Canny guidance follows visible boundaries, so it works well for product silhouettes, architecture, line art, and compositions with clean edges. Depth guidance follows near and far relationships, which is more useful for rooms, people, vehicles, and scenes where perspective matters more than every contour.

Start with one control signal. If the result is too rigid, simplify the source or reduce the amount of detail in the prompt. If the result drifts, use a cleaner reference with a strong subject outline and fewer competing background elements.

The control image already describes much of the layout. Use the prompt to name the new subject, material, environment, lighting, color palette, and finish. Add preserve instructions for the parts that must remain recognizable, such as camera angle, pose, product outline, or room perspective.

Avoid asking the model to keep the structure and completely replace every object at the same time. Change one visual layer first, compare the result to the reference, then make a second pass if the composition still holds.

Use ControlNet when geometry is the priority: pose, edges, depth, perspective, or placement. Use image to image when identity, texture, color, or a localized edit matters more than reproducing the exact structure.

Review hard edges, repeated lines, fingers, product labels, and contact shadows at full size. A result can match the broad map while still breaking small geometry, so keep the source beside the output during approval.