HN Explorer

Ask HN: What is the state of the art in AI photo enhancement?

5 points by brap ↗ HN
I have a few face photos that I really like, but they’re a tiny bit blurry. The lighting is great but the issue is that they were taken a little far away, so some digital zoom is needed. They’re good enough, as you can see every important detail, but they’re not “sharp”.

I looked into several consumer apps, including paid ones. Some options that exist:

* Topaz AI

* Photoleap

* Facetune

But I couldn’t really compare nor do I know if better tech exists these days. I don’t mind running my own model.

What’s the state of the art?

6 comments

[ 2.8 ms ] story [ 20.9 ms ] thread
tlack
GFPGAN is decent but kinda heavy handed: https://github.com/TencentARC/GFPGAN

Worth a shot!

brap
Thanks! Just tried it, but it made my skin very smooth, like it was obviously filtered. And it gave me perfect teeth. Kind of uncanny.
tlack
how I wish it had a strength setting..

you can try blending the before and after with PIL but it will probably look even more disturbing

    strength = 0.5
    horrific_image = Image.blend(image_with_normal_skin_and_teeth, 
    kinda_awful_gfpgan_image, strength)
let me know if you find something better
shipwright
Fortunately, that is indeed a thing :) several post-processing tools have some sort of gfpgan_strength (searching as-is should yield results on a few repos) param being passed, also it's been a while since diffusion-based restoration methods are equipped to diminish stochasticity enough to produce great results, worth the time as well! Also noteworthy: https://openmodeldb.info/?t=faces+restoration+texture-genera...
tlack
In Auto1111, that just uses Image.blend. :)

https://github.com/AUTOMATIC1111/stable-diffusion-webui/blob...

shipwright
Hah, so high tech- Turns out the said repos used to do the same, then one of them later added CodeFormer under the hood, conflated the strength with affinity and didn't bother renaming (calling codeformer --gfpgan_strength looks very unhinged to me but oh well haha). It's definitely doable though! What I had in mind as a GAN equivalent (used to apply in WSI context while back) is akin to Rayleigh EigenDirections in terms of latent space navigation https://www.ecva.net/papers/eccv_2022/papers_ECCV/html/7277_... There's some code out there, unsure if it's streamlined for this use case but can lend a hand if you decide to give it a shot :)