Tip : Upscaling....

#21
by BobbyLH44 - opened

So this might be obvious...but I just realized I can use the various neural upscale or super-resolution models to take these Dalle Mini outputs from 256x256 to 512x512 with virtually no loss of quality at all. In fact they sometimes look BETTER after being upscaled to 512x512, depending on the image and the model used. You can also upscale to 1024x1024 but there does seem to be some minor loss of quality most of the time. (but still impressive considering thats a 4x upscale)

I’ve done that already. :)

In my experience upscaling does not work. I think it depends on the image you are trying to upscale. For example if you have a flat plane or surface, like a plain painted wall, then upscaling works fine. But if you have a detailed structure, like tree roots for example, then the upscaler cannot 'fill in the gaps'.

What we need I think is a context aware upscaler, trained on the same model that produced the original image, using the same prompt, that can 'zoom in' on a structure, add relevant context aware greeble detail and then stich the outputs together into a larger output image. Context-free upscaling simply can't add appropriate details that don't exist in the original output image. That is to say that the original image contains a finite amount of information and the upscaler doesn't know enough about the original image (or prompt) to add additional context-aware information.

Maybe the tool I am using IS context-aware, or maybe I am just not picky enough to see the quality loss. But at 2x upscaling I am perceiving virtually no loss in quality. Mileage may vary depending on the image and the model used, etc.

mbug90, comment ? Comment passer de 256x256 à 512x512 ? J'ai cherché dans le code de la page html et dans les scripts, mais je n'ai pas trouvé ou modifier les valeurs !
... how? How do I change from 256x256 to 512x512? I searched the code of the html page and the scripts, but I couldn't find where to change the values!

Take a generated image, then put it through this rapid ad fairly subtle/nondestructive upscaler: https://deepai.org/machine-learning-model/waifu2x
Followed by this one, by far the most uncannily efficient free upscaler I've found so far*: https://zyro.com/tools/image-upscaler

  • On the downside, it can be buggy with some images, and occasionally destructive of grain, minor detailing, and certain subtle textures. But imho for most generative content, especially 256x256, it's worth it.
    For me, it's been, as they say, a "game changer" in terms of producing decent quality imagery, thereafter usable for making digital artworks and collages; specially with further edits using more hands on tools. Btw, as far as free art/photo editing apps, I would easily recommend PhotoScape X. With these tools + Dall-E Mini, as well as the BigSleep GAN notebook on Colab, and another guided diffusion notebook I've already generated illustrations for two autonomously designed and edited books of poetry. For desktop publishing, and also image collaging/poster-design, etc, I'd recommend Affinity publisher over the pricier Indesign.

Wow, that Zyro upscaler is VERY good. Maybe the best 4x upscale I have seen, its damn near flawless if the original 256x256 image is fairly crisp to begin with. Thanks for the link.

Agreed about editing the Dalle Mini results. I just started playing with telling it to produce images with "green screen" backgrounds so that I can easily edit the file to make the background transparent and stuff like that. Although a lot of the newer photo editing apps can strip backgrounds automatically anyway. The mind boggles at all the possibilities, just not enough time in the day to try everything ;) I am starting projects I could not have dreamed of previously, just the sheer scale and volume of content that can be generated, then combed through to grab the best results, further refine and edit in photo editing tools, etc.

Thank you very much for your tips @AlekseyCalvin Zyro works remarkably well! I tried Googling for some of your montage images but I could find nothing. If you are able to share I'd love to see what you are doing. Thanks again! Cheers

Sign up or log in to comment