Three text-to-diagram prompts, each rendered by Princigram under two training regimes. Without structured physical annotation, the scene, geometry and topology become incorrect, while the on-figure symbols collapse into illegible gibberish. With SP-CoT, the same architecture generates the entities, relations and values specified by each prompt.
Structured Physical Chain-of-Thought annotates each training image with an explicit chain of physical entities, quantities and relations. This ablation isolates that single ingredient — everything else in the pipeline is held fixed.
A pulley system with two masses: mass m1 rests on an inclined plane at angle θ, connected by a rope over a pulley to a hanging mass m2.
An 8.00 kg block on a horizontal table, connected by a rope over a pulley at the table’s edge to a 6.00 kg block hanging vertically.
Two blocks 1 and 2 on a horizontal surface: block 1 moves toward block 2 with initial velocity v1, while block 2 connects through a spring to a fixed wall.
Across all three cases the pattern is identical: the ablated model produces plausible-looking artwork whose physics and labels are meaningless, while the SP-CoT model places every mass, pulley, angle and velocity where the prompt demands and renders each symbol (m1, m2, θ, 8.00 kg, 6.00 kg, v1) legibly. Structured physical labeling is what makes the generations physically faithful rather than merely photorealistic.