Controlled generation (ControlNet)

Generate a new image whose composition matches a source image while its style, subject, or finish changes freely. You extract a structural map (pose, Canny edges, depth, or segmentation) from the source, then condition generation on that map. This is the "same composition, new look" lever: interior redesign, sketch-to-render, and repeatable product or game-asset variations.

Inputs to collect

• The source image to take structure from (a photo, sketch, render, or existing asset). Required.
• Which structure to preserve: silhouette/outlines (Canny), human pose (pose), 3D layout/perspective (depth), or scene regions (segmentation). If unstated, infer from the goal: outlines for objects/assets, pose for people, depth for rooms/scenes.
• What changes: the new style, subject, material, or palette described as a prompt.
• How tightly to hold the structure (loose restyle vs strict trace) and how many outputs.

Models

Two pieces, both confirmed live via runware-models before calling, never hardcoded: