search

xaCheng1996 / DiFF

Diffusion-generated Facial Forgery Dataset
32 stars 0 forks source link

What's wrong within LoRA datasets? #4

Closed wolo-wolo closed 16 hours ago

wolo-wolo commented 3 months ago

After unzip LoRA and get: /test/I2I/__MACOSX/LoRA /test/I2I/LoRA

However, the image seems to have a format error. Could you check and provided correct version? Thks

xaCheng1996 commented 3 months ago

Hi, You can just delete the __MACOSX. This folder is an auto-generated directory that is used only in MAC OS.

wolo-wolo commented 3 months ago

Hi, You can just delete the __MACOSX. This folder is an auto-generated directory that is used only in MAC OS.

Thanks for your quick reply. I found images that start with the '._,' prefix, e.g., /test/I2I/LoRA/id_0224/._lora_001_000.png, are less than 4kB in size, which cannot be read and visualized. I wonder if I should just delete those files.

Another question: Table 7 in the paper provides four detection methods trained on FF++ and reports AUC generalized to T2I I2I FS FE in test subdatasets. To benchmark this table in my current work, I need to strictly follow its settings. Can you tell me: 1) When using FF++ for training, do you follow its official 820 videos as a training set? How many frames are extracted from each video, and how are faces extracted? 2) As for T2I I2I FS FE test data, it seems that the downloaded test set only has forged images. Are there no real images in the test subset? Or is it convenient for you to provide them?

wolo-wolo commented 3 months ago

@xaCheng1996 How do we reproduce the results in Table 7, test in DiFF(w/o crop face?) w.r.t. train on FF++(compression? frame sample ratio? face extraction)?