IlluNoob 3.0
New features in 3.0:
Lots of changes:
Improved anatomy accrossed the board, no more objects going through the body magically, fingers, toes generally improved
Better NSFW
Huge improvements on poses, should get rid of the rigidness that SDXL sometimes has
Better faces, better gestures, characters should feel more alive
Better prompt adherence, compared to 2.0
Improved artist/character knowledge
Should be generally easier to use
Some minor furry improvements
What's new in v1.2?
Aesthetic finetune of v1.1, it has all the artists/characters from v1.1 and:
Prompt adherence
Artist Knowledge
Pose Knowledge
Natural Language Adherence
Based on my QuantumMerge method, this model is the result of leveraging the much more stable clip of Illustrious 1.1 and the amazing artist/character knowledge of NoobAI EPS 1.1. The resulting model has:
All the artist/characters knowledge of NoobAI
All the NL prompt abilities of Illustrious 1.1
In addition, due to how this merging method works, the best weights were selected automatically between the two models. What does this mean?
The model will, for any given prompt, choose the best weights between the two models. In practice, based on my tests, this means that if a character is better recognized by Noob, the weights from Noob will be used for it, same goes for artists too. Likewise, if Illustrious has better grasp of a pose/clothing element, the Illustrious 1.1 weights will be used.
The merging method is highly experimental, so feedback is very welcome.
Support:
I tend to not relase 10s of models weekly but rather focus on releasing the models I truly like and provide the best for the community.
I highly recommend that you support both OnomaAI and LAX, as this model heavily relies on their contributions.
If you'd like to, you can also support me by paying the early access fee or on Patreon:
https://patreon.com/Lilogummie
This is entirely optional, if you'd like to check out the model prior to early access ending, you can download it from huggingface:
https://huggingface.co/lylogummy/illuNoob-1.1
Description
Trained the model a bit further and merged with personal mix, this is basically an aesthetic finetune - should be easier to use and generate higher-quality images
Core Settings:
Steps: 50
Sampler: Euler Ancestral CFG++
Schedule type: Karras Dynamic
CFG scale: 1.5
Size: 1024x1728
VAE: sdxl_vae.safetensors
Clip skip: 2
Emphasis: No norm
Advanced Settings/Experimental Features
guidance_limiter_enabled: True
guidance_sigma_start: 25
guidance_sigma_end: 0.28
rescale_cfg_enabled: True
rescale_cfg_version: Normal
rescale_cfg_multiplier: 0.5
Prompt Structure
meta tags (masterpiece, best quality), rating:###, 1girl, character, series, other tags
FAQ
Comments (12)
Version 1.2 Post (open for details)
Due to a CivitAi error, I cannot edit the description, I've had the best results with the below config:
Core Settings:
Steps: 50
Sampler: Euler Ancestral CFG++
Schedule type: Karras Dynamic
CFG scale: 1.5
Size: 1024x1728 (or any SDXL-compatible res)
VAE: sdxl_vae.safetensors
Clip skip: 2
Emphasis: No norm
Advanced Settings/Experimental Features
guidance_limiter_enabled: True
guidance_sigma_start: 25
guidance_sigma_end: 0.28
rescale_cfg_enabled: True
rescale_cfg_version: Normal
rescale_cfg_multiplier: 0.5
Prompt Structure
meta tags (masterpiece, best quality), rating:###, 1girl, character, series, other tags
Dropped the 45,000 buzz remaining to unlock it early for everyone interested. Also, your QuantumMerge utility looks incredible and I cannot wait to try it out. Thank you for your contributions!
Thank you so much!! This truly means alot! I am really grateful for the amazing support, for the Quantum Merge method feedback is very appreciated and welcomed as it's still very experimental, I've deployed an update yesterday which added an UI to make it easier to use so please do give it a try and let me know your thoughts 😊
this model is pretty good as well. I just try it on 1344x2048 and it not bad pretty well bro
Appreciate it! Haven't tried that resolution yet as I gen mostly at 1024x1732 but I'll try - glad to hear you like it
In theory the model should do both low res (1024x1024) and hires without a big drop in quality but hires seems to be more detailed
I like the 1.2 version very much, but the contrast of the generated images are a bit too high for my taste. Is there any way to adjust it? Thank you!
Hi, Thanks so much for your feedback, I'm glad to hear you like 1.2. I think the most straightforward way to lower contrast is to lower the cfg. If using a cfg++ sampler, lowering to cfg 1 should have quite an effect. I'll check to see if I can also lower the contrast via the negative prompt and get back to you asap
Ask for merged recipes.
And looking for cooperation projects.
Hi, the original recipe is a merge between illustrious 1.1 and NoobAi 1.1 epspred using the default settings from here: https://github.com/GumGum10/QuantumMerge-sdxl/tree/main
V2 is merged with an unreleased model that I've trained on score:>200 booru images, and then it is further trained on an aesthetic dataset to align
I'd love to collaborate, please dm me the details
@LyloGummy The merge algorithm I'm currently using is a custom algorithm, and I feel it would be too sad if only I were using it. I'd like to invite you to try it out.