Custom models - predictions don't seem to be influenced by camera input

[newt-beach]

Hello :)

I'm encountering an issue running my own models generated through pix2pix-tensorflow. Check out the following video:

Untitled.mp4

In the video I'm switching between two pre-trained TF models: The model you provided from Google Art, and generated from pictures of animals. Notice that when I switch to the animal model, the prediction window looks like a fast slideshow of existing animal pictures.. it doesn't seem to be affected much by the input from the camera. When I switch to the GArt model, it looks as expected.

I have tried creating models from animals, flowers, and even other artworks. Image sets have ranged in size from 100 to 8,000 pictures. I have tried different numbers of epochs (ranging from 10 to 500), as well as different batch sizes. When I put any of these through webcam-pix2pix, I get the same "slideshow" effect. I just can't seem to make the predictions look like the input from the webcam. Tweaking the capture/processing parameters on the options panel doesn't do anything, either.

When I train the models, I follow the instructions listed here and on the pix2pix page, with one caveat: I am not doing the "export" step because doing so causes webcam-pix2pix to throw an error. Instead, I am using the models as they are exported from "train". This is my first suspect for why my output is looking weird, but I'm not certain. The error I get when running an exported model is this:
InvalidArgumentError (see above for traceback): You must feed a value for placeholder tensor 'input' with dtype float and shape [256,256,3]

My commands for training/exporting models vary but are generally of the form:
python pix2pix.py --mode train --input_dir ../meta-mirror/image-data/animals_canny --output_dir ../meta-mirror/models/animals-temp --which_direction BtoA
python pix2pix.py --mode export --input_dir ../meta-mirror/image-data/animals_canny --output_dir ../meta-mirror/models/animals --checkpoint ../meta-mirror/models/animals-temp

I have tested the models as well and they all look as expected:

Would love some advice on what I could be doing differently. I am using Tensorflow v.1.13.1 for both training and prediction. Thank you.