Multimodal models are moving toward structured spatial reasoning

Canvas-based generation, VibeEdit, and RISE-Agent treat layout, region selection, planning, and verification as explicit interfaces rather than relying on text prompts alone. The 1,000-case, 58-system visual-editing benchmark reports only a 56.6% ceiling, while V-CoLA shows that vision-token compression can improve efficiency without immediate large quality loss. Planning, calibration, and robustness under compression remain open.