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 120s." 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 (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)