Jennifer

Model description

Jennifer is a SD1.5 lora trained with SDXL generated images. The dataset was made a while ago with SDXL of a latin woman, while using controlnet for vareous poses, and detailers for the hands and eyes. (comfyui). Training was done with AI-toolkiit on the SD1.5 base model. Not the easiest model to fully control. It was trained on an exreme small data set consisting of roughtly 7 images.

the lora is very weak !!!

works with a range of 32 - 46 steps DPM++SDE, DPM 2M SDE, UniPC (32 steps), Euler a sheduler: karras and simple

lora weights: 0.85 to start with, you can drop to 0.65 or up to 1.25, to support the lora one could use an IPAdapter with an example image from the page to improve quality (weight: 0.1 <> 0.5) otherwise the lora will collapse Adetailer: eye detailers work, for the hands it's seldom a succes.

Controlnets for openpose can be used, again low weights

My advise is to generate with a realism model and after the Ksampler/VAE decode -> encode again with an SDXL VAE -> Ksampler for SDXL -> upscale with model -> ESRGAN X2 plus gives good results

use a realism SDXL model, latent text encoders: (positive latent: quality prompt)(negative prompt: negative quality) Ksampler settings: (denoise 0.1 <> 0.3) (samplers: DPM++SDE or DPM++2M SDE)
(shedulers: simple, exponentional, SGM_uniform may give nice results to) (use seperate seed) (septs: 32 - 42 don't push it to high)

Trigger words

You should use Jennifer to trigger the image generation. Aditionally "latin", "woman","1girl", "seductive" can be used to support the lora

Download model

Download them in the Files & versions tab.

Downloads last month: 10

Model tree for HeartandSoul/Jennifer

Base model

stable-diffusion-v1-5/stable-diffusion-v1-5

Adapter

(646)

this model