Make sure to use UI that supports v-prediction, otherwise you will get a noisy blob!
NoobAI-XL (NAI-XL) based.
Main parameters:
Steps: 28
CFG: 3-5
Sampler-scheduler: euler/euler a-normal, DDPM-exponential (better noses)
CFGZeroStar and sometimes Self-Attention Guidance: default for difficult backgrounds
Try to BREAK prompts into parts by 75 tokens max, this goes for most models..
Positive prompt:
number of characters (1girl/1boy), character name, artist prompt, scene/environment/camera angle, action, expression, items, quality prompts (masterpiece, best quality, absurdres)some heavily stylized characters might require prompt: nose, realistic.. you can put nose into Adetailer
Negative prompt:
worst quality, low qualityeven longer negative should work, but I prefer simplicity.. tags worth noting include sketch, abstract, url
Enjoy..
Description
Added SPO-SDXL_4k-p_10ep_LoRA_webui.
Changed dataset marginally.
FAQ
Comments (8)
I might have done something wrong, but my tests with anime coloring (flat) always generated an excessive amount of detail. It seems like a very powerful model in principle, but very hard to control.
You are very right, i sacrificed flat colors for a shift in default look, i might try to reconsider though cause i improved my dataset drastically also. If i manage to get the style looking nice to me while preserving flat colors i ll post v3. Cheers and thanks for trying out my model! Your images on the previous version are very cool.
@blinkdotleh Thanks for the update. The flat color is kept nicely, to be honest. What happens is that there are loads of extra details, as if I had cranked denoise up to 0.7 (I used 0.25, so, I didn't). For example, loads of creases on a beach towel. You might notice also on Eiko's belly in v1 posts, that shadows truly are a lot (and this happens before hires, so it's likely intrinsic in your model).
@blinkdotleh FTR, I posted an image which shows better what I was talking about (although v.3.0 so far makes it better than 2.0).
@einar_rainhart less crunches in fabric than v3 produces wont be possible im afraid, it stems from the dataset.. i ll try lowering learning rate in kohyass and see what it does cause that seems to making some issues.. EXPERIMENTING, HEHE
Yooo I can't wait to try V4! (After I fix my venv lol).
Haha, yes, I would say it's my best fine tune yet.. About venv, I grew tired of git pulling and downloaded the comfy.exe, works great!
@blinkdotleh I thought of doing that too! I might look into that. :)