<pre><code></code></pre>This workflow allow to apply Pulid with Flux.2-Klein to three characters simultaneously. It's my first shared little work, so be kind :-) You can apply three contemporary Pulid for three different characters simultaneously. I've added a refinement group because with three applypulid nodes there is noise in the final image, but the refinement works well.I've set the parameters for klein 9b distilled so 4 steps and cfg 1.You'll find the text encoders of the suite comfyui-prompt-control because it give me better adherence at the prompts, since it's better not to raise cfg. It's easy to change them with the standard text encoders but if you want i'll provide one already setup.The TextEncodeQwenImageEditPlus node is there because klein need the image injected also in the latent and with only 1 image and 1 applypulid you can just use a vae encode but join the 3 vae encode AND the empty latent was difficult to me. You can try to adjust the weight of each applypulid from 1 to 1.4, the higher the value more noise the final image will have (at least if you work with all three the applypulid nodes) but then the refinement will take care of that. Obviously you can de-activate 1 image and the relative applypulid and work with just 2. I hope you find it useful. Update Version 2: I updated the loading of the image reference with crop face node and remove backgroud node. Now in the ID injection there will be just the face so as result: - the background of the new image follow the prompt - the clothes of the new images follow the prompt - the expressions of the characters follow the prompt easily - seems to me the final result is better, you can try in the example included in the workflowkeeping the seed fixed and generate the image with the crop face and remove background bypassed and then unbypassed. Note: I didn't use nodes grouped into subgraphs because the workflow isn't exactly "point-and-click," and I love seeing what's happening and having control. Enjoy :-)

Flux 2 Klein - 3 Contemporary Pulid

Version 2: I updated the loading of the image reference with crop face node and remove backgroud node. Now in the ID injection there will be just the face so as result: - the background of the new image follow the prompt - the clothes of the new images follow the prompt - the expressions of the characters follow the prompt easily - seems to me the final result is better, you can try in the example included in the workflow keeping the seed fixed and generate the image with the crop face and remove background bypassed and then unbypassed. Enjoy :-)

Files with SHA256: 4b6cd32df55d5dff20275bbcdbf174063fff6d14e972bd3d5e6f0d00306bb393