🌕 LUMA
This is anime-based semi-realistic mix of checkpoints and loras tweaked for good results on simple prompts.
Model is fully compatible with all LoRA that I made and will make in the future.
All the content I make is and will be completely free on CivitAI.
But if you want to support me with a coin, then you can do it here.
Luma A3 VAE Style and Artist List (please note, that some of styles need additional tags like "retro artstyle, faux traditional media, painting", etc):
Please note that tags "3D, REALISTIC, ANIME" in positive or negative may change look of your images significantly.
I also higly recommend to add "score_5, score_4, ps1 screenshot, 3d, niji" in negative prompt, if you're not going to use it in positive.
Also note, that i added somme extra tags for concepts.
For example:
altshaped - will help you make humanoids with alternative anatomy
overdesigned - will help you make some complex designes if your prompt assumes it.
Complete style list:
123wsty
ahriman, anime
albrecht dürer
alex ross, retro artstyle
alexkonstad
aloija
amadmir
anato finnstark
anbudehuang
anime
arcane
asahiart, automaton, altshaped
boris valejo
brenda burke
burda, grungetech
castlevania, anime
cher-ro
cherrygig
chibi herowars
claude monet
cmg
darkest dungeon, high contrast
dashiana
degenesis
demessance
deus ex, cyberpunk
disney, flat color
dkeeno
drakonoart
dravacus
dreelrayk
eddie mendoza
edgerunners, anime
elden ring
fanfoxy
galyosef, 3d
ganas
greg rutkowski
hades
hearthstone
henry asencio
hizipro, high contrast
ikenaga yasunari
indus
ivan kramskoy
jandrew, scenery, environment
jeremy mann
jim lee
john william waterhouse
jojo
jose parra
jparked
junejenssen
kael ngu
krrat, overdesigned
kylepunk
laurelin
leicai
lol
lovk1y
makoto shinkai, style
mangamie
marta nael
max kostin
mengxuanli
mixppl
mohrbacher, divine, altshaped
mtg
nadar
nicolasaviori
niji
nixeu
othalam
painting \(medium\)
papsuev
peterskore
photo, realistic, cinematic
pixar, 3d
ps1 screenshot
qichao wang
rakavka
reliah
retrofuturism syd mead
saiya-art
sanchare
scenery, environment
shinjuku
smite
stephen pan
stumpyfongo
sugarpunk, cyberpunk
tck, altshaped
templarpainter
tixnen
tomas duchek
trufanov
valentin serov
valorant
vegera
viktor vasnetsov
vincent van gogh
warcraft
wayfinder
wlop
wroniec
wushenyou
yumingli
yuriiiko, flat color
zack snyder style, movie still, cinematic, realistic
zexiguo, overdesignedPDXL Usage:
Clip Skip 2
Sampler: Euler a / DPM ++3M SDE (SGM Uniform) (or any your favorite)
Steps: 30-40
CFG Scale: 5-7
Hires.fix: x1.5
Upscaler: 8x_NMKD-Superscale_150000_G for realism or R-ESRGAN 4x+ Anime6B for illustrations (Denoising strenght 0.2-0.3)
Hires steps: 15
Never use Face Restoration!
If you want better faces just use AutoChar or AfterDetailer.
Useful tags (negative):
wrinkled skin, double head, mutant, bad anatomy, fewer digits, bad hands, extra limbs, twisted, cropped, out of frame1.5 Usage
Recommended embeddings:
1. easynegative
2. easynegativev2
3. bad-hands-5
4. badhandv4
Recommended extensions:
1. AutoChar
2. AfterDetailer
My regular settings:
Clip Skip 2
Necessary negative prompt: easynegative or easynegativev2
VAE: kl-f8-anime2
Sampler: any
ENSD: 31337
Hires.fix: x1.5 and higher
Upscaler:
Option 1: 8x_NMKD-Superscale_150000_G for realism or R-ESRGAN 4x+ Anime6B for illustrations (Denoising strenght 0.3-0.5)
Option 2: Latent (any) or Lanczos (Denoising strength 0.5-0.6)
Never use Face Restoration!
If you want better faces just use AutoChar or AfterDetailer.
In 0.9X version you may use tags like 3d, realistic in positive or negative prompt to increase/decrease realistic shading. (Check sample images for examples)
But you may feel free to experiment.
Please note that this model is horny sometimes, so I strongly recommend to tag some clothes (pants or something) and using these tags in a negative prompt to avoid NSFW: nsfw, naked, nude
Useful tags (not necessary, just minor tweaks): volumetric light, realistic, dark atmosphere, volumetric light, caustics, absurdres, eye focus, contrast, dark scene, bright scene
Description
Upgrades, people, upgrades:
Based on 0.97.
Completely reworked lighting, volumetric and caustic effects! (Again)
Baked VAE!
More artistic (less realistic) shading.
Better contrast.
Slightly more and better details.
Slightly warmer gamma.
Minor tweaks.
FAQ
Comments (33)
I'm not getting anywhere close to the kind of fidelity showcased on the model page. I am doing 1:1 on prompts... what am I missing here!?
Are you read model description? There is all info that you need. Please check, that you have easynegative embedding, clip skip 2, and AfterDetailer extension.
I have added the easy negative embedding to my file directory, clip skip 2 has been set, and AfterDetailer to the best of my knowledge is installed and functioning... I think ADetailer is working??? I click a tick box and go.
@Mathiesin you may add some images to post, so i will gone it through pnginfo and see what's wrong, if you want.
@sadxzero
It looks so odd.... I added the easynegative embedding and it looks so much worse. first render was SUPER dark.
@Mathiesin mate, i can't tell you much just from text, gimme some examples with metadata if you want some help.
@sadxzero I just deleted my venv folder and cleared out the extensions folder. Getting better results. What upscaler should I be using with Hires?
@Mathiesin 4x UltraSharp, 4x Valar or any like (Denoising strenght 0.3-0.5)
changes between .99 and .98?
0.99 is a really strong update! It's hard to decide if I like it more than 0.97 or .98 but that's not about .99 not being good enough, and more that .97 and 0.98 were really really good on their own!
Posted some comparisons below, as well a comparison to some of my favorite other 2.5d models as a reference, and I would say 0.99 definitely holds it's own as one of the best 2.5d on CivitAI atm
Oh man, there is not enough words to describe how much i appreciate your feedback. Thank you for your words, your pics, and all your tests! You're an angel.
Love the model, been using it throughout the different iterations. A small, specific, query though on this last update. Your cherry-gig LoRa worked amazingly well on 0.98. But I've been having some issues with it on 0.99. The lighting seems to be hit or miss especially at higher weights, and inpainting seems to become a lot darker than the image. I may be doing something wrong but not sure.
Thank you! Yeah, Cherry-gig lora was trained specifically on 0.98 but it really different in base than 0.99. But this lora still works for me on 0.6-0.8 with DPM++ 2K, Karras and Alt, but yeah, lighting is quite agressive. So i probably will retrain it on 0.99 and some onther loras too. Maybe i will make a separate branch with loras only for luma or will train all styles on Luma only, dont know yet, will see about that. But i will retrain cherry-gig one for sure.
Yeah, and about inpainting. As far as i can tell, when you youse lora in inpainting mode you should slightly decrease the weight, cause image may burn out because of extra weight (it adds or multiplies, i dont know for sure, really).
I like the lightning and shading or ver 99 but in terms of following prompts it falls apart compared to the previous versions, also inpainting is also issue where the skin colour is slightly off, creating visible seems
Thank you. Will upload inpaint version soon.
@sadxzero Looking forward to it!
Version 99 is very lovely, but also very dark, and doesn't seem to respond to prompts to alter the lighting - face is always in shadow/backlit, even with a combination of regular and negative prompts against this.
Honestly, i have no idea, why you guys keep telling that 0.99 dark or something. I keep testing it still, and i haven't bump into something like that, and you may see it on sample pics, so please be sure, that you read model description and usage, or gimme some examples so we can find out whats wrong with this version.
Here's an example of the same prompt/seed side by side. Left is 98, Right is 99. Notice how the face is so much more shadowed/dark in 99? That's what we can't get rid of even with trying to prompt to change the lighting: https://cdn.discordapp.com/attachments/1010354333658910801/1115736438693179493/image.png
Otherwise I absolutely love the new model, everything else seems a great improvement style wise.
@firewraith oh, yeah, i see. But it's not critical is it? This type of contrast was a goal, honestly, so it's more of a feature than a bug. I also can recommend to use in positive/negative propmt stuff like volumetric lighting, cinematic, realistic, dark, bright, black, white, sepia, monochrome, bw in any combination. Or/and lower CFG scale for less contrast.
I can't seem to get models to hold weapons like guns, katana etc. I'm using everything as instructed, keeping the prompts simple like 1girl, holding katana/gun. Can you please help? Thank you.
How can I use this model's face just like in the first image? With what prompts? PLEASE HELP:(
Are you still going to work on version 0.98? I absolutely love it (the more realistic style compared to 0.99). Regardless, great job!
Any chance for 1.99gb ver?
I wouldn't want the quality to suffer from too much pruning. I've seen other models where the pruned version was garbage. Still, 5.5GB is quite a lot for a pruned version.
I really like the artwork you published but for some reason I can't seem to replicate it, not even close. I see a secondary resource listed, which is unavailable, could that be the culprit? What is it and how can we get it?
Thank you for the hard work and the models, and loras, I really like the styles, I hope to replicate it fully :)
Could you please upload full model currently there is only pruned version
Great. No words))) What is that knight lora people use in sample images ?