credits to Thyri Perturbed Version of SD3 using script (replace input and output location)<ol><li>download sd3 from huggingface</li><li>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%)</li><li>execute the .py using "python your_py_<a target="_blank" rel="ugc" href="http://script.py">script.py</a>" (maybe you need to activate your virtual environment)</li></ol> import torchfrom safetensors.torch import load_file, save_file# Load the safetensors modelmodel = load_file('__INPUT LOCATION__')# Extract the keys and tensorskeys = model.keys()dic = {key: model[key] for key in keys}# Define parts to filterparts = ['diffusion_model']count = 0# Perturb the specified partsfor 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 += 1print(f'Total perturbed tensors: {count}')# Save the perturbed modelsave_file(dic, '__OUTPUT LOCATION__')

Files with SHA256: a612ab6f0ab88c1c811585b0a92edf3c11acc5a7e50d39d2bbedcb4b8f591e24