【Real-ESRGAN】Stable Diffusionで4K画像を生成する方法【Python】

from diffusers import StableDiffusionXLPipeline
import torch
import time

model_id = "stabilityai/stable-diffusion-xl-base-1.0"

pipe = StableDiffusionXLPipeline.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
).to("cuda")

pipe.enable_attention_slicing()
pipe.enable_vae_slicing()

prompt = "cinematic photo of a futuristic tokyo street at night, neon lights, ultra detailed"

#解像度指定（4K）
width, height = 3840, 2160

#seed固定
seed = 1234
generator = torch.Generator(device="cuda").manual_seed(seed)

#処理時間計測開始
start_time = time.time()

image = pipe(
    prompt,
    num_inference_steps=30,
    guidance_scale=6.5,
    width=width,
    height=height,
    generator=generator,
).images[0]

#処理時間計測終了
end_time = time.time()
elapsed = end_time - start_time

image.save("sdxl_4k.png")

print("saved: sdxl_4k.png")
print(f"seed: {seed}")
print(f"time: {elapsed:.2f} sec")

from diffusers import StableDiffusionXLPipeline
import torch
import time
from PIL import Image

model_id = "stabilityai/stable-diffusion-xl-base-1.0"

pipe = StableDiffusionXLPipeline.from_pretrained(
    model_id,
    torch_dtype=torch.float16,
).to("cuda")

pipe.enable_attention_slicing()
pipe.enable_vae_slicing()

prompt = "cinematic photo of a futuristic tokyo street at night, neon lights, ultra detailed"

#解像度指定(1024pxに近い解像度)
width, height = 1536, 864

#Resize後の解像度(4K)
width_4K, height_4K = 3840, 2160

#seed固定
seed = 1234
generator = torch.Generator(device="cuda").manual_seed(seed)

#処理時間計測開始
start_time = time.time()

image = pipe(
    prompt,
    num_inference_steps=30,
    guidance_scale=6.5,
    width=width,
    height=height,
    generator=generator,
).images[0]

#Resize
image_4K = image.resize(
    (width_4K, height_4K),
    resample=Image.LANCZOS  # 高品質リサイズ(LANCZOS補間)
)

#処理時間計測終了
end_time = time.time()
elapsed = end_time - start_time

image_4K.save("sdxl_4K_resize.png")

print("saved: sdxl_4K_resize.png")
print(f"seed: {seed}")
print(f"time: {elapsed:.2f} sec")

import os
import time
from urllib.request import urlretrieve

import cv2
import torch
from PIL import Image
from diffusers import StableDiffusionXLPipeline
from realesrgan import RealESRGANer
from basicsr.archs.rrdbnet_arch import RRDBNet


# =========================================
# 設定
# =========================================
sd_model_id = "stabilityai/stable-diffusion-xl-base-1.0"

prompt = "cinematic photo of a futuristic tokyo street at night, neon lights, ultra detailed"

# SDXL生成サイズ（最初に大きめで生成）
width, height = 1536, 864

# 最終出力サイズ（4K）
final_width, final_height = 3840, 2160

# seed固定
seed = 1234

# Real-ESRGAN設定
realesrgan_model_name = "RealESRGAN_x4plus.pth"
realesrgan_model_url = "https://github.com/xinntao/Real-ESRGAN/releases/download/v0.1.0/RealESRGAN_x4plus.pth"
realesrgan_scale = 4

# ファイル名
sd_output_path = "sdxl.png"
esrgan_output_path = "sdxl_x4.png"
final_output_path = "sdxl_4k_realesrgan.png"


# =========================================
# デバイス設定
# =========================================
device = "cuda" if torch.cuda.is_available() else "cpu"
print("device:", device)

# =========================================
# Stable Diffusion XL パイプライン
# =========================================
print("loading SDXL pipeline...")
pipe = StableDiffusionXLPipeline.from_pretrained(
    sd_model_id,
    torch_dtype=torch.float16 if device == "cuda" else torch.float32,
)
pipe = pipe.to(device)

# メモリ節約
pipe.enable_attention_slicing()
pipe.enable_vae_slicing()

generator = torch.Generator(device=device).manual_seed(seed)


# =========================================
# Real-ESRGAN モデル準備
# =========================================
if not os.path.exists(realesrgan_model_name):
    print("downloading Real-ESRGAN model...")
    urlretrieve(realesrgan_model_url, realesrgan_model_name)
    print("downloaded:", realesrgan_model_name)

print("building Real-ESRGAN model...")
esrgan_model = RRDBNet(
    num_in_ch=3,
    num_out_ch=3,
    num_feat=64,
    num_block=23,
    num_grow_ch=32,
    scale=realesrgan_scale
)

upsampler = RealESRGANer(
    scale=realesrgan_scale,
    model_path=realesrgan_model_name,
    model=esrgan_model,
    tile=0,
    tile_pad=10,
    pre_pad=0,
    half=(device == "cuda"),
    device=device,
)


# =========================================
# 1. SDXLで画像生成
# =========================================
print("generating image with SDXL...")
start_total = time.time()
start_sd = time.time()

image = pipe(
    prompt=prompt,
    num_inference_steps=30,
    guidance_scale=6.5,
    width=width,
    height=height,
    generator=generator,
).images[0]

if device == "cuda":
    torch.cuda.synchronize()

sd_elapsed = time.time() - start_sd

image.save(sd_output_path)
print("saved:", sd_output_path)
print("SDXL output size:", image.size)


# =========================================
# 2. Real-ESRGANでアップスケール
# =========================================
print("upscaling with Real-ESRGAN...")
img = cv2.imread(sd_output_path, cv2.IMREAD_COLOR)
if img is None:
    raise ValueError(f"failed to read image: {sd_output_path}")

start_esrgan = time.time()

output, _ = upsampler.enhance(img, outscale=realesrgan_scale)

if device == "cuda":
    torch.cuda.synchronize()

esrgan_elapsed = time.time() - start_esrgan

cv2.imwrite(esrgan_output_path, output)
print("saved:", esrgan_output_path)
print("Real-ESRGAN output size:", output.shape[1], "x", output.shape[0])


# =========================================
# 3. 4Kへ高品質縮小
# =========================================
print("resizing to exact 4K...")
upscaled_pil = Image.open(esrgan_output_path).convert("RGB")

final_image = upscaled_pil.resize(
    (final_width, final_height),
    resample=Image.LANCZOS
)

final_image.save(final_output_path)

total_elapsed = time.time() - start_total

print("saved:", final_output_path)
print(f"seed: {seed}")
print(f"SDXL time: {sd_elapsed:.2f} sec")
print(f"Real-ESRGAN time: {esrgan_elapsed:.2f} sec")
print(f"total time: {total_elapsed:.2f} sec")