I know, there are several embeddings and loras already. I trained one of the embeddings, and this is my updated lora version of that dataset. trained in Kohya on 41 images, roughly 2000 steps at 0.00007 with 40 repetitions per epoch, if memory serves.
Edit: the cyberstories model used for the second half of the post is just a 50/50mix of cyberrealistic v32 and children's stories, in case you like the style of those pictures.
Description
FAQ
Comments (6)
For what its worth, I have thought the results from the previous Hailee models were either flawed or difficult to get a good result from. This LoRa works pretty well, inpainting definitely can boost it. The overall likeness is strong although my initial tests, the width of her face is occasionally off. Might need to test with checkpoints beyond Cyberrealistic.
There are certain celebs that seem to present a bigger challenge than others with SD, I don't get why, but really pleased you released this. Thanks!
Thanks. One thing that I've found works really well (and credit goes to malcolmrey for this) is using two loras for the same person or character at lower strengths. I bet if you take this one and another trained on a different dataset, and use a weight of 0.4 - o.6 for each of them, you can get a pretty good result. I admit I didn't do much testing beyond cyberrealistic, because it's my preferred photo model, but between 0.7 and 0.8 has been most consistent for me.
edit: also, use a non-latent upscaler at 0.4 -0.45, as opposed to latent. personal preference for photos, anyway.
@balbrig I've tried combining models but have had mixed results, but when it works it works. Your Hailee LoRa is very good off the shelf, but I'll give that a whirl.
Sorry, but not sure what qualifies as a non-latent upscaler.
@jacklaughed the default upscaler in hires fix is Latent, which requires a high denoise (.6 or higher) for a clear image, but that can sometimes distort a lora-based character. If you use a different upscaler like https://civitai.com/models/116225/4x-ultrasharp or https://civitai.com/models/147759/remacri or https://civitai.com/models/141491/4x-nmkd-superscale, you can set hires fix to a lower denoise setting, to preserve much of what has already generated, while still adding enough noise to improve detail. I usually use ultrasharp or superscale at .35-.45 denoise for hires fix, or something lower for upscaling after the fact.
@balbrig Thanks. I apologize, I thought you were talking about something else, should have looked before commenting.
@jacklaughed All good. Enjoy, and have a good one.