Easy to Use Request

[Konerusudhir]

Currently, the setup is divided into 2 steps. Step 1: Copute_dire.sh to generate an DIRE image from Input image(Real or Fake) (Distributed computing on N GPUS) Step 2: demo.py to generate probability from DIRE image.

This setup makes it very hard to test and use the model.

I think if we can simplify it to one batched function call to execute both steps will make the model more usable to other researchers or end users.

Example: Create inference.py file which takes a directory as input and generates probabilities in a CSV file on a Single GPU or M2 Mac.

[RichardSunnyMeng]

Hi, have you already reproduced compute_dire.py on your dataset? I didn't see so significant reconstruction error...

[Konerusudhir]

Hi, have you already reproduced compute_dire.py on your dataset? I didn't see so significant reconstruction error...

I tested it only for 4 images so far. I am planning on testing it with CIFAKE dataset in the coming days. https://www.kaggle.com/datasets/birdy654/cifake-real-and-ai-generated-synthetic-images

[shanface33]

Hi, thanks for your comments. May I inquire about the time taken to generate DIRE for each image?

[Konerusudhir]

It took 4.5 Seconds per image on T4 GPU. DIRE approach does not look practical for the production use case if it takes 4.5 seconds per image.

[shanface33]

Thank you so much for your quick response! I really appreciate it!

[happy-Moer]

Have you ever encountered this problem: ModuleNotFoundError: No module named 'mpi4py' how can I slove this problem? thanks in advance

[aditipanda]

I used demo.py on raw (original, not DIRE-computed) fake and real images from my own dataset, and got strange results. The celebahq_sdv2 ckpt classifies everything as fake - even real images from FairFace dataset.

The lsun_adm and lsun_iddpm ckpts classify most real images as real, but also fake images as real. In fact, lsun_iddpm classifies all fake images as real images - 100%!

This compute_DIRE step seems compulsory to be done at our end to see the improvements with DIRE. But I shall follow what's suggested, and get back here.

[aditipanda]

I used demo.py on raw (original, not DIRE-computed) fake and real images from my own dataset, and got strange results. The celebahq_sdv2 ckpt classifies everything as fake - even real images from FairFace dataset.

The lsun_adm and lsun_iddpm ckpts classify most real images as real, but also fake images as real. In fact, lsun_iddpm classifies all fake images as real images - 100%!

This compute_DIRE step seems compulsory to be done at our end to see the improvements with DIRE. But I shall follow what's suggested, and get back here.

So 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]

both R

I downloaded the celebAHQ dire images and ran test.py - got the same results as Table 3 of the paper.

But for computing dire for my own dataset, I am facing this problem #35 - someone pls help!

[XCQ001204]

在 T4 GPU 上，每张图像需要 4.5 秒。如果每张图像需要 4.5 秒，则 DIRE 方法对于生产用例来说似乎不切实际。

May I ask if you are using distributed training or training on a single GPU?anks very much!