Vers2.0 I removed the Sam3 node and connected the "krea2 edit model" node from the first sampler. In the second sampler, the face improves even more. 3. sampler for details.
Hi everyone! I'm sharing my custom ComfyUI workflow, built on the Krea2 architecture, which combines identity editing and advanced painting/masking processes. With Image to Image, you can slightly alter an artwork you're inspired by to fit it into your desired identity. You can turn a pose that inspires you from an influencer into your own pose. It's an experimental process designed to create visually stunning compositions by blending realistic photographic textures or your preferred styles. I tried to avoid identity confusion on the face by masking it with SAM3.
Workflow Highlights:
Base & Model: Krea2 Jib Mix https://civarchive.com/models/2799984/jib-mix-krea-2
Text Encoder: I prefer the heretic one https://huggingface.co/DreamFast/Qwen3-VL-4b-Heretic-ComfyUI/tree/main
Masking & Inpaint: SAM3 (
sam3.1_multiplex_fp16.safetensors) Native NodeLLM Prompter: Krea's own text encoder Qwen3 VL
Optimization: Speed-optimized pipeline with Sage Attention support.
☕ Buy Me a Coffee https://buymeacoffee.com/ugurdoyduk
Happy generating, and I'd love to see your results in the gallery!
Description
I removed the Sam3 node. I connected the "krea2 edit model" node from the first sampler. In the second sampler, the face improves even more. 3. sampler for details.
FAQ
Comments (3)
The facial consistency is extremely poor. Even with the "Maintain perfect facial consistency" setting, it only "barely" looks somewhat similar. Adjusting Lora's proportions didn't help. I even downloaded the same model as the author. I also tried over 10 KREA2 models; while the author's model did have the highest consistency, the overall results were still poor… The essence of face swapping is achieving extreme similarity, and currently, the KREA2 workflow with the highest face swap consistency still requires adding a corresponding character like Lora.
Yes, unfortunately, Krea's Lora is very good, but since we had to lower the denoise setting in Img2Img, its effect decreases relatively. We need a new WF that replaces the face with Klein.
Unlike other posts, I am having other problem. No matter what reference image (body) I use. It makes totally different body, background. Or, create some kind of creature from John Carpenter's Thing.
In prompt,
I put "Replace only the head of woman in image 1 with the head from image 2." tried to switch the number 1 and 2. Also, I tried to use the only default prompt you made. Both didn't work.
I did use the latest version of jim mix,
Jib Mix Jalapeno INT8 Convrot - Krea2 - er_sde,6-12steps.safetensors
I saw other face/head swap workflow for krea2, but nothing worked for me, even face/head swap workflow for other models worked with the same prompt