Would this work with some training about waifu2x

[nagadomi]

Use --validation_rate 0.17 option.
number of validation images = validation_rate * number_training_images.
validation_rate is 0.05 by default. (int(0.05 * 6) = 0, no validation image)

I think training on 6 images is unreasonable.
If you want to use photo model, you can use models/ukbench or models/photo.

th waifu2x.lua -model_dir models/ukbench -m scale -i input_image.png -o output.png

when you use photo branch,

th waifu2x.lua -model_dir models/photo -m scale -i input_image.png -o output.png

models/ukbench was trained on the ukbench dataset (It contains 10000 small photo). models/photo was trained on 5000 high-resolution raw photograph. Denoising is not supported (work in progress).

[bmaltais]

Cool, thank you!

[bmaltais]

I did a test run from 19 source pictures I use on and off as texture using a scale of 4 and I am amazed at how good the scaling work. I think running it on a small set work well as long as the same source of info was used to create the original that need scaling.

Can I keep adding pictures to the model or is it that once you are done making it you have to start from scratch? For example, if I add a new picture to my set, do I need to run learning against the 20 images or can I only run it against the 1 new to add the data to the model?

[nagadomi]

I think self-similarity based approach(e.g. SelfExSR) is better for your use case.
And single image super-resolution has its limit.

[bmaltais]

Interesting. I will have to get a matlab trial to give it a try. Too bad there is no free matlab starter version for testing purpose. I don't feel like buying the software just to play around and figure out if it work.

Based on how it is trained and what it does this should produce stellar resolution increase for what I intend to do... but perhaps not. Maybe your neural solution work best. Only testing would validate this.

[bmaltais]

Also, how should I set the following values since I use hi-res input images:

-max_size
-crop_size

[nagadomi]

-max_size and -crop_size are not need to change.
If your dataset is very small (less than 100), I recommend following option,

-random_half_rate 0.5 -random_color_noise_rate 0.5 -patches 1024 -epoch 200

But maybe it will overfit.