After actual testing, the images generated by stable diffusion cannot be recognized.

[TheBloodthirster]

give the samples from test result , which predict score are all 0.0000

[TheBloodthirster]

give one more all predict 0.0000 test result

[lukovnikov]

Can you please be more specific about what exactly you did? What model were these generated with, how did you save these images?
Could you try saving them as .jpg and try again?

[aditipanda]

@lukovnikov

I ran compute_dire for CelebAHQ images - both Recons and DIRE, and strangely enough, the celebahqsdv2 ckpt classifies all these real images as fake!

lsun-adm, lsun_iddpm and lsun_stylegan correctly classify real images but not the other three ckpts. Very strange as the celebahq checkpoint is trained on celebahq real dataset I presume?

[aditipanda]

@lukovnikov

I ran compute_dire for CelebAHQ images - both Recons and DIRE, and strangely enough, the celebahqsdv2 ckpt classifies all these real images as fake!

lsun-adm, lsun_iddpm and lsun_stylegan correctly classify real images but not the other three ckpts. Very strange as the celebahq checkpoint is trained on celebahq real dataset I presume?

You have to run with dire versions and not original/raw images. See guided-diffusion/compute_dire.sh for more. I could get almost same results as Table 3 of the paper.

[IIyaJ]

@lukovnikov
我为 CelebAHQ 图像运行了 compute_dire - 包括 Recons 和 DIRE，奇怪的是，celebahqsdv2 ckpt 将所有这些真实图像归类为假图像！
lsun-adm、lsun_iddpm 和 lsun_stylegan 可以正确分类真实图像，但不能正确分类其他三个 CKPT。非常奇怪，因为 celebahq 检查点是在 celebahq 真实数据集上训练的，我猜？

您必须使用可怕的版本而不是原始/原始图像运行。有关更多信息，请参阅 guided-diffusion/compute_dire.sh。我可以得到与论文表 3 几乎相同的结果。

I am also repeating this content. I don't understand what you mean by using the dire image training model? Whether to use dire image in training and testing model or use dire in calculating reconstruction error, I always think it is the latter.

[IIyaJ]

@lukovnikov
我为 CelebAHQ 图像运行了 compute_dire - 包括 Recons 和 DIRE，奇怪的是，celebahqsdv2 ckpt 将所有这些真实图像归类为假图像！
lsun-adm、lsun_iddpm 和 lsun_stylegan 可以正确分类真实图像，但不能正确分类其他三个 CKPT。非常奇怪，因为 celebahq 检查点是在 celebahq 真实数据集上训练的，我猜？

您必须使用可怕的版本而不是原始/原始图像运行。有关更多信息，请参阅 guided-diffusion/compute_dire.sh。我可以得到与论文表 3 几乎相同的结果。

In addition, I would like to ask whether the program starts to run after pre training on the real image or before. I understand that if it only needs to do reasoning work and calculate the error after pre training, it would be good. So the program includes the pre training process, so it only needs to replace its own data set with its own data set. I look forward to your reply!!

[IIyaJ]

After reading your content, is there a problem with setting the dataset like this? The result is 1 in most cases

[IIyaJ]

How to understand the three sets of training test and verification sets in the data set?