Many anime Loras have extravagant clothing and styles compared with realistic ones. Have you had trouble applying these Loras to realistic checkpoints? Look no further! It is essentially a txt2img and img2img merge. This can also be helpful to guide checkpoints into concepts they don't understand.
Let's focus on just the Model Swap portion using an otherwise empty workflow.
The concept is simple. Load a checkpoint that gets your character looking correct, after the Ksampler send that latent image to another Ksampler. You will also want to send the original CLIP and create a new conditioning from it, you can use the same text prompt. (Sometimes it helps to use the same lora with much less weight as well.) Load a realistic checkpoint and send the 'model' to the Ksampler. Set denoising below 0.5, I find 0.4 to be a good starting point. CFG will need modified along the way, higher on the anime Ksampler and lower on the realistic.
At this point you will have a photo that looks similar to the original anime, but with a few realistic-ish features like face and hands. We will need to run another pass with either adding noise or upscale (1.5x works well) then send to the 3rd Ksampler. Since we are changing the quality (noise/upscale) we will need to run a denoise closer to 0.5 just like a normal upscale.
In the example photos, I am using 'Edea' from Final Fantasy 8. With a realistic checkpoint SourceoftheGoddess-v3-LCM it draws a woman in a dress with decent representation of her headpiece. However, it misses the essence of Edea.
Using DreamCrafter-v10-LCM, it properly includes the metallic half-circle frame on her back as well as the fur trim and better looking headpiece.
We start with DreamCrafter, draw the first photo, send it to SourceoftheGoddess, and finally upscale and resample. If your final image does not have the proper likeness, we can add the same lora with lower weight in the 2nd pass after loading the realistic checkpoint.
Let me know your thoughts in the comments!