Show HN: Image Upscaler AI | Hacker News Reader

Show HN: Image Upscaler AI

imageupscalerai.com

96 pointslou_alcala57 comments

If you prefer to run something like this locally while maintaining 100% ownership and control of your files here is my android app I wrote 6 months ago: https://play.google.com/store/apps/details?id=com.luksapps.i...

It uses the best open model on the market to run up scaling (multiple versions of Real-ESRGAN) and runs them in 10s~up to few min per photo on relatively modern mobile phones. This is a free ad supported versions. There is also a paid version with no ads. Both versions can be used while 100% offline.

Nice! however, it says not available for my device (Galaxy Flip 4). Hopefully at some point it will be?

Thank you. This is the kind of information I need to make it better :-)

This is a fast phone with a modern Gpu (Adreno 730), it should run much better than on my devices. Perhaps it's because I haven't released Android 13 version yet? Are you running Android 13 by any chance? If you'd like you can check by clicking on the Settings, then on the very bottom click on "About phone" then on "Software information". There should be "Android version" in there.

I'll make a new release in few days.

I tried it out and was surprised how well it worked on a photograph my friend sent me over WhatsApp the other day (pretty sure WhatsApp downsampled/applied extra compression).

I'm definitely at least considering purchasing the paid version for $5:

https://play.google.com/store/apps/details?id=com.luksapps.i...

Really nice job, sir.

By the way, do you plan to create versions for iOS and/or iPadOS?

Similarly, there's SuperImage [0] on F-Droid, which uses Real-ESRGAN.

[0] https://github.com/Lucchetto/SuperImage

I guess it's fine if it's free or cheap. But why would I pay anywhere from $29-232 per month for what is basically a glorified web UI wrapping open source models? I can get the same or superior functionality for free from FOSS running on my local machine, including mobile devices.

https://www.upscayl.org/

https://superimage.io/

> But why would I pay anywhere from $29-232 per month for what is basically a glorified web UI wrapping open source models?

Why would you? No idea. But there's obviously a huge market for slapping an easy-to-use interface + hosting onto something available for free.

This website is just a wrapper around https://github.com/brycedrennan/imaginAIry which is a wrapper around various neural networks, e.g. https://github.com/xinntao/Real-ESRGAN in this case.

To be fair they're not hiding it. There's a link to imaginAIry in the footer, even on the pricing page.

Are you using the word "just" in a dismissive manner?

my test image is number 201 in the queue, so might as well just run upscayl locally. https://github.com/upscayl/upscayl

I'm increasing the workers, didn't expect this growth

Congrats on the overwhelming success

That's great. You already know this but ignore the people trivializing it. Yes, we hackers could get this running without much effort but most people are not hackers and have absolutely no clue where to start. Even as a hacker there are occasional tools that I don't have on my system at that moment so I'm very thankful to have easy point and click websites people have built around the tools, because it saves a ton of time. Time is incredibly valuable to me now, and I'm happy to pay for things that help me do things faster.

To me, "just a wrapper" makes it sound like it's something completely trivial but I'm an experienced web dev and I wouldn't even know where to start. How do I run this stuff on the server and how do I expose it in a web UI like this?

RealESRGAN is available as a command line tool you can run on a gaming PC. So you'd build a site that does image uploads and a worker process that upscales the images. That hardest parts are off the shelf open source libraries.

Your automatic italian localization is incredibly bad.

Examples: "Stiamo usando un fork dell'immaginazione per alimentare la nostra intelligenza artificiale, e il nostro progetto è sviluppato con Django per il sito web." -> "We are using a fork of imagination for feeding our artificial intelligence, and our project is developed with Django for the website".

"Migliora le tue foto, cartoni animati, immagini e più facilmente e rapidamente utilizzando l'intelligenza artificiale" -> "Improve your pics, cartoons, images and easier and faster using AI"

"Aumento della scala delle immagini" -> "Increase in the scale [alternative reading: in the staircase/ladder] of images"

I haven't read your term of services, but I'm not sure they should be left to the same automatic translation; give them in English, or at least make sure you say that the English version is binding and the automatic translation is provided "for convenience only". Not a lawyer at all, but if this was my website I would write something like this before the translated ToS: "La versione vincolante dei termini di servizio è quella originale in lingua inglese [link to ENG version]; questa versione in italiano è tradotta in maniera automatica e pertanto può contenere inesattezze. La versione italiana non ha alcuna valenza legale, ed è fornita unicamente per comodità."

The Spanish translation has similar errors. In particular the two "y" (and) in the initial sentence make it difficult to read. (I think it's grammatically correct, but it's too difficult to understand.) Also the name of the project imaginAIry is translated to "imaginación" (imagination), but it should have been left unchanged.

The term of service are in English, anyway.

Not the Op, but I have recently been working on applying "state of the art" AI for translations into languages other than English and I was pretty disappointed with rather mediocre results from Chatgpt, normal gpt and Google translate API. As it is, just running fairseq M2M100 locally (even just on a cpu) gave me significantly better results for English->Polish translation. And I was properly amazed by the quality of Microsoft's translation service.

It wasn so much that the output ChatGPT / Google created was wrong, but that it used wrong type of language for the context. For example speaking using very basic/common words no real person would use in a business document. MS translation service was really good in this regard, only making mistakes on genders of people in the translated text (too small context window?).

If you liked M2M100 you might also want to try NLLB 200, which is basically a more recent version (from Fairseq too) that supports more languages and that gives better results in my opinion.

These image upscalers always seem to remove noise in ways that make the result look terrible. You should be adding reasonable texture, not smearing surfaces in vaseline.

This is especially obvious for the Machu Picchu image where the cobble stone walls are turned into different walls.

Original: https://imageupscalerai.com/static/images/test-machupicchu.j...

Upscaled: https://imageupscalerai.com/static/images/result-machupicchu...

Yeah, that upscale result is horrendous. It would never be acceptable in professional situations. It looks like something rendered in a 3D system with a cell shader plugin

Does the result get better if you upscale multiple times, in smaller intervals, so that the single jump extrapolation of detail is not as large going from the original to 8x?

I would wonder if increasing the size of the image keeping the pixelation and then using a pixelation removal network might do better than this. Those, at least, are already trained to add detail where it isn't.

They're using established open source models and code, just under their own web UI. Specifically Real-ESRGAN for the upscaling, and Codeformer for the face enhancements, which is what most people are doing. Though my company is using GFPGAN for the facial enhancements at the moment.

"AI Image Enhancement" websites are all using the same stuff.

You are just jealous of the Machu Picchu super lawn.

But seriously, I think this is much harder to do and probably a second step after the edges are defined.

If you want to provide good upscaling results you need to use a latent scaler but these are compute intensive so the easier way is to just blow it through a gan like here.

What are some examples of latent scaler upsamplers?

There are a lot but the one implemented as LDSR in most stable guis is this one. https://github.com/CompVis/latent-diffusion

upscale wiki is really the place to explore everything image scaling:

https://upscale.wiki/

Not always. One notable example which left me really impressed was letsenhance.io (admittedly I almost started working for them at one point but ultimately took a different route).

The same for video.

Even the "state of the art", Topaz AI's video upscaler, makes everyone look like a porcelain doll.

there’s something like the uncanny valley for upscaled images

some of them make me uncomfortable for some weird reason

This is super cool, and congratulations for launching it. But the pricing is obscene.

Reasonable would be 50ct per image, or at least a lower level tier for the occasional use, something around $3-5/month.

I'm adding a new pricing logic, 10 imgs pack, 50ct/img sound reasonable to me

Awesome. Thanks for listening!

Gigapixel AI works fantastic (https://www.topazlabs.com/gigapixel-ai) and has a perpetual license.

Gigapixel is no longer in active development, they’re focusing on Topaz Photo AI (which rolls all of their AI photo tools into one app, including upscaling) and Video AI. Both are pretty decent considering they’re just recently out of beta, and they do roll out updates with fixes and new features fairly frequently. I’d say they’re both pretty good at what they do but on the photo side you can definitely get some decent results using free tools and models if you have the time to play around.

Topaz Labs makes a stand-alone upscaling app with proprietary models, or you can use something like Cupscale and try various models that are freely available. You usually need to chain multiple models together with the latter option in order to get decent results.

Edit: here are some links to get started…

Cupscale - https://github.com/DrPleaseRespect/cupscale/releases

Models - https://nmkd.de/?esrgan

More models & info - https://upscale.wiki/wiki/Main_Page

I haven't tested it yet, but I suggest that you have a "per image" price for those who use it sporadically.

I could be a potential user, but not for 24/m for the 50 images I can do per month.

If wanting to use ESRGAN I've used this colab [0], set up for video.

[0] https://colab.research.google.com/github/MSFTserver/AI-Colab...

That's a great idea for pricing logic, I'll put it live, I might even be able to create bundles, something like 10-image packs, and you just keep adding more and more if you need.

Yeah, that is also good. X images for Y money.

I came across another open source upscaler https://github.com/upscayl/upscayl

Not an honest comparison, but Pixelmator Pro [0] has a similar "Super resolution" option when resizing images. (Among other great features, of course, being a full-featured photo editor)

http://pixelmator.com/

The upscaled Mickey Mouse image does look "better" to my eyes, but the other 3 examples on the homepage, particularly Machu Pichu, look worse. The original owl on grass looks like a composite image, so it's difficult to discern what to even expect as "better".

The up-scaled ones always look overly smooth to me. Looks unnatural in a lot of cases.

UI looks identical to https://backgroundremoverai.com/ ?

So? They both use bootstrap and they're made by the same people with the same contact information and company name and everything.

Is it the same people though? The Github link is different.

Yes, it's similar because Nadermx and I created a template that we share to launch our apps and sometimes we help each other out.

> Is it the same people though?

Do you not see the "VPS.org LLC" and "Contact us" links on the bottom that I see?

Absolutely amazing. Thank you!

Reply on news.ycombinator.com