DiffUTE icon indicating copy to clipboard operation
DiffUTE copied to clipboard

This repository is the code of our paper "DiffUTE: Universal Text Editing Diffusion Model" (NeurIPS'2023).

Results 4 DiffUTE issues
Sort by recently updated
recently updated
newest added

您好,我在渐进式训练vae的时候出现了一些问题, S=64和S=128的阶段,将和训练集一样处理的图片经过vae encode再decode之后,图片上的文字可以辨认, 但是当S增加到256时,解码出来的图片上面文字非常不清晰,不能辨认,可是我的模型已经收敛了…… 想知道vae训练的时候参数是怎么设置的呢?

Thank you for the meaningful work here. Could you please provide sample files for `doc.csv` and `doc_select.csv` with 1 line of placeholder data in each column? This will greatly assist...

Could you please tell me if this work employs classifier-free guidance (CFG)? If so, what is the negative prompt, and what is the CFG scale value?

作者你好,拜读了你的文章,我对训练推理过程有几个问题: 1. 训练是用 concat[x, x_m, z_t] 来预测 z_t 的 noise,这样理解对吗 ? 2. 推理的时候输入是 concat[x, x_m, z_T], 那么这一过程是对谁去噪呢 ? 是对 z_T 还是对 concat[x, x_m, z_T] 3. 一般 SD 是直接采样一个噪声作为初始输入,我推理的时候直接把 z_T 换成一个随机噪声,还能达到原来的效果吗...