Upscale model specifically for AI (e.g. MidJourney)

AI art generation is becoming increasing common.

AI programs like MidJourney create fairly small images (often no larger than 1024x1024).

Stable Diffusion allows you to upscale… but only if you have a VERY powerful CPU (or pay for cloud based diffusion e.g. RunDiffusion etc).

MidJourney now offers 4x upscaling… but at a high cost (5-7min of processing time = 45 cents per upscale… plus 7 minutes waiting!)

Photo AI does the same job much quicker, with some control… but the upscales can be overly bright, with “unusual” skin and eye issues.

I’ve found that the Standard model with 15% face recover works best. High Fidelity model does well if there are no faces, but if there are, skin and eyes don’t fare well. 25% face recover helps, but not much. More than 25% face recovery leads to poor focus.

Since this market is growing, could there be the possibility of creating a model better suited to AI… surely this would be popular for all the Dall-E3, Adobe Firefly, SD, and MidJourney users…

I’ve included an AI photo from MidJourney, with Midjourney’s own 4x upscale, and some Topaz upscales for comparison


  • List item



For AI generated image, I think it is better upscale with Stable Diffusion, because it can generate new detail. It is completely free and doesn’t require a powerful CPU.

You may try the image below and see whether TPAI can generate the same amount of detail from the image.

Thanks for this! It’s very good advice!

I use Stable Diffusion (using an online cloud version), but my CPU isn’t large enough to install it locally… to use A1111 AND to get above 1024x1024 you need 12GB vram (and I only have 4)… beyond my budget for this single purpose.

It makes me sad!