Important discovery (crazy)
Yeh so i just discovered if you put a white square in the character reference image box then use a well trained character LORA you get a perfect swap with very minimal artifacts. I use between 1-1.5 strength on my character lora. I use the v3 distill and the fro adaptive head swap v3. I'm using bongtangent 8 step and LCM or ER SDE.
I trained a character LORA on AI toolkit with 20 static images. I imagine if you did it with videos it might be even better.
But yeh, its really crazy. The swap is much more accurate and realistic for me and can handle distance a bit better. Also, there may be solution to use both reference and lora together but i didnt do that yet. I just know a white square + character lora is amazing
White square picture: https://encrypted-tbn0.gstatic.com/images?q=tbn:ANd9GcSBEKRYXicmzzIRS4wBubLX71V7KOGCL0Fy-w&s
Bottom line is I think theres a great alternative future for this LORA with a character lora rather than a ref image. I think it could be perfect and the best open source face swapper.
That's great! I thought about trying out the idea of using a white square as a reference, but I didn't end up testing it. It's wonderful that you tested it! Do you have a workflow you could share?
workflow! please! thks! :)
Which face-swapping effect is better compared to Wan2.2 Animate combined with a character LoRA?
Allisonerdx, please try this and put out a new workflow with the lora loader in there!
white square? Don't understand... where? Your image provided doesn't help I don't see the image only the white square... lol
white square? Don't understand... where? Your image provided doesn't help I don't see the image only the white square... lol
Instead of sending an image of a reference face, you send a blank image and use a Lora of the character to transfer the new identity.
Nice! Seems to work even better with an Empty Latent Image instead of a white square. ymmv