An 8 vector embedding that emulates the style of a particular Japanese artist, trained on the NAI model. V1.0 is recommended version.
You may notice in the image metadata different variations, like 'henreader-8' and 'henreader girl'. It's all the same embed renamed for various testing purposes, such as training the same dataset on different initialization tokens and at different sizes; the presented embedding was the best result.
Prefer LoRA?
https://civarchive.com/models/108757/henreader-lora
You can also find me on pixai.
Description
Working on another embedding, I found that 3-4k steps may result in overtraining. While I had no issues with V1 of the Henreader TI, this is an alternative version with half as much training for those who'd prefer it.
FAQ
Comments (10)
Henreader is my fav artist. Have you considered making a lycoris/lora version of this to improve the generations?
I have tried multiple times to create LoRAs but am incapable of doing so; no matter what guides I follow or which extension I install, the training process always fails.
excuse me, have u seen this ?
I don't really want to rely on Google...
HOWEVER
@AInonymous oh, i am on the same boat , i have even attempted to fully "de-googlify" my devices ..
BTW, have you tried making other artist LoRAs recently? may i suggest this artist perhaps..?
@YazawaOkeLoRA I'm always open to suggestions, particularly for artists/characters/etc that don't have a model yet. I'll give it a shot.
@AInonymous oohhhh, thank you, best wishes !!!
@YazawaOkeLoRA Alright, I now have both a TI and LoRA uploaded for Yazawa Oke; personally, I'm not all that satisfied with the results, but I did what I could.
@AInonymous oh, but it's pretty good ! of course there is some room for some improvements i guess it was because the artist has not many full color stuff besides fanwork covers and whatnot (last time i chequed hes only doing some monochrome vtuber fanart), maybe you needed more full color pics or maybe i should have asked for something less niche, Mamezou or yuihara, maybe..
@YazawaOkeLoRA There were a few problems with this artist.
1: A small dataset
2: Much of that dataset is monochrome (this is especially bad for TI embeddings)
3: Much of that dataset is rough and unfinished (also bad for TI embeddings)
Out of the ~200 images I was able to easily find, the final dataset contained only 82 of them, and even those had rough lineart which is why the TI version came out so muddy; fortunately, the LoRA seems to have largely avoided that issue.
All of that said though, it was still worth doing as a learning experience if nothing else. Also, since you mentioned Mamezo(Mamezou?), I briefly looked it up, but there seem to be a few artists under that name. If you can provide a link, I wouldn't mind at least going glancing through their work to see if it's worth making a model.
Details
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.