Updated: Mar 31, 2025
base model
This is a model based on @Creeper_MZ's "Soft_Noob". The original model was trained on Noob version 0.5. @solid_snek requested that someone "modernize" Soft Noob for "a more modern version". This model has been "modernized" with Noob Epsilon 1.0 and 1.1 (and 1.1 version will be uploaded soon... it takes 6 hours to upload this thing...)
Please checkout, download and give lots of love to the original models!
Soft Noob can be found here: https://civitai.com/models/880074
Noob Epsilon 1.0 is here: https://civitai.com/models/833294?modelVersionId=1022833
Recommended settings:
Sampler: DPM2 Ancestral (also called: DPM2 a) (for general and flat colored images), LCM (for simpler-looking images and for the occasional funky style), PLMS (for realistic colors, might get muddied, use more steps if the image is muddied and/or Turbo sampler with less steps and it might still mess up... but the images that come out right are worth it)
Scheduler: Karras, Align Your Steps (11/32/GITS), Laplace
ClipSkip 1 for more 2.2-2.5D colors
ClipSkip 2 for flatter colors
In Automatic1111/forge/reForge interfaces you have an option to "upcast cross attention layer to float32"
(search the settings for "txt2img" > type in "upcast_attn" > Apply Settings > Reload the UI).
It will appear in your settings section under the seed. I recommend you upcast your attention layer. This model will benefit from it, sharpening the details immensely.
The use of a VAE is recommended, especially if the colors are not looking good. Adding the SDXL VAE makes a substantial difference.
You can find it here: https://civitai.com/models/296576/sdxl-vae
Anime-stylized SDXL VAEs work very well also.
I recommend this pack by @atw44qb: https://civitai.com/models/152040/xlvaec
If the pack is too confusing: use C1 for cel-shaded and flat images, F2-2 for smooth shaded and G9 for 2.5D and 3D. You can also be creative and mix-and match your prompting with the VAEs.
Thanks Everyone!