Prompt Alignment measures semantic agreement between protocol anchor frames of a generated video and the target instruction T=sample.prompt_target. It evaluates whether the prediction realizes the requested scene content at benchmark-sampled time points, not whether the full rollout is temporally coherent or visually realistic.
Let A={fi}i=1M denote the anchor-frame set selected by the evaluation protocol from the decoded prediction. With CLIP ViT-B/16 image encoder ϕI and text encoder ϕT, define L2-normalized embeddings ei=ϕI(fi) and t=ϕT(T). The native CLIPScore for frame i is
si=max(100⋅⟨ei,t⟩,0)(1)
WorldAtlas maps each anchor score to [0,1] via s~i=clamp(si/100,0,1) and reports the mean
Sprompt=M1i=1∑Ms~i(2)
Higher Sprompt indicates stronger instruction following at protocol anchors. The metric is official when explicit instruction supervision is available and conditioning is not self-referential (conditioning_strategy ≠ self). Image suites may condition on a single reference frame, but the model prediction must still be a video.
The examples below are curated from the complete 2026-07-17 image_static snapshot: 20 models evaluated on the same 3,372 cases. Every side-by-side pair is strictly verified to share the sample ID, conditioning-image SHA-256, and actual generation-prompt SHA-256; both generation records must also identify that image as the conditioning input. “Good” and “poor” refer only to Prompt Alignment—not overall video quality—and each score is the mean over four segment_random protocol anchors.
