How to Upscale Midjourney Images to 4K Without Losing Details

I ruined a gallery submission last year. Not because the art was bad. I had spent three days generating and curating Midjourney outputs for a local group show. The compositions were exactly what I wanted. Then I upscaled them all using Midjourney's built-in upscaler and sent them off to the printer.

When the test prints came back, I almost didn't recognize them. Every piece of fabric texture was gone. The atmospheric haze I loved had been compressed into flat, smooth gradients. The skin on every portrait looked like a cell phone beauty filter had been cranked to maximum. I had to pull out of the show.

That was when I realized Midjourney's built-in upscaler and a proper external upscaler are doing completely different things. If you are making images for Instagram, the built-in tools are fine. If you want to print, sell prints, or display your work at full resolution, you need a different approach.

What Midjourney's upscaler actually does

Midjourney offers a few upscaling modes. Subtle, Creative, and the older V5 variations. They all work on the same principle. The model looks at your image and tries to improve it holistically. It sharpens, smooths, and adjusts the overall composition.

The problem is that “holistic” means it cannot tell the difference between things you want smoothed and things you want kept sharp. A noisy background should be smoothed. The weave pattern on a linen shirt should not be. Midjourney treats them the same.

I noticed this most on images with complex materials. Leather came out looking like brown plastic. Wool textures disappeared completely. Anything with subsurface scattering, like skin or candle wax, got flattened into a single uniform surface.

A dedicated external upscaler fixes this

External upscalers like Real-ESRGAN are trained differently. Instead of trying to improve the whole image at once, they were trained on millions of photo pairs to learn a specific skill. Given a low resolution image, what should the high resolution version look like? They learn to predict where detail should exist and generate it, pixel by pixel.

The difference is most obvious on texture. Here is what I saw when I tested the same 1024x1024 Midjourney output through both methods.

Midjourney upscaler on a portrait of an old man in a wool coat. The face looked fine at first glance, but zoom in. The pores and fine lines that made him look real were gone. The wool coat became a brown blob with no visible weave. It looked like a very good video game render from 2016, not a printable piece of art.

Same image through Real-ESRGAN at 4x. The pores were there. You could see individual wool fibers. The stitching on the coat collar was sharp enough to count the threads. It still looked like the same man and the same coat. Just clearer. More real. At 4096x4096, it was ready for a 13 inch print at 300 DPI.

The workflow I use now

I keep it simple because I have learned that adding steps usually makes things worse, not better. Here is what I do for every piece I plan to print or sell.

First, I download my Midjourney output at the highest resolution available. Usually this is 1024x1024 or 1664x1088 depending on the aspect ratio. Always download as PNG. JPEG compression adds artifacts at this stage that get amplified when you upscale.

Second, I upload to ClarifyPix's AI upscaler and run a 4x upscale. This takes about five to eight seconds per image. Cost is 4 credits on most platforms that use Real-ESRGAN. The result is a 4096x4096 image that prints at 13.6 inches at 300 DPI. Good enough for most gallery shows and art prints.

Third, I check the faces. If the Midjourney generation had people in it, the upscaling step sometimes softens facial details slightly. In that case I run a separate face restoration pass using GFPGAN. This costs another 4 credits. If the image has no faces or the faces were already sharp, I skip this entirely.

That is the whole workflow. Export PNG, upscale 4x, optionally fix faces. Two steps for most images, three if there are people. Total cost is 4 or 8 credits depending on whether faces are involved. Total time is under 15 seconds of processing.

One thing I wish I knew earlier

Never upscale a JPEG if you can avoid it. I did this for months because I did not realize Midjourney defaults to JPEG for some outputs. The compression artifacts are subtle at 1024x1024 but when you blow the image up 4x, they become visible. You start seeing blocky artifacts in smooth gradients and ringing around sharp edges.

Always export as PNG from Discord or whatever Midjourney interface you use. If you only have a JPEG, it is still worth upscaling. The results will be better than not upscaling. But PNG source files give you noticeably cleaner output at 4K. I learned this the hard way after wasting about 20 credits on JPEG upscales that looked mediocre at full resolution.

What about Midjourney V7 and newer versions

I have tested this workflow on outputs from V5, V6, and V7. The improvement from external upscaling is consistent across all versions. V7 generates better base images than V6, which generates better base images than V5. But the fundamental limitation of the built-in upscaler is the same. It smooths textures you want to keep.

I suspect this will be true for future versions too. Midjourney is an image generator first. Their upscaler is a convenience feature, not a professional tool. As long as you treat it that way and use an external upscaler for anything you care about, you will be fine.

If you have been relying on the built-in Midjourney upscaler and wondering why your prints look soft, try running one image through an external 4x upscale and compare them side by side. The difference is not subtle.