Extended sizing has the clearest need in this whole list. Ranges are photographed on straight-size models far more often than not, sizes are inconsistent between brands, and shoppers order multiples to try at home and send most of them back. Seeing a garment on your own body before ordering is worth more here than almost anywhere.
We are going to be careful about what we say next, because this is the category where a confident claim would do the most damage.
How the generation actually works
The model generates from the photograph it is given. It is not selecting from a set of predefined body types, and there is no "plus-size mode" — the shopper's own body in their own photograph is the input, and the garment is composed onto it.
That is the right architecture for this. It is not, on its own, evidence that the output is accurate.
What we have not measured
We have not run a quality benchmark across body types. There is no identity-drift or body-fidelity check on the output — that check is designed and not built. So we cannot tell you how often the result preserves a shopper's actual proportions rather than drifting toward whatever the model saw most in training.
That drift is the specific, well-documented failure mode in this category, and a vendor telling you their model handles extended sizing well without a measurement to show is telling you something they do not know.
What to do about it
Test on your own catalogue and your own photographs, in the admin playground, before you put it in front of shoppers. That runs the identical pipeline the storefront uses and spends your own credits. Your judgement of your own range is worth more than our assurance.
If the results are not right for your range, we would rather you found that in a playground than in your returns data.