From now on all future versions will be named after their base model, for example if I train with the YAAI dataset on Pony V6 the version of the model will be named "PonyV6", this is to clear up confusion since I am planning on training on a lot of other base models. The original V1 and V2 models will remain with their original names.
I do want to make it clear that every new version that I make is not supposed to be "better" than the old ones, they're just different, which is why I am training on different base models each time, the only constant is the dataset.
Each model is trained on a hand picked dataset of art that I liked, results should generally speaking be a little more stylized than their original base, with some improvements on creativity as well.
Here's the V2 (WAI-NSFW V15 base used) model on tensor
Comparison of YAAI v2.0 and WAI v15: https://imgsli.com/NDE4OTg5
Comparison of YAAI v1.0 and WAI v14: https://imgsli.com/NDA5NzU4
Recommended settings:
Steps:: 15-30
CFG scale: 5-7
Sampler: Euler
Use size larger than 1024x1024 for the original dimensions
Recommended upscale settings:
Hires upscale: 1.5
Hires steps: 20
Denoising strength: 0.35~0.5
Hires upscaler: R-ESRGAN 4x+ Anime6B
All the preview images have the workflow attached, it automatically calculates a 1.5x resolution for you so you don't have to change multiple fields each time you want to change resolution.
Everything is included in the checkpoint (vae, text encoders, unet)
Recommended prompts:
Positive Prompt
masterpiece, best quality, amazing quality, absurdresNegative Prompt
bad quality, worst quality, worst detail, sketch, censor, watermark, signature, text, multiple_poses, multiple_scenes, speech_bubble, patreon_username, multiple_images, multiple_angles, bad_hands, wrong_hand, bad_anatomy, extra_fingers, extra_digits
Basically all characters in the WAI-NSFW model should work here too, you can check all the characters available here: https://huggingface.co/spaces/flagrantia/character_select_saa
Of course due to the nature of full fine tuning some characters might be impossible to be generated anymore, but from my testing all characters that I tried worked fine.
Description
FAQ
Comments (9)
支持
This model is much better than I thought it will be. Good enough quality + have no visible bias (some models generate similar faces/poses etc.) + universal enough (style can be changed easily). It's not generating something completely unique on each new seed with the same prompt, but it's already much better than in 90% of IL models on CivitAI.
Thank you, I appreciate the feedback! My goal from the start was to nudge the model into a style that I like, but not overtrain it so people can still modify it easily, I am glad that I have successfully achieved that :)
This model is great! Are you planning on updating this from the newest version of WAI-NSFW (v15 has add more new data)?
I am planning on training with the same dataset on WAI-NSFW v15 with some updated settings which should capture smaller details a bit better and also on Illustrious xl 2.0 as a bit of an experiment
@jorkingtoncityshallwe Just wanna say I've been using v2&1 for a while now and whatever fintune you've done is just the best. It erase the greasy AI feeling WAI has, also with great creativity and versatility (especially the lighting, composition and dynamic part). Really looking forward what the same finetune will do on other anime models as well!
@niniconcon First of all thank you! If there's interest I could use the same dataset for finetunning other models too, which ones would you like to see?
My only concern is that the versioning might become too complicated with this many finetunes of the same dataset, so I will probably just have the base model in the version name instead (so instead of v1, v2, it would be "WAI-nsfw v14 base", "WAI-nsfw v15 base", etc).
Also I'm curious if you would like to see any of my style loras as full model finetunes as well, and if so, which ones?
@jorkingtoncityshallwe Sorry I missed the reply, some anime checkpoints I think will be interesting to try finetuning in the same way: WAI-illustrious-SDXL v16.0 (the newest WAI update), Nova Anime XL v15.0 (more stable than WAI but less aesthetically pretty imo), DREAMBOX v4.5 (stylish like your finetune and can understand some natural language), some more stable noobAI models like Chenkin Noob XL (CKXL) v0.2 or T-noobnai3 v9, and Illustrious XL 2.0 (very weird model with potential I think).
@niniconcon I will for sure be doing a new training on WAI V16, Illustrious XL 2.0 I've tried in the past and I didn't get good results but I might try again. As for the other models that you suggested I will check them out and see which ones are worth training on.
Also I would love some feedback on my "Saikyo" model, it has many art styles built in and personally I liked the images that it produces but the model itself didn't get nearly as much traction as "YAAI", any feedback would be appreciated! https://civitai.com/models/2214085/saikyo-illustrious
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.