can this thing generate bigger than 256 pixel result?

[FurkanGozukara]

I just tested super resolution yes it works but can it generate bigger than 256 or always 256?

how to generate bigger than 256?

and you call 256px super resolution? :D

also here my test results i think you really should write some proper instructions this is a shame

input image to test

and here the output of super_resolution.py

so it generated this simply

[FurkanGozukara]

@NielsRogge @hamadichihaoui

please some info regarding these

[Dazzastrous]

Hmm the results look good but 256px interested to see if this cab be improved

[hamadichihaoui]

@FurkanGozukara Thanks for your interest in our work! Our method can work with any resolution as long as i) you have enough memory because inference at high resolution needs more memory ii) the pre-trained diffusion model is trained using the same type of data and (approximately) the same resolution so there will be no domain gap. Now, why the result that you get for the example you should is not good? The answer is simply the domain gap. The pre-trained model that we use was trained only curated and centered faces (using FFHQ dataset if I am not wrong and did not see during training a person with a part of the body like the one you use in this example).

[FurkanGozukara]

@hamadichihaoui so we have to train ourselves to make it useful? Nothing good pretrained by you right? Like SUPIR

For example SUPIR is able to upscale literally anything with the very best results right now. Even better than commercial apps like Topaz AI

[Katehuuh]

@hamadichihaoui so we have to train ourselves to make it useful? Nothing good pretrained by you right? Like SUPIR

Centered faces only as shown in examples.:

i think you really should write some proper instructions this is a shame

More likely, combine in one blind_deblurring, blind_non_uniform_deblurring, inpainting, super_resolution with arg for path instead of example path already processed., (and maybe ensure convert img to 24bit to handle 32bit error.)

---

@FurkanGozukara ... you have enough memory because inference at high resolution needs more memory

high resolution not possible to modify as in config: https://github.com/hamadichihaoui/BIRD/blob/03dff34b46703eba678bb4d30e06927220b6b806/data/celeba_hq.yml#L4

[FurkanGozukara]

i see looks like another repo i spent time for nothing :D

meanwhile these are from SUPIR just upscaled to use in my newest tutorial

it is a chinese model by the way public open source

[hamadichihaoui]

@FurkanGozukara I think you are missing something here. SUPIR was trained with more than 20 millions images with text annotations. It is specialized in superresolution. The model that we use here for faces is trained using 60K images (without text annotations) and we can handle different restoration tasks (superresolution, inpainting..). Here we just present an idea and a proof of concept. At the end, the comparison with SUPIR is unfair. I know that you don't care about all of this but just wanted to point it out.

[FurkanGozukara]

@hamadichihaoui ty for replies