First pass of my Squeaky Toy LoRA in XL. Didn't really change the training much, so what worked there mostly works in XL, but there are definitely some differences in how XL likes to be prompted with it.
Depending on the model and prompt, I've been using weights of 1.0, 0.8, or 1.0:0.8 (which only moves the unet to 0.8) with the last being the most consistently decent.
Works really well with lightning models, but it has a hard time with faces there for full size images, I've been using it ~exclusively with ADetailer in that mode.
Starter prompt:
<subject>, squatting, arms raised, leather collar, chewing a squeakytoy
Description
FAQ
Comments (10)
Not able to get this model to produce the squeeky toy
Can you share a prompt and base model you're trying with?
https://civitai.com/posts/3402162 I made over 40 of them these are representative
@seruko862 Ah, I see, you're trying with a Pony model. This is trained against base SDXL, so it's not super surprising that it doesn't work with 3d pony models. As far as I can tell they seem to have 're-learned' what faces look like, so SDXL stuff doesn't seem to map well to them. I'll bump this thread if I do a Pony fine tuning with the same data, it might take out of the box.
@thegipper Really appreciate the insight, thank you
@thegipper is there a realistic checkpoint you would recommend besides Base SDXL?
@seruko862 I generally use Epicphotogasm (which it looks like you found), but I haven't' tested much with the newer revisions. I also generally use ADetailer to redraw the face with the squeakytoy prompt (and any details like "text" on the toy) at 0.4-0.55 strength, which makes the faces look better and really adds a ton of detail.
@thegipper Thank you very much! I really appreciate the help and inisght
@seruko862 closing things out, I tested out the newest version of that model and posted some examples based on your prompt. All of these use adetailer to do one additional pass that cleans up the face and adds higher detail for the squeakytoy itself. I'm not sure how you're generating images, but if you draw those in Automatic1111 into the "PNG Info" section, it'll let you replicate the specifics of the prompt and subsequent adetailer pass.
@thegipper that's more or less it. Thank you again