Compassmix XL
Check out frosting.ai to use this model!
8 steps are recommended.
Made by lodestone-rock for frosting.ai
Description
FAQ
Comments (21)
Of the little testing I've done, I'm liking this model.
What models when in to this mix?
This does not feel like Pony or at least not heavily Pony, its responding better and seems much less unstable.
To answer my own question: seaart + secret stuff + lighting.
Lodestone is involved, which I am not surprised. xD
You beauty!
This is very nice, do I need to do anything different to train loras for this model? I can't find much info about lightning XL model lora training online.
use seaart furryxl https://civitai.com/models/391781 as the base model for training because this model is merged from it
love uploads without explaniation or examples with prompts :)))
Some info about this model.
Compassmix XL is based on SeaArt Furry XL 1.0 + Lightning 8 step lora + lodestone-rock's "secret sauce" (something to do with GANs).
Using 10-12 steps seems to improve quality a fair bit over 8.
Use Euler a (or Euler) with SGM Uniform (other schedulers seem to work okay).
Use a CFG between 2-3.
Since the model is based on SeaArt, score_x tags are not needed, use quality tags like "best quality, detailed, masterpiece" and "worst quality" in negative. Artist names were not prepended with "by", so just typing artist names by themselves is fine. The model works well with longer natural language prompts in addition to e621 tags. It has also retained knowledge from base SDXL, allowing it to understand a larger array of concepts.
how does this compare with NAI diffusion furry V3?
@dfas No idea, I haven't used NAI before. The quality you get for the amount of steps is definitely this model's strong point, but it's not perfect. It doesn't do all artist styles perfectly and you can sometimes get a bit of a noisy or grainy texture in your images. But it's definitely worth trying out.
What Clip skip this should have?
Clip Skip 2 (a1111), stop_at_clip_layer = -2 (ComfyUI)
SDXL models always use clip skip 2, and A1111 actually forces it to 2 regardless of clip skip setting. ComfyUI also auto sets it to 2 (or -2) so the clip set last layer node isn't needed. The only time you need to set it is when using nodes like the efficient loader from efficiency nodes
@CosmicElement My ComfyUI does not auto-set it to anything.
@CosmicElement Comfyui does not do this. Stop sharing this false misinformation.
@ZUSIMO comfy is literally hardcoded to load sdxl models with a clip skip of 2 (layer_idx variable)
Plz return posible publick nsfw for frosting
This model works wonderfully well for the most part, but I struggle to get simple/plain color backgrounds as well as wood(walls, furniture, etc.) that doesn't look primitive, bulky, and reminiscent of a survival games' starter home.
Very fun model to work with otherwise though.
Which VAE is the best for this model?
Is it possible to get photorealism? Any tags/artists for that effect? Any recomendations please?
an uploader that does not answer the questions in the discussion area. None of your models are worth this. If you expect a reply from this uploader, don't get your hopes up. They are as inactive as grandpa's sex life, and probably does not care, like Oprah Winfrey's giving cars out and lieing to those people about them being free(they got fucked over because the taxes was not covered by her) Yeah Skip all these models created by this person.
What are you yapping about brotato chip 🙏🙏🙏