The claim
T2I-CompBench has a non-spatial relation category (interactions such as holding, wearing, looking at). Unlike its attribute and spatial categories, it is scored by CLIPScore, which the same paper shows tracks human judgement poorly, and human scores on it are near ceiling (0.95–0.99 for five of six models). So an interaction benchmark exists, but one that cannot separate a complicit exchange from a reprimand.
What it does not say
It does not weaken Case III; it sharpens its positioning. The manuscript should cite the non-spatial category as the baseline it extends, not describe the benchmark as object-centered throughout.
Since the journal version
In T2I-CompBench++ (LIT-798) the interaction category is judged by GPT-4V, which matches human rankings better than CLIPScore (Kendall τ 0.48 against 0.25). The “scored by CLIPScore” half of this claim is true of the conference version only. Human scores on the category are still 0.95–0.99, so the near-ceiling half stands.
Fu et al. (LIT-938, read in NOTE-813; THEORY-250) extend the point about automatic scoring to vision-language-model judges. In their App. A.6, VLM judges correlate with human ratings of relational images at r ≈ 0.43–0.56, against a human split-half reliability of 0.91.