Hi!
This lora should produce more photorealistic images, with more knowledge on skin features, and darker, dramatic lighting.
The results change acording to the checkpoint, sample and scheduler you use, so you may need some testing to find the best config for you...
Be careful with negative tags, on tests on civitai, it seens that negatives may make the skin too smooth.
It seens to work better with Anima Turbo, so i'm not sure the best config for use without it yet. (It seens that Er_sde SGM_UNIFORM, with lower cfg 2.5~4 works nice)
Update:
Lora str can be good up to 1.15 for stronger skin effects
er_sde linear_quadratic OR Euler Normal worked best on my tests (Focusing on skin features)
The triggers are:
photorealistic, real skinAnd skin features:
heavy freckles, light freckles, body freckles
moles, mole on body, butt moles
acne, acne on body
goosebumpsOther recomendations:
depth of field, rim lighting, dim lighting, sidelighting, soft sunlight, window shadow, shade
face focus, upper body focus, cleavage, close-upPS:
I'm short of time lately, so i tested using LLM to make the prompts, so they can be kinda messy, but you can get some ideias.
I tested mostly on Anima Turbo, as my pc take 8+min to generate images without it, and the results change a lot. Euler Normal, Euler Simple, Er_sde linear_quadratic, seens to get more smooth results on Anima, but this change when generating without it.
If you generate some images wih it, please post it so more people and I can get a better view of what it can do! :D
Description
Quick version for Anima.
FAQ
Comments (10)
Based on your showcase images, this LoRA seems to further accentuate one of the main drawbacks of any Krea 2 model: its tendency to add fluids and skin imperfections of all kinds.
Yeah, the extra fluid effects introduced by the LoRA are definitely a drawback; even though I was careful with the dataset, one image slipped through.
But based on my tests so far, the images only became extremely messy when adding "sweat" or "wet" effects, also, sampler and scheduler make a huge difference it seems...
Also, some of the showcase examples were using Sam Anima Real Turbo + Anima Turbo (I didn't notice Sam's already had Turbo in it), so that multiplied those problems xD.
In my opinion, v1 turned out really well. Depending on the checkpoint, you need to adjust the strength to get a good result, but a value between 0.6 and 1.5 will almost certainly give you what you're aiming for. I am very satisfied with the LoRA. Thanks for it.
Thanks for the feedback! I didn't try it up to 1.5 yet, which checkpoint you tested with this str?
As I said, it depends on what you want to achieve; with semi-models, you need more, though I think 1.5 is the absolute upper limit.
https://civitai.red/models/2677907/pornmaster-animasemi-realistic
https://civitai.red/models/2717940/nova-reality-am
With checkpoints that are already good at realism, you tend to use values of just 0.6–0.8 like with Sam Anima.
https://civitai.red/models/2409949/sam-anima-realistic?modelVersionId=3060073
@Gorean I see, i'll give it a try on those. Thanks!
@DomPanda One thing I’ve noticed now is that it darkens the image significantly. Perhaps you could create a version that preserves the color and lighting.
@Gorean Oh, i forgot to actually explain that on the model description, but this is one of the focus of it too, both realskin and darker/dramatic light
@DomPanda aaaaah ok... is there a way to get one without darken? :)
@Gorean Yep, i think it is possible with a new dataset more focused on natural lighting, i'll give it a try when i have more time ^^