Why no resolution in megapixels?

[light-and-ray]

Is there any technical reason why resolution aspect ratio for image edit models is fixed for entire dataset? Why not to scale input image (control) and result image both to desired total number of pixels, ignoring aspect ratio. Is it because nobody thought about it, or it is impossible due to how training algorithm works

Upd. Not only about image edit models btw. Image generation also could have different result aspect ratios

[kohya-ss]

The aspect ratio of the training image is always maintained as the aspect ratio of the original image. Images are resized to maintain their aspect ratio so that their total number of pixels matches the number of pixels in the selected resolution.

[light-and-ray]

Ah, I see. So resolution option [w, h] is technically an option for total pixels = w * h? I'm just preparing my dataset, I never worked with this project yet (and with any lora training project in general)

[kohya-ss]

Yes, that's right. I apologize for the lack of documentation.
You can check the actual training videos or images by using the --debug_mode video or --debug_mode image options of the latent caching script, so please use this to verify.