Type | |
Stats | 511 183 |
Reviews | (89) |
Published | Apr 13, 2023 |
Base Model | |
Trigger Words | man, male, woman, female, turbant |
Hash | AutoV2 4A9193EAD2 |
Happy easter! (late, I know!)
Copy pasting from our Patreon:
"Welcome to the wearable LORA series. And welcome to the Experimental part of LORAs by my side.
This embedding is a spinoff of our patreon's character Seizo, the mage desert.
It focus on certain ethnics and headgear, in this case turbant. Classic and iconic item.
I have focused so we have diversity when prompting and even going higher values, it remains like that.
How did I train this?
This LORA is trained in 3 steps. Normal training, re-training (there is a free tutorial in our patreon and lora mixing for final product).
More info in the future since this is a thing I'm working on but these are the basics (always changeable as need).
How this works?
These type of LORA are designed to get good results with minimal prompting. You can see them as "lazy"
LORAs but, they also work with larger prompt, however if you want to get "lazy" you will need to do few things. Of course this lora works on any custom model (1.5 based).
You can either use Clip Skip 1 or 2. Clip Skip 2 will make AI to get more freedom in the generation.
Settings for inferencing:
For painterly style (lazy) (very good for make upscaling and inpainting)
CFG Scale = bellow 5, from 2.5-5 is good value.
Steps = Above 40 necessary (on Euler Ancestral you can try less)
Sampler = Euler Ancestral for variability and Karras Samplers for technical details
For normal style derivative to the model (normal)
CFG Scale = 7 and you can go above.
Steps = 20-30 are good values, 40 could work too.
Sampler = Same as above but SDE works good here.
How to prompt?
This is a "lazy" LORA. It works the same way of other LORAs you may have seen but this lora due to a LORA custom merge can extend it's possibilities:
- It can get good results with just 1 or 2 words. Since it's a wearable with turbant stylish you just have to put man/male woman/female and words like turbant (you can also choose color).
If you use the painterly technique remember to drop way way more than normal CFG Scale and up the steps. You can also use Hi-res Fix scaling to 1, denoise 0.15-0.20 to maintain images a bit cleaner.
- You will get variability even at higher LORA values (over 1 and till 1.5 from my tests) but you can also go 0.X values.
Auxiliar Lora that I recommend even it's not necessary: Epinikion's Offset Noise
This LORA is good for...
Portraits, standing situations, black background images, dark images, high contrast,
normal actions like drinking, sitting, etc. and MIXING with other character LORAs
This LORA is NOT good for
Any sort of landscape. Is a wearable, is not designed for this.
Putting your character in any special situation. It suffers when you don't want to do portraiture or medium standing situations but if you tweak values from 1.2 to 0.8 probably it will be fixed."
If you liked this content please consider support us in our patreon (https://www.patreon.com/user?u=90661954) for more quality material so we can continue doing tutorials, workflow and embeddings.
PS: All of our models, embeddings, etc. are free-use. We don't trust in license and don't want to limit open sourcing.