The AI depth presets had never been measured against each other. They have been now, on real frames, and two of them were not producing a depth map at all.
The whole ladder. Top row is the source frames; below them every Distill preset from the cheapest down to the reference, with what each one costs and how closely it agrees with it. Two presets were broken. They have been replaced. Distill Fastest and Distill Very Fast scored 0.60 and 0.82 against the largest model on the part of the picture that actually holds shape - a railing, a figure, a branch. At 0.60 there is no depth left: in the street frame above, a person with an umbrella dissolves into the road. They are now 210x210 and 280x280, at 0.86 and 0.87. Both kept their place in the list, so whatever you had selected still works and simply has a better model behind it. The new Fastest costs 0.2 ms more than the old one; the new Very Fast is better and no dearer.
A new preset: Distill Balanced (322x322). There was a jump from 2.1 ms straight to 4.3 ms with nothing in between, so a machine that could not quite afford Fast dropped a long way. Balanced sits in that gap at 2.8 ms. The automatic picker has shorter steps. It now walks 1.6, 2.4, 2.8, 4.3, 7.7, 11.6 and 24 ms. It picks by fitting your frame budget, and a ladder with a hole in it makes it choose between too little and too much. Everything above is in all 16 languages.
The presets above Fast are untouched, and on these four frames they add very little over it - the difference is in thin structures that these particular frames do not contain. If you chose Normal or higher for something specific, keep it. Rolling back: right-click FlatVR in your library -> Properties -> Betas -> last_stable. Settings are kept either way.