Reply
Tue 29 Sep, 2026 09:20 pm
When working with reference-image workflows (single img2img or multi-image fusion), what is the most reliable way to separate what should change from what must stay preserved?
For instance, when feeding two or three reference images (e.g. one for subject pose/structure, one for style/palette, and one for background context), diffusion models often blend features unpredictably—such as lighting from the style reference bleeding onto the subject, or fine facial details getting altered.
Beyond adjusting prompt weights and boundary phrasing, what practical steps or preprocessing checks do you find most effective to keep control over the output before running iterative generations?
@sumytbe,
The biggest win for me is giving each reference only one job. Rather than feeding the pose image in raw, convert it to a depth or pose map first so it can't pass on colour or lighting. For the style reference, crop out any people and check its lighting direction roughly matches your subject, otherwise it'll fight you. Masking helps too: apply the style adapter to the background region only and keep the face area protected. Then lock the seed and change one variable per run. If faces still drift, finish with a low-denoise inpaint on the face.
Stumble It!
Tweet This
Bookmark on Delicious
Share on Facebook
Share on MySpace