Krea 2 head swap in ComfyUI by Stable Yogi
Two photos in. The clothes, the pose, the background and the light all stay. Only the head changes and it changes properly, hair and all, instead of averaging the two people together.
That averaging is the point of this post. It is what the recipes going round give you by default: blonde hair that still has the first woman's curl in it, and a face that is neither of them. Two things decide it.
One LoRA, not two.
There is a head-swap LoRA and it is tempting to stack it on the identity one. Six versions on the same seed and the same two photos: every version with two LoRAs blended the two people, at every ref_boost. Every version with the identity LoRA alone swapped cleanly. It is not a strength you can tune your way out of — it is the wrong stack.
The instruction has to say REMOVE.
"Put the head from photo two onto the woman in photo one" never tells the model to get rid of the first head, so it averages them. Three lists, not one: hold the base, remove the old head completely, then name what travels with the new one — the hair, its colour, its length, the way it falls, the eye colour, the nose, the jaw.
Two more worth knowing. The model list going round for this names the wrong VAE — Krea 2 ships exactly one and it is not the Wan one; use the wrong file and it finishes with no error and saves coloured static. And a source photo that is not a clean 16:9 leaves a thin smeared band along the bottom. That band is arithmetic, not the model.
The workflow is free, the checkpoint is my own and also free, and the walkthrough video shows every node in order.
Workflow, model download links and the full write-up: https://forgebun.com/go/civit_headswap_wf
The checkpoint: https://forgebun.com/go/civit_muse_krea2
Credit where it is due: the identity-edit LoRA is conradlocke's and the node pack is lbouaraba's. The checkpoint, the workflow, the settings and the write-up are mine.
Description
FAQ
Comments (5)
great workflow, but its 50/50 when it comes to whether or not it works, some of the time it outputs head swap properly but sometime it just a image of ref headshot super imposed onto scene
Thanks, that's useful to know. That pasted-on look usually means the swap step didn't really re-render the area, it just laid the reference over the top instead of rebuilding it into the scene.
Two things worth trying. Nudge the denoise up a little on the swap pass, if it sits too low the model leaves the reference more or less untouched, which is exactly what you're seeing. And check the head is actually being detected, because if the detection misses, it can quietly fall back to just compositing.
One thing that would really help me pin it down. When it goes wrong, is it the same reference image failing every time, or does the identical setup sometimes work and sometimes not?
I made the following modifications to your workflow based on my own needs, and it’s working very well.
・I added LoraLoader again after “Krea2Edit” to support multiple Lora models.
・I applied masks to Image 1 and the face image, and used ref_boost_mask and ref_boost_mask_a to preserve the face and emphasize the areas of Image 1 to be retained.
• I align the orientation of the original image and the face image as closely as possible.
• I match the ratio of the face size in Image 1 to the face size in the face image relative to the image. Trying to use a large, standalone face image for the face in a full-body image sometimes doesn’t work well.
I’ve tried various models, but within my environment and the scope of my testing, your Realism model works best. Next is the Muse series—though performance varies by version—but I found v1.5pro to be the easiest to use. With models from other creators, I sometimes ran into issues, such as the model not maintaining the pose from the source image.
I love Realism.
Yeah, the head swap is a 50/50. Even with modyfing the workflow I still get those 50/50 results. I tried with auto crop head, inpaint, prompting and lowering the denoise.
@SwapThatLad yeah, thats honest and it matches what i see. most of it comes down to the source picture, a clear face pointing roughly the same way lands far more often. still working on making it less fussy.