post_process/upscale/super_resolution.py

import torch
import PIL
import numpy as np
# from post_process.upscale.RealESRGAN import RealESRGAN
from diffusers import StableDiffusionPipeline, StableDiffusionImg2ImgPipeline

class Upscaler():
    def __init__(self, text2img: StableDiffusionPipeline, realesrganer):
        self.device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
        # self.scale = scale
        # self.model = RealESRGAN(self.device, scale=self.scale)
        # self.model.load_weights(f'CodeFormer/CodeFormer/weights/realesrgan/RealESRGAN_x{self.scale}.pth')
        self.model = realesrganer
        self.img2img = StableDiffusionImg2ImgPipeline(**text2img.components).to(self.device)
        
    def upscale(self, imgs: list[PIL.Image]) -> list[PIL.Image]:
        # torch.cuda.empty_cache()
        upscaled_imgs = []
        for img in imgs:
            upscaled_img = self.model.predict(img)
            upscaled_imgs.append(upscaled_img)
        return upscaled_imgs

    def hires_fix(self, imgs: list[PIL.Image], prompt: str, negative_prompt: str) -> list[PIL.Image]:
        upscaled_images = self.upscale(imgs)
        results = self.img2img(
            prompt=[prompt]*len(imgs), 
            negative_prompt=[negative_prompt]*len(imgs),
            image=upscaled_images, 
            strength=0.2, 
            guidance_scale=7.5, 
            num_inference_steps=20, 
            schduler="EulerAncestralDiscreteScheduler",
        ).images
        return results