Much of the discussion about AI in the 3D community has so far focused on fears about computers eventually replacing artist. But the tone at an event I attended in Berlin last week struck a completely different note by asking how artists could harness AI to make their work faster and better.

I found myself agreeing with SOOII‘s Marc Gruber-Laux when he posited that most of us care much more about the images we produce than about how we produce them. If the output is exactly how we intended, then it doesn’t matter if it was made in 3D, in Photoshop or with watercolors.

In that respect, generative AI isn’t quite there yet. Not because its output is qualitatively inadequate–on the contrary, Midjourney and Stable Diffusion are often much more skilled than we could ever hope to become–but because it is so hard to control.

This limited control is a problem that will eventually be solved. In the meantime, generative AI can do a lot to help 3D artists speed up their workflow and improve the quality of their final images.

That’s where tools like Magnific AI come in. If you are familiar with Topaz Labs’ Gigapixel, this does something similar in a web editor. The main difference is in the amount of additional detail Magnific creates when enlarging an image. While not designed specifically for 3D artists, the tool can be harnessed to add a final layer of details to a finished 3D render, ideally before the last post-production step. As Commonpoint‘s Artur and Bartosz put it in their presentation in Berlin, it does in one click that last 10% of the image which normally takes 50% of the work.

Below are a few before-after examples using some of my older renders. As you can see, the effect ranges from subtle to quite dramatic. Magnific can also screw up quite badly, so if you are going to use it in production, it will likely be only for specific parts of your image, typically those that show organic details or distressed materials. Some of the difference may only be visible on a large screen at 100% resolution.

The example below (ignore the difference in white balance, which is my mistake) shows how Magnific excels at vegetation and rugged buildings. While the two images aren’t radically different, the very fine details of leaves and branches is much better on the after image. It would takes days to get this extra level of polish in 3D or with overpainting in Photoshop.

In this image of the Goldstein House in LA, Magnific does a great job of adding extra vivid details in the plants and trees but it can’t quite figure out the interior. In general, it doesn’t seem to process reflections very well and there is a good chance it will screw up anything that is behind glass. You’ll want to mask out those areas in your project.

One area where Magnific can excel (and sometimes screw up spectacularly) is with people. Note how it makes this 3d character more convincing while preserving the lighting. The vegetation in the sand, the trees, and the ground around the pillar are also much improved.

This example shows the best and the worst of Magnific. On the one hand, the vegetation, rocks and water are much more natural, on the other, the tool mangles the fine geometry of the music kiosk and even completely transforms some of its features. All this would have to be masked out in a final composite.

Magnific can do wonders in interiors, especially in grungier ones, potentially saving days of modeling and texturing work. Check out the details of the leather sofa, where extra wrinkles appear, giving it a much more realistic look. The fabrics are generally much better, as are some of the pots and vases. There are some obvious glitches too: one of the candles on the table becomes a weird pint glass, and the wood frame of the pendant light turns into some kind of brass construction.

Generally speaking, Magnific won’t do much good on very clean scenes. In this example, it gives the towel on the Eames chair a much more naturalistic look, but the rest of the image doesn’t see much of an improvement.

Food is another area where Magnific can come up with interesting results. In this case, the madeleines and the bread appear more realistic while the glass and cup are largely left unchanged. Note that the tool sometimes struggles to decide what is in focus and what isn’t adding fine details in areas that should be blurry.

Another still life. The roses are much improved, but everything else is arguable worse.

Magnific does a good job in urban scenes. This one is quite detailed to start with, but it adds another layer of grime and realism that works quite well in this case, though I have some issues with the road. The bridge construction on the left is much more realistic, except for one area of the pillar in the foreground that comes out strangely soft. In general, I’ve noticed that the tool can sometimes remove or soften details in areas that are slightly out of focus, rendering them as flat surfaces.

The tool’s landscape mode does wonders with Gaea scenes, especially if you give it quite a low-res image, which gives the algorithm more space to improvise. The improvements here are quite striking even though Magnific takes a lot of questionable liberties with the background. The clouds too are much improved, which much sharper definition.

I’m quite impressed how Magnific’s ability to invent details where there aren’t works wonders with old scenes, papering over the sloppy modeling and texturing to deliver something much more high-end.

This means it can breath new life in old images that show their age by elevating the vegetation, textures, surface details and even the skies.

Magnific is at best when dealing with realistic renders and struggles a bit more with Sci-fi and generally anything that lacks a counterpart in the real world, which I assume is an artefact of its training data. Some of its decisions are sometimes baffling. In a few different instances, it insisted of turning hard hats into cowboy hats. Why?

In summary, anyone whose job is to produce still images is likely to find Magnific very good value in terms of time saved, which is just as well because it is eye-wateringly expensive. Those on a tighter budget will want to check cheaper alternatives, such as Krea and Leonardo, which do similar things, though not as convincingly.