Can any Ai based tools could make a clean looking sign/logo from a photo?

gnoop

Like some random phone taken  picture of  a sign,  low res and blurry  to something ready to turn in vectors ?    I tried some Ai  upscalers , Ai image generators , chat gpt .  

Upscalers  are useless until you have  already nice photo .     Generators  do something fancy  and never  what you actually need  .      I know Photoshop has some font recognizing feature but it's not very reliable  and  basically pain you a.. either.  

Tiles

Well, imho AI upscaling is the way to go here in general. And turning it into a vector grahphics afterwards. Inkscape is capable of that step. And then you can correct it to your needs.

What will not work is to turn crap into gold. The better the input image the better the output. And when you really have just a washy input image then you will go nowhere with it. In this case it might be easier to use this blurry image as a reference image in your vector software, and redo it manually.

You could play around with different upscalers. There are several AI Upscaling models around. And so you can achieve different results. There are a few models available for Automatic1111 or ComfyUI for offline needs. A popular one and one of the oldest is the realesrgan for example. Works also as standalone. Or Upyscayl, a offline tool just for upscaling. And then there is stuff like Topaz Labs Gigapixel. It costs money. 

https://github.com/xinntao/Real-ESRGAN
https://civitai.com/models/116225/4x-ultrasharp
https://upscayl.org/
https://www.topazlabs.com/gigapixel

There are also upscaling workflows for ComfyUI around which does not go across a upscaling AI, but across a (usually second) ksampler. But it will alter the input.

poopipe

you don't need AI - there's a plethora of traditional computer vision techniques that will do the job 

the same caveats apply though - shit data in, shit data out

gnoop

Thanks  Tiles   for interesting links to look in .  Yeah.   with a nice photo I wouldn't need anything  for sure.     Just hope  for something  that could quickly do that nice photo from a total crap.   You tube video  frame and such . 

Tiles

Maybe scale it even more down to get a better result. When you feed in a washy image then it will try to upscale this washy look ...

there's a plethora of traditional computer vision techniques that will do the job 
Not this job, unfortunately. Traditional methods will not fill in fine details like hairs or surface structure. Traditional methods will upscale what exists, not add in more details. Just have a look at the results of Topazlabs for example.

poopipe

no but that's not what he asked for in the first post. 

Standard edge detection and hull generation techniques do a decent job of finding shapes. Unlike an ML they can't  magically invent things that aren't there and can't infer meaning from the shapes but in terms of vectorising something like a photo of a sign they're generally fairly capable given reasonable quality source data. 

gnoop

Wasted half a day  trying to force chat GPT version  o1 to do it .  It's the one with "advanced reasoning" and slow as hell.      At least it can recognize what's written in such washed signs from a track race video screens .   I sometimes can't .   No help for recreating it .  just fancy pictures.     Even when it seems to understand  the task pretty well . 

Tiles

poopipe said:

no but that's not what he asked for in the first post. 

Standard edge detection and hull generation techniques do a decent job of finding shapes. Unlike an ML they can't  magically invent things that aren't there and can't infer meaning from the shapes but in terms of vectorising something like a photo of a sign they're generally fairly capable given reasonable quality source data. 

Yeah sure, but the quality is the culprit here. Too much missing bits from what i understand.
Would be of course interesting to see the image that we talk about here.

@gnoop, i would at this point have a look how far you can come with the traditional methods instead. It might indeed be the easier approach here. Stay pragmatic, and choose whatever the job does best.