Reviews
NoobAI-XL (NAI-XL)
Overwhelmingly Positive
99% of 2,194 reviews recommend it · 2,194 reviews
2,17717
Elegh
· Epsilon-pred 0.75-Version
Recommends
Still in love with this version of the checkpoint! Super versatile on its own, and it really brings out the best in any LoRA.
0 comments
Read review by Elegh
Sennke
· Epsilon-pred 0.75-Version
Recommends
it will be new base model !
0 comments
Read review by Sennke
Great, versatile, more knowledge than NovelAI Diffusion V3 even within characters existing within its database (like Roll.EXE)... Handles artist styles & directions better than any other anime model out right now (try adding eyeshadow/eyeliner and such to Roll.exe on NovelAI, it will not work. But aside from that things like theme, backgrounds, etc all way better here)... Welcome to the future, I have high hopes for the proposed V1.0 model and hope this continues to be trained after that for newer concepts as well, but even if it isn't this could likely be the best local model for quite a while.
0 comments
Read review by fizalpher
CuauhtemocI5MAL
· Epsilon-pred 0.75-Version
Recommends
There're two things I would like to see in the next version of this model: The ability of replicating the style of artist which their artworks has been uploaded to e621 and the ability of replicating the style of the first NovelAI's furry model.
And it appears its compatible with Lora designed for Pony Diffusion V6 XL.
0 comments
Read review by CuauhtemocI5MAL
rerolls26
· Epsilon-pred 0.75-Version
Recommends
a huge improvement to 0.5, if you are having problems with the model check the strength and mix of your artist tags
0 comments
Read review by rerolls26
0.75 version is worse than 0.5 in pretty much every way
less colors, bodily anatomy is less consistent, anime style is suffering and most artist tags look less accurate than they did before
drop the e621, it's not worth training on it, it just makes the model ugly
0 comments
Read review by Kewtie
Similar to the previous version, this model has a good understanding of anatomy, concepts and artist styles.
In the latest version 0.75, I noted that it performs better in terms of anatomy, lighting as well as in lesser known concepts / styles (low occurrence count in dataset)
However, in my testing I found the images generated to be less aesthetically appealing than that of the images generated with version 0.5.
I observed that the images generated tend to lean towards a yellowish (sepia) tint. Additionally, the colors appeared to be flatter (lesser color variation) and that the outlines appear to be thicker.
Following this, I found that while artist styles and concepts with lower tag counts have generally improved in performance, there are some styles / concepts that were previously performing well in 0.5 that are showing degraded performance in the newer version of the model.
Could all this perhaps be due to the model's lack of convergence at the current point in time? (I noted that Text Encoder training has begun in this phase.)
0 comments
Read review by r3c