On this tutorial, we design a sensible image-generation workflow utilizing the Diffusers library. We begin by stabilizing the setting, then generate high-quality photographs from textual content prompts utilizing Steady Diffusion with an optimized scheduler. We speed up inference with a LoRA-based latent consistency strategy, information composition with ControlNet underneath edge conditioning, and eventually carry out localized edits by way of inpainting. Additionally, we deal with real-world methods that steadiness picture high quality, velocity, and controllability.
!pip -q uninstall -y pillow Pillow || true
!pip -q set up --upgrade --force-reinstall "pillow<12.0"
!pip -q set up --upgrade diffusers transformers speed up safetensors huggingface_hub opencv-python
import os, math, random
import torch
import numpy as np
import cv2
from PIL import Picture, ImageDraw, ImageFilter
from diffusers import (
StableDiffusionPipeline,
StableDiffusionInpaintPipeline,
ControlNetModel,
StableDiffusionControlNetPipeline,
UniPCMultistepScheduler,
)
We put together a clear and suitable runtime by resolving dependency conflicts and putting in all required libraries. We guarantee picture processing works reliably by pinning the right Pillow model and loading the Diffusers ecosystem. We additionally import all core modules wanted for era, management, and inpainting workflows.
def seed_everything(seed=42):
random.seed(seed)
np.random.seed(seed)
torch.manual_seed(seed)
torch.cuda.manual_seed_all(seed)
def to_grid(photographs, cols=2, bg=255):
if isinstance(photographs, Picture.Picture):
photographs = [images]
w, h = photographs[0].measurement
rows = math.ceil(len(photographs) / cols)
grid = Picture.new("RGB", (cols*w, rows*h), (bg, bg, bg))
for i, im in enumerate(photographs):
grid.paste(im, ((i % cols)*w, (i // cols)*h))
return grid
machine = "cuda" if torch.cuda.is_available() else "cpu"
dtype = torch.float16 if machine == "cuda" else torch.float32
print("machine:", machine, "| dtype:", dtype)We outline utility features to make sure reproducibility and to arrange visible outputs effectively. We set world random seeds so our generations stay constant throughout runs. We additionally detect the accessible {hardware} and configure precision to optimize efficiency on the GPU or CPU.
seed_everything(7)
BASE_MODEL = "runwayml/stable-diffusion-v1-5"
pipe = StableDiffusionPipeline.from_pretrained(
BASE_MODEL,
torch_dtype=dtype,
safety_checker=None,
).to(machine)
pipe.scheduler = UniPCMultistepScheduler.from_config(pipe.scheduler.config)
if machine == "cuda":
pipe.enable_attention_slicing()
pipe.enable_vae_slicing()
immediate = "a cinematic picture of a futuristic road market at nightfall, ultra-detailed, 35mm, volumetric lighting"
negative_prompt = "blurry, low high quality, deformed, watermark, textual content"
img_text = pipe(
immediate=immediate,
negative_prompt=negative_prompt,
num_inference_steps=25,
guidance_scale=6.5,
width=768,
peak=512,
).photographs[0]We initialize the bottom Steady Diffusion pipeline and change to a extra environment friendly UniPC scheduler. We generate a high-quality picture straight from a textual content immediate utilizing rigorously chosen steering and determination settings. This establishes a powerful baseline for subsequent enhancements in velocity and management.
LCM_LORA = "latent-consistency/lcm-lora-sdv1-5"
pipe.load_lora_weights(LCM_LORA)
strive:
pipe.fuse_lora()
lora_fused = True
besides Exception as e:
lora_fused = False
print("LoRA fuse skipped:", e)
fast_prompt = "a clear product picture of a minimal smartwatch on a reflective floor, studio lighting"
fast_images = []
for steps in [4, 6, 8]:
fast_images.append(
pipe(
immediate=fast_prompt,
negative_prompt=negative_prompt,
num_inference_steps=steps,
guidance_scale=1.5,
width=768,
peak=512,
).photographs[0]
)
grid_fast = to_grid(fast_images, cols=3)
print("LoRA fused:", lora_fused)
W, H = 768, 512
format = Picture.new("RGB", (W, H), "white")
draw = ImageDraw.Draw(format)
draw.rectangle([40, 80, 340, 460], define="black", width=6)
draw.ellipse([430, 110, 720, 400], define="black", width=6)
draw.line([0, 420, W, 420], fill="black", width=5)
edges = cv2.Canny(np.array(format), 80, 160)
edges = np.stack([edges]*3, axis=-1)
canny_image = Picture.fromarray(edges)
CONTROLNET = "lllyasviel/sd-controlnet-canny"
controlnet = ControlNetModel.from_pretrained(
CONTROLNET,
torch_dtype=dtype,
).to(machine)
cn_pipe = StableDiffusionControlNetPipeline.from_pretrained(
BASE_MODEL,
controlnet=controlnet,
torch_dtype=dtype,
safety_checker=None,
).to(machine)
cn_pipe.scheduler = UniPCMultistepScheduler.from_config(cn_pipe.scheduler.config)
if machine == "cuda":
cn_pipe.enable_attention_slicing()
cn_pipe.enable_vae_slicing()
cn_prompt = "a contemporary cafe inside, architectural render, comfortable daylight, excessive element"
img_controlnet = cn_pipe(
immediate=cn_prompt,
negative_prompt=negative_prompt,
picture=canny_image,
num_inference_steps=25,
guidance_scale=6.5,
controlnet_conditioning_scale=1.0,
).photographs[0]We speed up inference by loading and fusing a LoRA adapter and show quick sampling with only a few diffusion steps. We then assemble a structural conditioning picture and apply ControlNet to information the format of the generated scene. This enables us to protect composition whereas nonetheless benefiting from artistic textual content steering.
masks = Picture.new("L", img_controlnet.measurement, 0)
mask_draw = ImageDraw.Draw(masks)
mask_draw.rectangle([60, 90, 320, 170], fill=255)
masks = masks.filter(ImageFilter.GaussianBlur(2))
inpaint_pipe = StableDiffusionInpaintPipeline.from_pretrained(
BASE_MODEL,
torch_dtype=dtype,
safety_checker=None,
).to(machine)
inpaint_pipe.scheduler = UniPCMultistepScheduler.from_config(inpaint_pipe.scheduler.config)
if machine == "cuda":
inpaint_pipe.enable_attention_slicing()
inpaint_pipe.enable_vae_slicing()
inpaint_prompt = "a glowing neon signal that claims 'CAFÉ', cyberpunk model, practical lighting"
img_inpaint = inpaint_pipe(
immediate=inpaint_prompt,
negative_prompt=negative_prompt,
picture=img_controlnet,
mask_image=masks,
num_inference_steps=30,
guidance_scale=7.0,
).photographs[0]
os.makedirs("outputs", exist_ok=True)
img_text.save("outputs/text2img.png")
grid_fast.save("outputs/lora_fast_grid.png")
format.save("outputs/format.png")
canny_image.save("outputs/canny.png")
img_controlnet.save("outputs/controlnet.png")
masks.save("outputs/masks.png")
img_inpaint.save("outputs/inpaint.png")
print("Saved outputs:", sorted(os.listdir("outputs")))
print("Executed.")We create a masks to isolate a selected area and apply inpainting to switch solely that a part of the picture. We refine the chosen space utilizing a focused immediate whereas retaining the remaining intact. Lastly, we save all intermediate and remaining outputs to disk for inspection and reuse.
In conclusion, we demonstrated how a single Diffusers pipeline can evolve into a versatile, production-ready picture era system. We defined find out how to transfer from pure text-to-image era to quick sampling, structural management, and focused picture modifying with out altering frameworks or tooling. This tutorial highlights how we are able to mix schedulers, LoRA adapters, ControlNet, and inpainting to create controllable and environment friendly generative pipelines which can be simple to increase for extra superior artistic or utilized use instances.
Try the Full Codes right here. Additionally, be happy to comply with us on Twitter and don’t overlook to hitch our 100k+ ML SubReddit and Subscribe to our Publication. Wait! are you on telegram? now you’ll be able to be a part of us on telegram as nicely.
