Added two more girls, increased details and vividness.
The model is based on the FenrisXL model. I trained on 10.000 photos from my archive. These are photos that I took myself on Nikon D610. The goal is to create a model depicting figures more similar to the people next door, rather than supermodels. I have also included many interesting backgrounds shot in Slovakia, Czech Republic, Romania, Croatia, Hungary, as well as many photos from Russia. I plan to develop the model further.
Sampling method:DPM++ 2M Karras
HiResFix: 8x_NMKD-Superscale_150000_G /day:0,4
CFG:6-9
VAE is baked in (sdxl_vae_fp16_fix.safetensors)
negative prompt: ugly, watermark, [deformed|disfigured], poorly drawn, [bad : wrong] anatomy, [extra | missing | floating | disconnected] limb, (mutated hands and fingers), blurry
Recommended workflow:
create your basic prompt, 768x1024 or 768x1152 recommended, with hiresfix 1,25x
in I2I double the resolution CFG7 and denoising 4, same prompt
in Inpaint correct details, cfg 7, denoising 4, only masked, 1024x1024, prompt only what you are changing
resize in Extras, do not use codeformer, 2x is enough
I will add some new girls, new locations to the new version.
Description
Added two new girls (1000 photographs), added mid journey pictures, increased details and vividness. Little better with hands. Still I recommend use InPaint and hand refiner.
FAQ
Comments (3)
Amazing work!
very beautiful landscapes , but more info on what this model is assure better at ?
, these generated landscape looks worth it amazingly looks realistic.
I've trained this model on photos I've taken over several years on the Nikon D610. So the landscapes are coming out well. But also only with the help of gradual enlargement and adding details. I have included unretouched photos of the girls in the model, so the faces come out realistic. I was not able to remove the many-fingeredness. So you need to use control net.