Hacker News Reader

Comment by thorum | Hacker News Reader

DALLE images are still only 1024 px wide. Which has its uses, but I don’t think the stock photo industry is in real danger until someone figures out a better AI superresolution system that can produce larger and more detailed images.

View on HN

You can obtain any size by using the source image with the masking feature. Take the original and shift it then mask out part of the scene and re-run. Sort of like a patchwork quilt, it will build variations of the masked areas with each generation.

Once the API is released, this will be easier to do in a programmatic fashion.

Note: Depending on how many times you do this... I could see there being a continuity problem with the extremes of the image (eg: the far left has no knowledge of the far right). An alternative could be to scale the image down and mask the borders then later scale it back up to the desired resolution.

This scale and mask strategy also works well for images where part of the scene has been clipped that you want to include (EG: Part of a character's body outside the original image dimensions). Scale the image down, then mask the border region, and provide that to the generation step.

Another commenter mentioned Topaz AI upscaling, and Pixelmator has the "ML Super Resolution" feature; both work remarkably well IMO. There are a number of drop-in and system default resolution enhancement processes that work in a pinch, but the quality is lacking compared to the commercial solutions. There are still some areas where DALL-E 2 is lacking in realism, but anyone handy with photo editing tools could amend those shortcomings fairly quickly.

On-demand stock photo generation probably is the next step, particularly when combined with other free media services (Unsplash immediately comes to mind). Simply choose a "look" or base image, add contextual details, and out pops a 1 of 1 stock photo at a fraction of the cost of standard licensing. It'll be very exciting seeing what new products/services will make use of the DALL-E API, how and where they integrate with other APIs, use cases, value adds like upscaling and formatting, etc.

I've been using this app to upscale the images to 4000x4000, and it works amazingly well (there is also a version for Android):

https://apps.apple.com/us/app/waifu2x/id1286485858

I paid extra to get the higher quality model using the in-app purchase option. It crushes the phone's battery life, but runs in only ~10 seconds on an iPhone 13 Pro for a single 1000x1000 input image.

I mean, waifu2x and similar waifuxx libraries are free and open-source, there's really no reason to pay for it if you're working on a desktop.

what's a desktop? /s

Things have moved on a considerable amount since waifu2x

Try https://github.com/n00mkrad/cupscale

Thanks, that’s actually the “better” model that I referenced. You can buy it with an in-app purchase using the waifu2x app.

Buy what?

I think he refers to the premium version of Waifu2x.

I've recently updated waifu2x and I've seen it now supports lots of algorithms for different use cases and contexts and it also supports other tasks like frame interpolation. So could you briefly explain in what is cupscale better than it?

Cupscale supports multiple models tuned for different types of content (including Anime like waifu2x): https://upscale.wiki/wiki/Model_Database

As I said Waifu2x too supports many different models aimed at different content, so what's the big improvement with Cupscale?

Well in that case I might be wrong.

Considering waifu2x is the name of an algorithm I assumed it was just that algorithm. There's also no mention of other models on the demo page or the Github page as far as I can see.

My bad! Yes Waifu2x is just a single algorithm, you are right.

The confusion originates from the fact that I was using a GUI project for Waifu2x called "Waifu2x Extension GUI" (https://github.com/AaronFeng753/Waifu2x-Extension-GUI) which other than Waifu2x also supports other algorithms like Real-ESRGAN, Real-CUGAN, SRMD, RealSR, Anime4K, RIFE, IFRNet, CAIN, DAIN, and ACNet.

So as you said Cupscale is surely more advanced than Waifu2x (the single algorithm), but do you think it's also better than Waifu2x Extension GUI?

Yes, but I usually find myself playing with this stuff when I have some free time and relaxing outside or on the couch, and it’s nice to be able to do it all on the phone.

DallE2 + Topaz Gigapixel AI works amazingly well.

Wow

https://www.topazlabs.com/gigapixel-ai

No kidding.

Photoshop also has super resolution

Reply on news.ycombinator.com