# Can LearnDistribution Train and Generate Images Without Defaulting To 256x256?

In spite of anything I try it seems the ConformImages[] function nested in the LearnDistribution[] function resizes images to 256x256.

If I convert an image into its "Data" representation before training I can avoid this problem, but then that conversion causes other problems besides the point, and so I thought I’d ask:

Is there any option that can be set such that when LearnDistribution[] is given an image for training that image is not automatically resized to 256x256 but maintains its original dimensions (e.g., 512x512, 1024x1024, etc.) instead?

Maybe there's a line of code somewhere in the system files that simply needs adjusting? Thanks.

• How do you see that it resizes the image to 256x256? E.g. when running LearnDistribution[{ExampleData[{"TestImage", "Lena"}]}] - the response is a LearnedDistribution object, but I see no "ImageSize" property or anything like that. – C. E. Sep 22 '19 at 1:19
• Feeding the Information[ ] function a LearnedDistribution[ ] gives a "Distribution Information" panel that displays the dimensions of the training examples used (an expanded LearnedDistribution[ ] object might already include this information but I'm unable to check just now). If the LearnDistribution[ ] function's input type is an image the dimension is always 65536 (i.e., 256x256) regardless of that image's initial dimension. – agibilium Sep 22 '19 at 4:40
• Ok, I see. So that size is only used internally. Do you have any reason to believe that this is a bad choice? I’m not sure that I understand what the problem is. – C. E. Sep 22 '19 at 7:30