search

phillipi / pix2pix

Image-to-image translation with conditional adversarial nets
https://phillipi.github.io/pix2pix/
Other
10.14k stars 1.71k forks source link

Using trained models outside of test.lua #71

Open Multiboxer opened 7 years ago

Multiboxer commented 7 years ago

I'm attempting to create a model that makes small adjustments to pictures however test.lua produces extremely distorted, cropped, or pixelated images. Is there any way to use a process similar to test.lua that outputs at an identical resolution to the input image?

Quasimondo commented 7 years ago

The model only processes an image in the fineSize dimensions. So if you pass in a bigger image it will get scaled down and potentially cropped.

To my knowledge there is no perfect solution. The optimum would be to train your model with a fineSize that is big enough to fit your original image, but except if you have a lot of GPU memory that is not always feasible.

You can also increase the fineSize on a pre-trained model, but that will often give you different results and not capture the bigger image structures so well (it usually works ok on textures)

A third option is process your image in fineSize-sized tiles, but that will also give you different results if you trained your model only on scaled down full-size images and often adjacent tiles will differ in color based on their individual content.

PolarisD commented 6 years ago

The model only processes an image in the fineSize dimensions. So if you pass in a bigger image it will get scaled down and potentially cropped.

To my knowledge there is no perfect solution. The optimum would be to train your model with a fineSize that is big enough to fit your original image, but except if you have a lot of GPU memory that is not always feasible.

You can also increase the fineSize on a pre-trained model, but that will often give you different results and not capture the bigger image structures so well (it usually works ok on textures)

A third option is process your image in fineSize-sized tiles, but that will also give you different results if you trained your model only on scaled down full-size images and often adjacent tiles will differ in color based on their individual content.

How can I modify the pre-trained model, by editing which file? For example I want to use the model on human face dataset?

junyanz commented 6 years ago

The 256x256 model can run on 512x512 images. See this post for more details.