composable_stable_diffusion.ipynb

Kernel: Unknown Kernel

Composable Stable diffusion

Composable Stable Diffusion proposes conjunction and negation (negative prompts) operators for compositional generation with conditional diffusion models. This script was contributed by MarkRich and the notebook by Parag Ekbote.

pip install torch numpy torchvision diffusers

import torch as th
import numpy as np
import torchvision.utils as tvu
from diffusers import DiffusionPipeline
import argparse
import sys

# Simulate passing arguments explicitly (bypassing Jupyter's arguments)
sys.argv = [
    "ipykernel_launcher.py", 
    "--prompt", "mystical trees | A magical pond | dark",
    "--steps", "50",
    "--scale", "7.5",
    "--weights", "7.5 | 7.5 | -7.5",
    "--seed", "2",
    "--model_path", "CompVis/stable-diffusion-v1-4",
    "--num_images", "1"
]

parser = argparse.ArgumentParser()
parser.add_argument("--prompt", type=str, default="mystical trees | A magical pond | dark",
                    help="use '|' as the delimiter to compose separate sentences.")
parser.add_argument("--steps", type=int, default=50)
parser.add_argument("--scale", type=float, default=7.5)
parser.add_argument("--weights", type=str, default="7.5 | 7.5 | -7.5")
parser.add_argument("--seed", type=int, default=2)
parser.add_argument("--model_path", type=str, default="CompVis/stable-diffusion-v1-4")
parser.add_argument("--num_images", type=int, default=1)
args = parser.parse_args()

# CUDA Setup
has_cuda = th.cuda.is_available()
device = th.device('cpu' if not has_cuda else 'cuda')

# Assign parameters
prompt = args.prompt
scale = args.scale
steps = args.steps

# Load pipeline
pipe = DiffusionPipeline.from_pretrained(
    args.model_path,
    custom_pipeline="composable_stable_diffusion",
).to(device)

# Disable safety checker (if intentional for internal use)
pipe.safety_checker = None

# Generate images
images = []
generator = th.Generator("cuda").manual_seed(args.seed)
for i in range(args.num_images):
    image = pipe(prompt, guidance_scale=scale, num_inference_steps=steps,
                 weights=args.weights, generator=generator).images[0]
    images.append(th.from_numpy(np.array(image)).permute(2, 0, 1) / 255.)

# Create and save image grid
grid = tvu.make_grid(th.stack(images, dim=0), nrow=4, padding=0)
tvu.save_image(grid, f'{prompt}_{args.weights}.png')

print("Image saved successfully!")

composable_stable_diffusion.py:   0%|          | 0.00/27.6k [00:00<?, ?B/s]

Loading pipeline components...:   0%|          | 0/7 [00:00<?, ?it/s]

composing ['mystical trees', 'A magical pond', 'dark']...

  0%|          | 0/50 [00:00<?, ?it/s]

Image saved successfully!