Group Leader
- Joined
- Jul 29, 2020
- Messages
- 181
I'll go with 1.6-1.8 usually, but it depends on the original filesize. Mostly to avoid making the images themselves too large.Can you go into why you go with magnify 1.6-1.8? Waifu2x as the name implies does doubling like most NCNN upscalers. The second step after that is a downscale, but what algorithm do you recommend? Bicubic sharp? Spline36?
Larger file sizes start becoming redundant at a certain point, but a good rule of thumb is about 1300x1800 minimum.
Larger images are also a lot more taxing to edit in PS if you have lower specs.
Since this was an example image, I didn't go too heavily into upscaling just so I could present a clearer picture of what to do.
For the downscaling section, I've found it unecessary to do it so long as I upscale to a good margin. Extra steps for little benefit, but I'm not telling anyone not to, just that I don't think it's wholly necessary unless you're trying to salvage something that's horribly degraded.
Beyond that, I'd say I only have experience with using Supersampling, but the results weren't amazing by any means.
I use waifu2x-caffe since I've a good enough graphics card to support it, never had to bother with waifu2x or vulkan. To specify for others, -caffe's framework is exclusive to nvidia graphics cards, which have to support CUDA. (You CAN use your CPU, but who would do that?). As for Vulkan and Waifu2x, either or, I suppose.