app.py

import os
import sys
import gradio as gr
from PIL import Image

## environment settup
os.system("git clone https://github.com/codeslake/RefVSR.git")
os.chdir("RefVSR")
os.system("./install/install_cudnn113.sh")
os.mkdir("ckpt")
os.system("wget https://huggingface.co/spaces/codeslake/RefVSR/resolve/main/SPyNet.pytorch -O ckpt/SPyNet.pytorch")

os.system("wget https://huggingface.co/spaces/codeslake/RefVSR/resolve/main/RefVSR_MFID_8K.pytorch -O ckpt/RefVSR_MFID_8K.pytorch")
os.system("wget https://huggingface.co/spaces/codeslake/RefVSR/resolve/main/RefVSR_small_MFID_8K.pytorch -O ckpt/RefVSR_small_MFID_8K.pytorch")

os.system("wget https://huggingface.co/spaces/codeslake/RefVSR/resolve/main/RefVSR_MFID.pytorch -O ckpt/RefVSR_MFID.pytorch")
os.system("wget https://huggingface.co/spaces/codeslake/RefVSR/resolve/main/RefVSR_small_MFID_8K.pytorch -O ckpt/RefVSR_small_MFID.pytorch")

sys.path.append("RefVSR")

## I/O setup (creates folders and places inputs corresponding to the original RefVSR code)
# HD input
HR_LR_path = "test/RealMCVSR/test/HR/UW/0000"
HR_Ref_path = "test/RealMCVSR/test/HR/W/0000"
HR_Ref_path_T = "test/RealMCVSR/test/HR/T/0000"
os.makedirs(HR_LR_path)
os.makedirs(HR_Ref_path)
os.makedirs(HR_Ref_path_T)
os.system("wget https://www.dropbox.com/s/x33ka2jlzwsde7r/LR.png -O HR_LR1.png")
os.system("wget https://www.dropbox.com/s/pp903wlz3syf68w/Ref.png -O HR_Ref1.png")
os.system("wget https://www.dropbox.com/s/zl0h83x0le6ejfw/LR.png -O HR_LR2.png")
os.system("wget https://www.dropbox.com/s/9hzupmc3clt0f0e/Ref.png -O HR_Ref2.png")
os.system("wget https://www.dropbox.com/s/2u6lcfdhvcylklg/LR.png -O HR_LR3.png")
os.system("wget https://www.dropbox.com/s/a7bwfy3gl26tvbq/Ref.png -O HR_Ref3.png")

# 4x downsampled input
LR_path = "test/RealMCVSR/test/LRx4/UW/0000"
Ref_path = "test/RealMCVSR/test/LRx4/W/0000"
Ref_path_T = "test/RealMCVSR/test/LRx4/T/0000"
os.makedirs(LR_path)
os.makedirs(Ref_path)
os.makedirs(Ref_path_T)
os.system("wget https://www.dropbox.com/s/hkvdwm3grshjt0k/LR.png -O LR.png")
os.system("wget https://www.dropbox.com/s/4sv34su3kg1ifkp/Ref.png -O Ref.png")

# output directory
os.makedirs('result')

## resize if necessary (not used)
def resize(img):
    max_side = 480
    w = img.size[0]
    h = img.size[1]
    if max(h, w) > max_side:
        scale_ratio = max_side / max(h, w)
        wsize=int(w*scale_ratio)
        hsize=int(h*scale_ratio)
        img = img.resize((wsize,hsize), Image.ANTIALIAS)
    w = img.size[0]
    h = img.size[1]
    img = img.crop((0, 0, w-w%8, h-h%8))

    return img
    

#################### 8K ##################
## inference  
def inference_8K(LR, Ref):
    ## resize for user selected input (not used)
    LR = resize(LR)
    Ref = resize(Ref)
    
    ## Input setup (creates folders and places inputs corresponding to the original RefVSR code)
    LR.save(os.path.join(LR_path, '0000.png'))
    Ref.save(os.path.join(Ref_path, '0000.png'))
    Ref.save(os.path.join(Ref_path_T, '0000.png'))
    LR.save(os.path.join(HR_LR_path, '0000.png'))
    Ref.save(os.path.join(HR_Ref_path, '0000.png'))
    Ref.save(os.path.join(HR_Ref_path_T, '0000.png'))
    
    ## Run RefVSR model
    os.system("python -B run.py \
                --mode RefVSR_MFID_8K \
                --config config_RefVSR_MFID_8K \
                --data RealMCVSR \
                --ckpt_abs_name ckpt/RefVSR_MFID_8K.pytorch \
                --data_offset ./test \
                --output_offset ./result \
                --qualitative_only \
                --cpu \
                --is_gradio")               
    return "result/0000.png"
    
title="RefVSR"
description="Demo application for Reference-based Video Super-Resolution (RefVSR). Upload a low-resolution frame and a reference frame to 'LR' and 'Ref' input windows, respectively. The demo runs on CPUs and takes about 30s."

article = "<p style='text-align: center'><b>To check the full capability of the module, we recommend to clone Github repository and run RefVSR models on videos using GPUs.</b></p><p style='text-align: center'>This demo runs on CPUs and only supports RefVSR for a single LR and Ref frames due to computational complexity.<br>Hence, the model <b>will not take advantage</b> of temporal LR and Ref frames.</p><p style='text-align: center'>Moreover, the model is trained <b>with the proposed 2-stage training strategy</b>, but due to the memory and computational complexity, we downsampled sample frames to have the 480x270 resolution.</p><p style='text-align: center'>For user given frames, the size will be adjusted for the longer side of the frames to have 480 pixels.</p><p style='text-align: center'><a href='https://junyonglee.me/projects/RefVSR' target='_blank'>Project</a> | <a href='https://arxiv.org/abs/2203.14537' target='_blank'>arXiv</a> | <a href='https://github.com/codeslake/RefVSR' target='_blank'>Github</a></p>"

## resize for sample (not used)
#LR = resize(Image.open('LR.png')).save('LR.png')
#Ref = resize(Image.open('Ref.png')).save('Ref.png')

## input
examples=[['HR_LR1.png', 'HR_Ref1.png'], ['HR_LR2.png', 'HR_Ref2.png'], ['HR_LR3.png', 'HR_Ref3.png']]

## interface
gr.Interface(inference_8K,[gr.inputs.Image(type="pil"), gr.inputs.Image(type="pil")],gr.outputs.Image(type="file"),title=title,description=description,article=article,theme ="peach",examples=examples).launch(enable_queue=True)    

#################### low res ##################
## inference  
def inference(LR, Ref):
    ## resize for user selected input
    LR = resize(LR)
    Ref = resize(Ref)
    
    ## Input setup (creates folders and places inputs corresponding to the original RefVSR code)
    LR.save(os.path.join(LR_path, '0000.png'))
    Ref.save(os.path.join(Ref_path, '0000.png'))
    Ref.save(os.path.join(Ref_path_T, '0000.png'))
    LR.save(os.path.join(HR_LR_path, '0000.png'))
    Ref.save(os.path.join(HR_Ref_path, '0000.png'))
    Ref.save(os.path.join(HR_Ref_path_T, '0000.png'))
    
    ## Run RefVSR model
    os.system("python -B run.py \
                --mode RefVSR_MFID \
                --config config_RefVSR_MFID \
                --data RealMCVSR \
                --ckpt_abs_name ckpt/RefVSR_MFID.pytorch \
                --data_offset ./test \
                --output_offset ./result \
                --qualitative_only \
                --cpu \
                --is_gradio")             
    return "result/0000.png"

title="Demo for RefVSR (CVPR 2022)"
description="The demo applies 4xVSR on a video frame. It runs on CPUs and takes about 150s. For the demo, upload a low-resolution frame and a reference frame to 'LR' and 'Ref' input windows, respectively. It is recommended for the reference frame to have a 2x larger zoom factor than that of the low-resolution frame."

article = "<p style='text-align: center'><b>To check the full capability of the module, we recommend to clone Github repository and run RefVSR models on videos using GPUs.</b></p><p style='text-align: center'>This demo runs on CPUs and only supports RefVSR for a single LR and Ref frames due to computational complexity.<br>Hence, the model <b>will not take advantage</b> of temporal LR and Ref frames.</p><p style='text-align: center'>Moreover, the model is trained <b>only with the proposed pre-training strategy</b> to cope with downsampled sample frames, which are in the 480x270 resolution.</p><p style='text-align: center'>For user given frames, the size will be adjusted for the longer side of the frames to have 480 pixels.</p><p style='text-align: center'><a href='https://junyonglee.me/projects/RefVSR' target='_blank'>Project</a> | <a href='https://arxiv.org/abs/2203.14537' target='_blank'>arXiv</a> | <a href='https://github.com/codeslake/RefVSR' target='_blank'>Github</a></p>"

## resize for sample
LR = resize(Image.open('LR.png')).save('LR.png')
Ref = resize(Image.open('Ref.png')).save('Ref.png')

## input
examples=[['LR.png','Ref.png']]

## interface
gr.Interface(inference, [gr.inputs.Image(type="pil"), gr.inputs.Image(type="pil")], gr.outputs.Image(type="file"),title=title,description=description,article=article,theme ="peach",examples=examples).launch(enable_queue=True)