Donate for civilians in Ukraine, Syria, Mali etc.: <a target="_blank" rel="ugc" href="https://www.icrc.org/en/donate">Red Cross</a> | <a target="_blank" rel="ugc" href="https://donate.doctorswithoutborders.org/secure/donate">Doctors without Borders</a> | <a target="_blank" rel="ugc" href="https://www.directrelief.org/">DirectRelief</a> | <a target="_blank" rel="ugc" href="https://www.feedthechildren.org/">Feed The Children</a>🔶 Description"What, another one of these?". Yes, I made it because everything I've tried had a minor flaw (nothing crazy, sometimes artefacting, sometimes face didn't look right, sometimes didn't abide by prompt). So I gave it a shot too. This is a doll aesthetic model with my finetuning for face and proper nudity. Pleasant semirealistic/2.5D style and very accurate to the prompt. Does not have clothing artefacts when prompted for nudity, upscaling artefacts or bias towards childish features.🔶 RuleDon't generate child porn/loli (this model doesn't do photorealism and does not churn out children by default unlike some other models but I don't want to see illustrated child porn either). <a target="_blank" rel="ugc" href="https://pedo.help/help">Find free, anonymous help</a> with pedophilic tendendies before a child is harmed.<h3>✨ Guide for best images</h3>(Update: Alternative try just High Res Fix with these settings, this was used for the example images of v1.1: Denoising strength: 0.55, Version: v1.2.1, Hires upscale: 1.8, Hires upscaler: Latent (nearest-exact))Step 1 - txt2img<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/2dfcde21-796e-4a92-b07a-5b4237bb0b5e/width=525/2dfcde21-796e-4a92-b07a-5b4237bb0b5e.jpeg" />Step 2 - img2img with UltimateSDUpscale extension<img src="https://image.civitai.com/xG1nkqKTMzGDvpLrqFT7WA/73d9bb3b-2559-4517-bc52-a7abeaeb7195/width=525/73d9bb3b-2559-4517-bc52-a7abeaeb7195.jpeg" />Credits for parts of modelV08_V08, breastinclassBetter_v13, fashionGirl_v36Fix, koreandolllikenessV20Credits for my LoRa training:I did it with <a target="_blank" rel="ugc" href="https://colab.research.google.com/github/hollowstrawberry/kohya-colab/blob/main/Lora_Trainer.ipynb">this colab</a> for training and <a target="_blank" rel="ugc" href="https://colab.research.google.com/github/hollowstrawberry/kohya-colab/blob/main/Dataset_Maker.ipynb">this one</a> for dataset cleaning. Source images are diverse so this model does not mimick a single individual.Pruning InfoThe model is pruned with the following settings using <a target="_blank" rel="ugc" href="https://github.com/lopho/stable-diffusion-prune">this repo</a> from 4GB to 2GB at virtually no difference in the generations.python <a target="_blank" rel="ugc" href="http://prune.py">prune.py</a> --fp16 --ema --no-depth delicate_v10.ckpt delicate-v1-0-full-ema-pruned.ckptSimply download the script <a target="_blank" rel="ugc" href="http://prune.py">prune.py</a>, create a new conda environment (conda create -n prune), then activate it with conda activate prune. Run pip install numpy safetensors torch!=1.13.0. Then run the command above and replace the first checkpoint name with your non-pruned one and the second checkpoint name with whatever you want your pruned model to be called. Don't use safetensors as the extension (use ckpt instead), otherwise it may fail during inference. You can convert pickle ckpt to safetensor in the Checkpoint Merging tab of Automatic1111.The model took 15 attempts or so probably and the LoRa in total 20 or so. I think it paid of that I didn't rush it though and hope you enjoy it too :)

Delicate v1.1 🧚‍♀️

Added Edge of Realism which refines the style and accuracy

Files with SHA256: 01e297cc754d0cbc7a831669cd58b31e194d1edd586876daa054151bf6d9989d