGroup/UniWorld-V1/blob/main/train_denoiser.py#L987 being empty?
If stage 1 is not using siglip features, is it fully trained for generative tasks, i.e. prompt with no image with the task being it to get the output image?
I find this bit a bit confusing and I found no reference in the paper to the data being used to train each stage. Could you please provide with some additional details to assist on reproducibility?
Thanks!