Build a ComfyUI Workflow for Consistent Pose and Style
Academy is temporarily unavailable while it is being rebuilt.
A common ComfyUI problem is getting both the layout and the look you want in one image. A text prompt may describe a product, object, or landscape, but the exact pose, outline, depth, lighting, or palette can drift. The workflow taught in this course solves that by separating structure from appearance: ControlNet guides the layout, while IP-Adapter uses an image prompt for visual influence.
Start with two references. Use the first as the structural guide: a pose image, an edge drawing, or a depth-based scene layout. Use the second as the appearance guide: a color palette, lighting mood, material style, or object reference. Keeping the roles separate makes troubleshooting easier.
Build the ControlNet branch first. Load the guide image, run the matching preprocessor, then connect the processed guide into Apply ControlNet. Choose the preprocessor for the job: OpenPose for body positioning, Canny for clear outlines, or depth for foreground-background structure. Load a ControlNet model that matches the base model family used by your checkpoint.
Next add the IP-Adapter branch. Load the appearance reference, encode it with CLIP Vision, load a compatible IP-Adapter model, and connect the adapter according to the node pack you installed. Use a general adapter for style, palette, or object appearance. Use a face-focused adapter only for permitted source material and when facial similarity is the actual goal.
Test in three passes. First, generate from the text prompt alone. Second, enable only ControlNet and check whether the pose, edges, or depth follow the structural reference. Third, enable IP-Adapter and check whether color, lighting, or appearance moves toward the second reference without overwhelming the layout.
Change one influence setting at a time, such as ControlNet strength or IP-Adapter weight. A successful result passes two visual checks: the composition still follows the structural guide, and the appearance reference is visible enough to matter. If one reference dominates, reduce its influence or limit when it is active during sampling.