Both models start from the same base model and are trained on the same data for the same number of iterations. Only the control representation differs. Some cases also contain static cubes or blobs for background control. Green means the subject turns as requested. Red means it does not turn, turns too little or too much, or shows hallucinated motion.