imgen-diffusers/flux-schnell.py

from diffusers import FluxPipeline
import torch
from pathlib import Path
import re
from datetime import datetime

def slugify(text):
    # remove non-word characters and foreign characters
    text = re.sub(r"[^\w\s]", "", text)
    text = re.sub(r"\s+", "-", text)
    return text

prompt = "the town center of a small futuristic town with a fountain and dodecahedral d20 buildings similar to small geodesic domes with a road leading out to grassy plains sparsely dotted with very broad tall trees"
height, width = 720, 1280

ckpt_id = "black-forest-labs/FLUX.1-schnell"

DIR_NAME="/nas/dockerdata/immich/local/"
dirpath = Path(DIR_NAME)
# create parent dir if doesn't exist
dirpath.mkdir(parents=True, exist_ok=True)

# denoising
pipe = FluxPipeline.from_pretrained(
    ckpt_id,
    torch_dtype=torch.bfloat16,
    use_safetensors=True,
)
pipe.vae.enable_tiling()
pipe.vae.enable_slicing()
pipe.enable_sequential_cpu_offload() # offloads modules to CPU on a submodule level (rather than model level)

output = pipe(
    prompt,
    height=height,
    width=width,
    num_images_per_prompt=4,
    num_inference_steps=4,
    max_sequence_length=128,
    guidance_scale=0.0,
)
print('Max mem allocated (GB) while denoising:', torch.cuda.max_memory_allocated() / (1024 ** 3))

# import matplotlib.pyplot as plt
# plt.imshow(image)
# image.save("./whitehenge.png")
# plt.show()
for idx, image in enumerate(output.images):
    timestamp = datetime.now().strftime("%Y%m%d%-H%M%S")
    image_name = f'{slugify(prompt)}-{idx}-{timestamp}.png'
    image_path = dirpath / image_name
    image.save(image_path)