A Causal Decomposition of Control Loss in Few-Step Diffusion: Steps, Operating Point, and Student Compensation

Few-step diffusion models are increasingly used as default generators, yet their prompt-level control remains unreliable, especially for attribute binding and spatial relations. A common shortcut is to blame these failures on distillation weights, while treating aggregate quality or overall compositional scores as evidence that semantic control has been preserved. We test this attribution with a matched factorial diagnostic: A runs the teacher with many steps, B runs the same teacher with few steps, B${}^{\prime }$ keeps the teacher and few steps but switches to the student operating point, and C runs the few-step student. The resulting contrasts separate step compression, operating-point change, and student-weight effects. Across SD3.5 and FLUX, the student-weight contrast is not negative after operating-point matching; instead, student weights compensate relative to the matched few-step teacher, while the dominant non-weight failure source is family-dependent. The same contrasts induce a conditional deployment recipe, validated by dose-response sweeps: SD3.5 recovers under guidance tuning, whereas FLUX recovers primarily by increasing steps rather than by guidance alone. The diagnostic also refuses attribution when the many-step teacher is near floor. Thus “few-step control loss = distillation damaged semantics” is a misattribution: matched factorial evaluation is needed because aggregate metrics can hide which causal factor failed.