save_lora.py

#!/usr/bin/env python3
import torch
from warnings import warn
from diffusers import (
    AutoencoderKL,
    DiffusionPipeline,
)
import hashlib

base = "stabilityai/stable-diffusion-xl-base-1.0"
adapter1 = 'nerijs/pixel-art-xl'
weightname1 = 'pixel-art-xl.safetensors'

adapter2 = 'Alexzyx/lora-trained-xl-colab'
weightname2 = None

inputs = "elephant"
kwargs = {}

if torch.cuda.is_available():
    kwargs["torch_dtype"] = torch.float16

#vae = AutoencoderKL.from_pretrained(
#    "madebyollin/sdxl-vae-fp16-fix",
#    torch_dtype=torch.float16,  # load fp16 fix VAE
#)
#kwargs["vae"] = vae
#kwargs["variant"] = "fp16"
#

model = DiffusionPipeline.from_pretrained(
    base, **kwargs
)

if torch.cuda.is_available():
    model.to("cuda")


def inference(adapter, weightname):
    model.load_lora_weights(adapter, weight_name=weightname)
    try:
        model.fuse_lora(safe_fusing=True)
    except ValueError:
        warn(f"{adapter} and {weightname} is broken. LoRA is not fused.")
        model.unload_lora_weights()

    data = model(inputs, num_inference_steps=1).images[0]
    model.unfuse_lora()
    model.unload_lora_weights()
    filename = '/tmp/hello.jpg'
    data.save(filename, format='jpeg')
    with open(filename, 'rb') as f:
        md5 = hashlib.md5(f.read()).hexdigest()
    print("Adapter %s, md5sum %s" % (adapter, md5))
    if md5 == '40c78c9fd4daeff01c988c3532fdd51b':
        print("BLACK SCREEN IMAGE for adapter %s" % adapter)


inference(adapter1, weightname1)
inference(adapter2, weightname2)
inference(adapter1, weightname1)
inference(adapter1, weightname1)