I’d love to see Recraft support structure-guided generation using control inputs like Canny edge maps, human poses, depth maps, or segmentation masks. This would provide much finer control over composition and layout, similar to techniques used in ControlNet and IP-Adapter in the ML community.

It would significantly improve precision in generating consistent characters, poses, or environments, especially in professional workflows. Allowing users to upload or generate these structures directly within Recraft would be a game-changer.