It generates less moiré like patterns and keeps details without oversharpening or blurring the image too much. It can upscale most pictures/photos (granted they are clean enough) without destroying as much detail as the aforementioned models. This model aims to improve further on what has been achieved by the regular 4圎SRGAN and also 4xBox. RRDB_ESRGAN_x4 replacement for stuff that's supposed to look realistic. I encourage everybody to mirror the model, distribute and modify it in anywway you want.
#Avs video editor 6.1 download#
The config file is provided on the download link above. Meant as an experiment to test latest techniques implemented on traiNNer, including: AdaTarget, KernelGAN, UNet discriminator, nESRGAN+ arch, noise patches, camera noise, isotropic/anisotropic/sinc blur, frequency separation, contextual loss, mixup, clipL1 pixel loss, AdamP optimizer, etc. Universal upscaler for clean and slightly compressed images (JPEG quality 75 or better) Trained using the patchgan discriminator, with cx loss, cutmixup and frequency separation, it produces good results with a slight grain due to patchgan, with some sharpening using cutmixup. Colors are pretty good as well as edges, but generated details seem slightly fuzzy hence the name.Ī universal model, that is aimed at prerendered images, but handles realistic faces, manga, pixel art and dedithering as well. Photographs, Artwork, Textures, Anything really - Tried out a new pixel loss idea based on ensuring the HR downscaled matches the LR.
The Misc model is trained on various pictures shot by myself ( Alsa), including bricks, stone, dirt, grass, plants, wood, bark, metal and a few others. NOTE: THIS WILL NOT WORK IN CUPSCALE, IEU, OR CHAINNER (yet)! You have to use my fork to use it for now. Basically it makes smaller ESRGAN models that theoretically can produce the same level of quality. You can read more about it in the github README.
Description: Pretrained model for the new architecture modification I made. Technically my previous experiment was the pretrained model, but for all intents and purposes this was trained from scratch. Supposed to help retain more details, but unfortunately due to the dataset (I think) still blurs details adjacent to other objects. DO NOT USE FOR COMPRESSED IMAGES, use the original UniScale or UltraSharp for that.īasically realesrgan-x4plus without the degradation training. These models work great on game textures when interpolated 50/50 with UniScale_Restore, and work amazingly on uncompressed images. Trained with BSRGAN_Resize and Combo_Noise in traiNNer. UniScale_Restore has strong compression removal that helps with restoring heavily compressed or noisy images. This model removes noise from images while upscaling. Version of UniScale trained with camera noise injection (NR = Noise Removal). It was originally intended to upscale game textures, but was expanded into a universal upscaler. This model can upscale almost anything well. UniScale strikes a nice balance between sharpness and realism. It's a bunch of interpolated models based around UltraSharp and my other models It has the ability to restore highly compressed images as well! If you want a more balanced output, check out the UltraMix Collection down below. It does work best on JPEG compression though, as that's mostly what it was trained on. It works on most images, whether compressed or not. This is my best model yet! It generates lots and lots of detail and leaves a nice texture on images.
#Avs video editor 6.1 skin#
This was, things like skin and other details don't become mushy and blurry. Outdoor scenes.Ī creation of BSRGAN with more details and less smoothing, made by interpolating IRL models such as Siax, Superscale, Superscale Artisoft, Pixel Perfect, etc. Streets with dense foliage in the background.
Image scaling and Video upscaling Universal Models Model Name These programs can be used to train your models: BasicSR, the official ESRGAN repository (old arch tag), victorca's traiNNer ( ), or sudo's colab-traiNNer ( ). The others are: IEU by Honh Cupscale by NMKD. The only actively maintained program is chaiNNer by Joey. There are various GUIs available to inference/upscale with these models. These are all models that use the "old" ESRGAN architecture. 1.2.12.2 Normal Map/Bump Map Generation.
0 kommentar(er)
