Qwen Edit don't really like to do face swaps by prompting only. But there's way to force him to do this. Whole trick is around marking some area in image in red color (just blob, square) passing additional image that would fit into that area (like face) and telling magic word "remove red area". Poor Qwen don't have a choice and have to put something in a place of red blob. <ul><li>workflow uses rembg node to color masked area as red and openpose node to generate pose</li><li>you can use any color, will probably work</li><li>you can pass collage of 2-12 photos of person (in one image) with different perspectives and emotions (only face closeups)</li><li>you can use this trick to replace anything, like a hairstyle only, shoe, car (without openpose then)</li><li>all can be more automated with some face auto masking, but I want release simple stuff</li><li>if red is somehow passed to result image, like a red santa hat, make blob more round</li><li>it will come out plastic fantastic and not really matching the skin color, but light/shadows will be correct (skin tone matching anybody?)</li><li>you can add to prompt stuff like "wearing sunglasses", will totally work</li><li>if you mask 2 faces and pass 2 images (in one) of two different persons there's a good chance it will work (probably depends how distinct are those faces)</li><li>to increase quality you can cut out only face and torso (to provide body size context) and pass only that, then stitch it</li><li>you can use this red area trick to just perfectly inpaint/outpaint (no additional image)</li></ul>In v2 I've added some stuff to use only masked area and leave rest of pixel untouched (exact same final image size). Sometimes qwen changes stuff outside red area, then this will not work.You can try remove bypass on "Set Latent Noise Mask" and/or "ReferenceLatent Image 1" if things get weird, maybe it will help.

Files with SHA256: 93e88442ec1a422d5f397c0ce8aea61eb9d6d78de0b078ab6322bc296786fa29