Why I Don’t Like the New AI AWB on the Fujifilm X-T5
Fujifilm used “deep-learning AI Technology” to improve Auto White Balance on the X-T5 (or, more accurately, on X-Trans V cameras—not just the X-T5). According to the promotional statement, the camera is able to more accurately identify warm tints, and adjust to compensate for that when using Auto White Balance. Sounds impressive, right?
When I first learned about this, I was a little concerned that the new Auto White Balance would affect Film Simulation Recipes that use AWB. So I took a few test shots with the X-T5 and an X-Trans IV model side-by-side to compare, and I didn’t notice any difference between the two regarding white balance. It looked the same to me. But now that I’ve used the X-T5 a little longer, I do, in fact, at times notice something that I initially overlooked.
In the banner above, which comes from Fujifilm’s promo materiel for the X-T5 (even though the X-H2 has this same feature, it wasn’t promoted with that camera), you can see the “conventional model” vs the X-T5 AWB rendering in identical light. I assume that the so-called conventional model wasn’t a Sony or Canon, but an X-T4 (or other X-Trans IV camera). I personally prefer the more golden rendering of the “conventional” AWB to the copper rendering of the AI AWB, but each has their own tastes, so there’s no right or wrong answer. Perhaps you prefer the image on the right over the one on the left. It’s definitely subjective.
Something I have noticed—and I don’t like—is that this new rendering is inconsistent. From one exposure to the next, with identical lighting and identical settings, you can get something more like the “conventional model” rendering or something more like the AI AWB rendering. I’ve noticed it in artificial light, and I’ve noticed it in golden-hour/sunset situations. Two exposures, one right after the other—nothing’s changed—but the camera produces two very different tints when using AWB. Take a look at the two pictures below for an example of this. They were captured under identical light with identical settings, but they clearly aren’t identical. This was in a set of 32 pictures (of my son opening birthday gifts); 19 had the golden-ish cast and 13 had the copper-ish cast (these are frames nine and ten, for those wondering).
Obviously if you are a wedding or event photographer, and you rely on Auto White Balance, this could be a big issue for you, because you want consistent results. You don’t want the white balance to be bouncing back-and-forth between two tints. I don’t even want it for my son’s birthday pictures! If the camera chose one rendering in the situation, and consistently applied that to each image, whether gold or copper or something else entirely, that’s fine—it’s what is expected to happen—but bouncing between renderings is bad and should not happen. If you can’t trust AWB, and if it’s a tool that you commonly use, the X-T5 (or any of the X-Trans V models) might not be the camera for you.
Of course, for many people this might not be an issue whatsoever. Maybe you don’t even use AWB. Perhaps you do but you don’t care if the results are different between exposures. It could be that you’re going to adjust white balance in software later anyway, so what the camera records makes no difference to you. If that’s you, and none of this matters to you, great! But I do want to point it out for those who it might matter for, because they should know. It’s better to find out now before dropping so much money on something that’s just going to frustrate you.
I imagine that this is something Fujifilm could fix fairly easily via a firmware update. A simple tweak to the code could possibly make this behavior happen much less frequently. Fujifilm should address this issue. I hope in a few months from now this will all be a past problem that was fixed and forgotten. Or it could be the expected behavior that all Fujifilm X-Trans V cameras will have, and it will only be fixed by an even more improved AI-AWB on X-Trans VI models. Time will tell.
See also: Five Fujifilm X-T5 AI AWB Workarounds