In general, after 30 hours of testing different Ponies. I came to the conclusion that realistic Ponies are not good friends with Lora. There will be no second version.
If I don't know something, let me know) Below is some information on how I manage to more or less avoid artifacts, but it doesn’t always work either, every other time, plus or minus.
Personally, I like how it works without Lora.
❗️Read below for a complete guide to avoid artifacts.
❗️The main emphasis is on creating beautiful compositions with girls.
❗️For portraits and near field, CFG works better at 5. Or set it higher, but use for example Adetailer or Img2img - Inpaint
This is a merger of some of my work in terms of Lora and Checkpoint Pony.
The result of the work gives more details to the clothes, body, environment.
Samplers:
Recommended:
DPM++ 2M SDE Karras
DPM++ 2S a Karras
DPM++ SDE Karras
DPM++ 2M SDE Exponential
DPM2 a
DPM++ 2S a
DPM++ 3M
DPM2 a (Best Detail)
Settings for generating Images:
Sampler: Euler a
Steps: 30-35 (But try different options)
CFG 5 - 12 (High CFG for long shots)
When using Lora, reduce its weight or play with CFG making it around 5.
Euler a - In general, he’s not good friends with Lora. In this assembly.
DPM++ 3M SDE - with Lora it causes almost no problems. If they still exist, reduce Lora's weight from 1 to 0.5. It all depends on Lora. You can also set CFG from 5 to 7
Description
FAQ
Comments (10)
Nice. But it kind of falls apart when used with Loras. Perhaps a version 2?
Which Laura do you use? Try the power of Lora to do less. If you use my Lora, then make the weights lower, they are already partially implemented here.
If I have information about the Lora used that is causing the problem, then I can debug.
@Neural_Lens It's the Desi Tabela by Desi_Cafe for Indian looking people.
@olternaut Here I made a small clarification on the use of the assembly. Now I'm testing other Lora, try it using my information.
When using Lora, reduce its weight or play with CFG making it around 5.
Euler a - In general, he’s not good friends with Laura. In this assembly.
DPM++ 2S a + Karras with Lora it causes almost no problems. If they still exist, reduce Lora's weight from 1 to 0.5. It all depends on Lora. You can also set CFG from 5 to 7
@olternaut <lora:Desi_Tabela_v3_PDXL_LoRA_prodigy_local:0.6> This works well even on Euler, the facial features are well drawn and recognition is also at the same level.
I'll add examples to the gallery. Very long-range plans in general, Pony’s face everywhere needs to be edited using Adetailer or Img2img - Inpaint
@Neural_Lens I like to use auto1111 to generate images then use a comfy workflow for upscaling and detailing. But I will try the desi_tabela at 0.6 and see what happens. Thanks.
@olternaut Added a list of best-suited samplers to the description. DPM++ 3M SDE (Best Detail)
I got really good results in tests today.
@Neural_Lens Thanks for testing. Going to follow the new guidelines.
@olternaut try yes, a lot depends on lora and its quality. I also often have problems with them.
Ihaven't tried this yet , but honestly, I like these examples. It has a place in my needs