Online generating: https://tensor.art/models/922075998860672113/NOWAI-(NOVA-WAI-model-merge)-12.5.15
This is an experimental merge between Nova Anime XL and WAI NSFW Illustrious SDXL.
Main purpose: for trial and feedback only, not intended as a final release.
P/s: I also want to learn more about controlling pose and color adjustments with this checkpoint. This merge is part of my personal experiments, so feedback or tips on improving pose accuracy and color consistency would be greatly appreciated.
Changes:
v0.3-test:
Nova Anime XL - IL v12.0 (a) × WAI Illustrious v15.0 (b)
v0.2-test:
Nova Anime XL - IL v11.0 (a) × WAI Illustrious v15.0 (b)
Created mainly for testing and comparison with the older merge.
Note: the new WAI version has been quite disappointing for many users, so this merge might be less stable or weaker than the previous one.
v0.1-test:
WAI Illustrious v14.0 (a) + Nova Anime XL (v10.0) (b)
Method: ComfyUI-DareMerge > Model Merger (Block)
Values: [1, 1, 0.96, 0.64, 0.05, 0, "cosin"]
Description
Nova 16 + Wai 16
CLIP / VAE: WAI
FAQ
Comments (3)
hi, dear Yuichi69 , you write :"controlling pose and color adjustments" - ok, I have a test arsenal for the most difficult poses that SDXL can handle at its limits. I'll post some pictures and see if the model can handle them. By the way, did you know that the pose you like, "calem_\(pokemon\), sitting on one knee," is considered one of the most difficult for SDXL because of the triangle? But, you also wrote about color and lighting—what's there you want to test?
Hi! Thanks for the test offer.
That note was written when I had just started merging SDXL checkpoints. Back then, I was wondering whether a checkpoint could handle difficult poses better without using ControlNet — mainly because I didn’t even know what ControlNet was yet 😅.
What I really wanted to understand was how to improve pose accuracy clearly without losing the original style of a checkpoint. I thought about approaches like applying a LoRA and exporting a new checkpoint, but that usually alters the style toward the LoRA itself.
The post is quite old now and hasn’t been updated yet, but I’m still interested in seeing how the model performs on challenging poses.
@yuichi69 I've fully tested your model. The safest poses are generated by the "dynamic pose, foreshortening, extreme perspective" token. Here are some examples: https://civitai.com/images/120632194.
There's also another token for generating extreme poses, but it's making the model struggle, "breaking" the character in about 40% of cases.