# Code taken (and slightly adopted) from https://huggingface.co/spaces/havas79/Real-ESRGAN_Demo/blob/main/app.py - credit where credit is due. I am not showcasing code here, but demoing my own trained models ;) import gradio as gr import cv2 import numpy import os import random from basicsr.archs.rrdbnet_arch import RRDBNet from basicsr.utils.download_util import load_file_from_url from realesrgan import RealESRGANer from realesrgan.archs.srvgg_arch import SRVGGNetCompact last_file = None img_mode = "RGBA" def realesrgan(img, model_name, face_enhance): global last_file # remove last upscale when doing this new upscale to prevent memory being full if last_file: print(f"Deleting {last_file} ...") os.remove(last_file) last_file = None if not img: return imgwidth, imgheight = img.size if imgwidth > 1000 or imgheight > 1000: return error("Input Image too big") # Define model parameters if model_name == '4xNomos8kSC': model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4) netscale = 4 elif model_name == '4xHFA2k': model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4) netscale = 4 elif model_name == '4xLSDIR': model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4) netscale = 4 elif model_name == '4xLSDIRplusN': model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4) netscale = 4 elif model_name == '4xLSDIRplusC': model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4) netscale = 4 elif model_name == '4xLSDIRplusR': model = RRDBNet(num_in_ch=3, num_out_ch=3, num_feat=64, num_block=23, num_grow_ch=32, scale=4) netscale = 4 elif model_name == '2xParimgCompact': model = SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=16, upscale=2, act_type='prelu') netscale = 2 elif model_name == '2xHFA2kCompact': model = SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=16, upscale=2, act_type='prelu') netscale = 2 elif model_name == '4xLSDIRCompactN': model = SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=16, upscale=4, act_type='prelu') netscale = 4 elif model_name == '4xLSDIRCompactC3': model = SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=16, upscale=4, act_type='prelu') netscale = 4 elif model_name == '4xLSDIRCompactR3': model = SRVGGNetCompact(num_in_ch=3, num_out_ch=3, num_feat=64, num_conv=16, upscale=4, act_type='prelu') netscale = 4 # Determine model paths model_path = os.path.join('weights', model_name + '.pth') # Restorer Class upsampler = RealESRGANer( scale=netscale, model_path=model_path, dni_weight=None, model=model, tile=128, tile_pad=10, pre_pad=10, half=False, gpu_id=None, ) # Use GFPGAN for face enhancement if face_enhance: from gfpgan import GFPGANer face_enhancer = GFPGANer( model_path='https://github.com/TencentARC/GFPGAN/releases/download/v1.3.0/GFPGANv1.4.pth', upscale=netscale, arch='clean', channel_multiplier=2, bg_upsampler=upsampler) # Convert the input PIL image to cv2 image, so that it can be processed by realesrgan cv_img = numpy.array(img) img = cv2.cvtColor(cv_img, cv2.COLOR_RGBA2BGRA) # Apply restoration try: if face_enhance: _, _, output = face_enhancer.enhance(img, has_aligned=False, only_center_face=False, paste_back=True) else: output, _ = upsampler.enhance(img, netscale) except RuntimeError as error: print('Error', error) print('If you encounter CUDA out of memory, try to set --tile with a smaller number.') else: # Save restored image and return it to the output Image component extension = 'jpg' out_filename = f"output_{rnd_string(16)}.{extension}" cv2.imwrite(out_filename, output) last_file = out_filename return out_filename def rnd_string(x): """Returns a string of 'x' random characters """ characters = "abcdefghijklmnopqrstuvwxyz_0123456789" result = "".join((random.choice(characters)) for i in range(x)) return result def reset(): """Resets the Image components of the Gradio interface and deletes the last processed image """ global last_file if last_file: print(f"Deleting {last_file} ...") os.remove(last_file) last_file = None return gr.update(value=None), gr.update(value=None) def has_transparency(img): """This function works by first checking to see if a "transparency" property is defined in the image's info -- if so, we return "True". Then, if the image is using indexed colors (such as in GIFs), it gets the index of the transparent color in the palette (img.info.get("transparency", -1)) and checks if it's used anywhere in the canvas (img.getcolors()). If the image is in RGBA mode, then presumably it has transparency in it, but it double-checks by getting the minimum and maximum values of every color channel (img.getextrema()), and checks if the alpha channel's smallest value falls below 255. https://stackoverflow.com/questions/43864101/python-pil-check-if-image-is-transparent """ if img.info.get("transparency", None) is not None: return True if img.mode == "P": transparent = img.info.get("transparency", -1) for _, index in img.getcolors(): if index == transparent: return True elif img.mode == "RGBA": extrema = img.getextrema() if extrema[3][0] < 255: return True return False def image_properties(img): """Returns the dimensions (width and height) and color mode of the input image and also sets the global img_mode variable to be used by the realesrgan function """ global img_mode if img: if has_transparency(img): img_mode = "RGBA" else: img_mode = "RGB" properties = f"Width: {img.size[0]}, Height: {img.size[1]} | Color Mode: {img_mode}" return properties def main(): # Gradio Interface with gr.Blocks(title="Self-trained ESRGAN models demo", theme="dark") as demo: gr.Markdown( """#
Upscale image
Here I demo some of my self-trained models (only those trained on the SRVGGNet or RRDBNet archs). All my self-trained models can be found on the [openmodeldb](https://openmodeldb.info/?q=Helaman&sort=date-desc) or on [my github repo](https://github.com/phhofm/models). """ ) with gr.Group(): with gr.Group(): model_name = gr.Dropdown(label="Model to be used", choices=["2xHFA2kCompact", "2xParimgCompact", "4xLSDIRCompactN", "4xLSDIRCompactC3", "4xLSDIRCompactR3", "4xNomos8kSC", "4xHFA2k", "4xLSDIR", "4xLSDIRplusN", "4xLSDIRplusC", "4xLSDIRplusR"], value="4xLSDIRCompactC3", info="See model infos at the bottom of this page") face_enhance = gr.Checkbox(label="Face Enhancement using GFPGAN (Doesn't work for anime images)",value=False, show_label=True) with gr.Group(): input_image = gr.Image(label="Source Image", type="pil", image_mode="RGB") input_image_properties = gr.Textbox(label="Image Properties - Demo will throw error if input image has either width or height > 1000. Output download is jpg for smaller size. Use models locally to circument these limits.", max_lines=1) with gr.Group(): output_image = gr.Image(label="Upscaled Image", type="pil", image_mode="RGB", interactive=False) output_image_properties = gr.Textbox(label="Image Properties", max_lines=1) with gr.Row(): upscale_btn = gr.Button("Upscale") reset_btn = gr.Button("Reset") with gr.Group(): gr.Markdown(""" **Examples are not pre-cached. You need to press the Upscale Button after selecting one**""") gr.Examples(examples="examples",inputs=[input_image, model_name, face_enhance],outputs=output_image,fn=realesrgan, cache_examples=False) gr.Markdown( """ **Model infos** *SRVGGNetCompact models - in general faster, but less powerful, than RRDBNet* 2xHFA2kCompact - use for upscaling anime images 2x, faster than 4xHFA2k but less powerful (SRVGGNetCompact) 2xParimgCompact - upscaling photos 2x, fast (SRVGGNetCompact) 4xLSDIRCompactN - upscale a good quality photo (no degradations) 4x, faster than 4xLSDIRN but less powerful (SRVGGNetCompact) 4xLSDIRCompactC3 - upscale a jpg compressed photo 4x, fast (SRVGGNetCompact) 4xLSDIRCompactR3 - upscale a degraded photo 4x, fast (SRVGGNetCompact) (too strong, best used for interpolation like 4xLSDIRCompactN (or C) 75% 4xLSDIRCompactR3 25% to add little degradation handling to the previous one) *RRDBNet models - in general more powerful than SRVGGNetCompact, but very slow in this demo* 4xNomos8kSC - use for upscaling photos 4x or can also be tried out on anime 4xHFA2k - use for upscaling anime images 4x 4xLSDIR - upscale a good quality photo (no degradation) 4x 4xLSDIRplusN - upscale a good quality photo (no degradation) 4x 4xLSDIRplusC - upscale a jpg compressed photo 4x 4xLSDIRplusR - upscale a degraded photo 4x (too strong, best used for interpolation like 4xLSDIRplusN (or C) 75% 4xLSDIRplusR 25% to add little degradation handling to the previous one) *Models that I trained that are not featured here, but available on [openmodeldb](https://openmodeldb.info/?q=Helaman&sort=date-desc) or on [github](https://github.com/phhofm/models):* 4xNomos8kSCHAT-L - Photo upscaler (handles little bit of jpg compression and blur), [HAT-L](https://github.com/XPixelGroup/HAT) model (good output but very slow since huge) 4xNomos8kSCHAT-S - Photo upscaler (handles little bit of jpg compression and blur), [HAT-S](https://github.com/XPixelGroup/HAT) model 4xNomos8kSCSRFormer - Photo upscaler (handles little bit of jpg compression and blur), [SRFormer](https://github.com/HVision-NKU/SRFormer) base model (also good and slow since also big model) 2xHFA2kAVCOmniSR - Anime frame upscaler that handles AVC (h264) video compression, [OmniSR](https://github.com/Francis0625/Omni-SR) model 2xHFA2kAVCOmniSR_Sharp - Anime frame upscaler that handles AVC (h264) video compression with sharper outputs, [OmniSR](https://github.com/Francis0625/Omni-SR) model 4xHFA2kAVCSRFormer_light - Anime frame upscaler that handles AVC (h264) video compression, [SRFormer](https://github.com/HVision-NKU/SRFormer) lightweight model 2xHFA2kAVCEDSR_M - Anime frame upscaler that handles AVC (h264) video compression, [EDSR-M](https://github.com/LimBee/NTIRE2017) model 2xHFA2kAVCCompact - Anime frame upscaler that handles AVC (h264) video compression, [SRVGGNet](https://github.com/xinntao/Real-ESRGAN) (also called Real-ESRGAN Compact) model 4xHFA2kLUDVAESwinIR_light - Anime image upscaler that handles various realistic degradations, [SwinIR](https://github.com/JingyunLiang/SwinIR) light model 4xHFA2kLUDVAEGRL_small - Anime image upscaler that handles various realistic degradations, [GRL](https://github.com/ofsoundof/GRL-Image-Restoration) small model 4xHFA2kLUDVAESRFormer_light - Anime image upscaler that handles various realistic degradations, [SRFormer](https://github.com/HVision-NKU/SRFormer) light model 4xLexicaHAT - An AI generated image upscaler, does not handle any degradations, [HAT](https://github.com/XPixelGroup/HAT) base model 2xLexicaSwinIR - An AI generated image upscaler, does not handle any degradations, [SwinIR](https://github.com/JingyunLiang/SwinIR) base model 2xLexicaRRDBNet - An AI generated image upscaler, does not handle any degradations, RRDBNet base model 2xLexicaRRDBNet_Sharp - An AI generated image upscaler with sharper outputs, does not handle any degradations, RRDBNet base model 4xHFA2kLUDVAESAFMN - dropped model since there were artifacts on the outputs when training with [SAFMN](https://github.com/sunny2109/SAFMN) arch *The following are not models I had trained, but rather interpolations I had created, they are available on my [repo](https://github.com/phhofm/models) and can be tried out locally with chaiNNer:* 4xLSDIRplus (4xLSDIRplusC + 4xLSDIRplusR) 4xLSDIRCompact3 (4xLSDIRCompactC3 + 4xLSDIRCompactR3) 4xLSDIRCompact2 (4xLSDIRCompactC2 + 4xLSDIRCompactR2) 4xInt-Ultracri (UltraSharp + Remacri) 4xInt-Superscri (Superscale + Remacri) 4xInt-Siacri(Siax + Remacri) 4xInt-RemDF2K (Remacri + RealSR_DF2K_JPEG) 4xInt-RemArt (Remacri + VolArt) 4xInt-RemAnime (Remacri + AnimeSharp) 4xInt-RemacRestore (Remacri + UltraMix_Restore) 4xInt-AnimeArt (AnimeSharp + VolArt) 2xInt-LD-AnimeJaNai (LD-Anime + AnimeJaNai) """) # Event listeners: input_image.change(fn=image_properties, inputs=input_image, outputs=input_image_properties) output_image.change(fn=image_properties, inputs=output_image, outputs=output_image_properties) upscale_btn.click(fn=realesrgan, inputs=[input_image, model_name, face_enhance], outputs=output_image) reset_btn.click(fn=reset, inputs=[], outputs=[output_image, input_image]) demo.launch() if __name__ == "__main__": main()