Reduce Attribute Bleeding in Two-Person AI Images
Academy is temporarily unavailable while it is being rebuilt.
A common failure in two-person generations is attribute bleeding: the prompt says one person wears a green coat and the other carries a black backpack, but the model mixes those details between them. A practical way to reduce this is to separate structure from description. Use OpenPose ControlNet to define where each person is positioned, then use regional prompting in a ComfyUI-style workflow to assign different text conditioning to each area.
Start with a clean pose reference or an OpenPose skeleton containing both people. The skeleton should clearly show which figure is on the left and which is on the right. Preview the OpenPose preprocessor output before generating. If arms overlap confusingly, one body is missing, or the two skeletons cross into each other, fix the reference first; text prompts usually will not solve a broken layout.
Next, apply the OpenPose ControlNet to the positive conditioning. Begin with moderate strength rather than pushing it to the maximum. The goal is for the model to follow the two poses while still having enough freedom to render clothing and the environment naturally. If the figures drift away from the skeleton, increase strength slightly. If they look stiff or distorted, reduce it.
Then create separate regional prompts. With rectangular areas, assign one prompt to the left side, one to the right side, and a lower-strength background prompt across the full image. For example: left region, “adult woman in a green coat, brown boots”; right region, “adult man in a navy jacket, gray scarf”; background, “quiet city park, soft overcast light.” Keep each regional prompt focused on only what belongs in that region.
If rectangular zones cut through a body, use mask-based regional conditioning instead. A painted mask can follow the silhouette of each person, which helps reduce hard seams and makes accessories less likely to jump across regions.
Check success by looking for three observable results: each person stays near the intended pose, clothing colors remain on the correct subject, and the background does not inherit subject-specific details. If attributes still mix, tighten the region or simplify the prompt before changing unrelated sampler settings.