v3.3
I mostly use 30 STEPS EULER CFG 8
v3.2
I mostly use 30 STEPS EULER CFG 8
high_contrast_evo_three
I mostly use 30 STEPS EULER CFG 6
dmd2_try_CFG_1
Another try with DMD2 LoRA this time I try to get something usable at CFG 1 because is the fastest CFG
Recommended Settings:
12 STEPS LCM CFG 1 -> this is faster
12 STEPS LCM CFG 1.1 -> Better quality, slower and a little more unestable
of course there are prompts that can work well with higher CFGs, each additional tenth can have a big impact in the final result
In this version I added a new ingredient:
https://civarchive.com/models/1001945/detail-slider-lora-or-illustrious-xl?modelVersionId=1122976
high_contrast_evo_two
This version adds a new ingredient to the merge, the great checkpoint mazina
https://civarchive.com/models/858810/mazina
I mainly use Euler CFG 8 with 30 Steps
high_contrast
I created a new merge and adjusted the contrast, brightness and saturation using the LoRAs from
https://civarchive.com/models/1093089
I mainly use Euler, CFG 8, 30 Steps
Other interesting option is to use UniPC with CFG 5 as all the images in following post:
https://civarchive.com/posts/27067545
Some samples LMS with CFG 5 in this post:
https://civarchive.com/posts/27078348
v3.1
This version is flatter and colorful, for some prompts can be anime or comic like, for others still have some 3D, I mainly use Euler cfg 6.5 and 30 Steps
second_dmd2_try
This second try with DMD2 lora is more conventional, first one was too overcharged
Settings:
Sampler lcm
cfg 1-2, 1.4 and 1.5 work well for most of the prompts
Recomended steps 8-12
first_try_dmd2
It is a version with DMD2 lora https://civarchive.com/models/1608870?modelVersionId=1820705
Settings:
Sampler lcm
cfg 1-2, 1.5 works well for most of the prompts, higher cfg produces brighter vivid colors, lower cfg is darker and more realist, in the showcase most of the images are with cfg 1.5 but few were cfg=1
Recomended steps 8-12
v2.9.5
Based in v2.9
More consistent and less 3D/realist than v2.9
v3.0
After the radical version v2.9, I wanted something less radical, this v3.0 is based v2.6_spo I added more 3D and realism but less than in v2.9, I think v3.0 is more versatil
I usually use 30 steps Euler with CFG 8, but for some prompts it is interesting to use CFG 16
Examples with CFG 16 are the image
https://civarchive.com/images/121192208
and the post
https://civarchive.com/posts/26627157
v2.9
Based in v2.8_spo I added more 3D and realism keeping still colorful
v2.8_spo
Merge of v2.8 with SPO-SDXL_4k-p_10ep_LoRA_webui
v2.6_spo
Merge of v2.6_alternative_clip with the great lora SPO-SDXL_4k-p_10ep_LoRA_webui
https://civarchive.com/models/510261?modelVersionId=567119
V2.6_alternative_clip
is v2.6 with a different clip that I have picked from
https://civarchive.com/models/2099347?modelVersionId=2569166
V2.8
More vibrant colors and a little more flatter
V2.7.5
based in V2.7 has more color saturation and contrast
V2.7
A new merge based on V2.6
The version trying_semi_realistic:
It is my first attempt of making a semi realistic version, it has a known issue that it tends to add a logo or gibberish url watermark in one or both bottom corners
V2.6 adds an additional pinch of realism and 3D on top of V2.5
V2.5 I added a pinch of 3D and realism to V2
New version V2, has same ingredients as the first version but merged different. It has less saturation and more details for most of the prompts.
I have not experience in creation of models, I took some of the models that I usually use in my workflow and merged them.
In case this produces any good output all the praise for the authors of the models that I used in this merge:
Project Pygmalion :
https://civarchive.com/models/2174541?modelVersionId=2583589
XXX-Comics:
https://civarchive.com/models/2278763?modelVersionId=2564763
Palladium:
https://civarchive.com/models/2082479?modelVersionId=2399153
Kurves:
https://civarchive.com/models/2115884?modelVersionId=2479552
Artaix Realism
https://civarchive.com/models/2197692?modelVersionId=2568346
AURALIS | Illustrious
https://civarchive.com/models/2121653/auralis-or-illustrious?modelVersionId=2562369
Description
I added a pinch of 3d and realism on top of v2
FAQ
Comments (5)
This is a really interesting checkpoint, thanks for it.
On a side note, i was looking at your prompts and couldn't help but notice that you use tons of tags that aren't Danbooru ones in your neg prompt. Also your showcased pics use prompts build upon natural language. Illustrious isn't good when it comes to natural language if you aim for accuracy regarding the content that you generate. Sure it ll render pics but it's the same as flipping a coin hoping it lands on the side you bet on: not accurate and build upon luck. For accuracy ILXL checkpoints needs Danbooru tags because it has been build upon it. The sole exception to this rule is for trained checkpoints that can have been trained on extra tag. But your checkpoint isn't trained, this is a merge. So unless you ve merged checkpoints that are trained ones AND that you know all of the extra tags they have been trained on, i struggle to see the interest in using tags that aren't Danbooru ones. Sure the resulting image is different from the one you would get without those "custom tags" but this isn't the result of the parser understanding the tag but just the result of the change in your prompt token weight. Just my 2 cents, tho.
Thanks for your comment, for the checkpoint the merit is for the ingredients used.
Regarding the images as I am not very creative I like the randomness to compensate that a little :-).
Recent high-quality models seem to have less output diversity than older models.
The instability of output due to natural language prompts somewhat compensates for this shortcoming.
At least, that's my opinion :-).
@aj_prime8517 Ultimately, the important thing is to get pics that you or the people you're generating them for appreciate. However, my comment regarding the use of non Danbooru tags (that aren't tags at all in fact) remains valid regarding their impact on generated pics.😜👍
I've had great success with using natural language. I think this depends a lot of sampler, clip, and scheduler and the other components of the workflow.