SD3 Medium Slightly Decensored (Perturbed)版本%2 (clips+t5 fp16) (clips+t5 fp16)

credits to Thyri

Perturbed Version of SD3 using script (replace input and output location)

1. download sd3 from huggingface

2. put script in a .py (replace input with sd3 model location and output location with where and what you want to name it). modify 0.01 to how much you want it to be perturbed (0.01 = 1%)

3. execute the .py using "python your_py_script.py" (maybe you need to activate your virtual environment)

import torch

from safetensors.torch import load_file, save_file

# Load the safetensors model

model = load_file('__INPUT LOCATION__')

# Extract the keys and tensors

keys = model.keys()

dic = {key: model[key] for key in keys}

# Define parts to filter

parts = ['diffusion_model']

count = 0

# Perturb the specified parts

for k in keys:

if all(part in k for part in parts):

v = dic[k]

print(f'{k}: {v.std()}')

dic[k] += torch.normal(torch.zeros_like(v), v.std() * 0.01)

count += 1

print(f'Total perturbed tensors: {count}')

# Save the perturbed model

save_file(dic, '__OUTPUT LOCATION__')

描述:

训练词语: