Open Source Video Upscalers - Page 2

This application was made to fulfill the assignment for the Data Structures course. The concept of the application is an application to upgrade/enhance image quality. The main theme is queues, we implement circular queues for pooling/storing a list of images to be upscaled. Gui creation is made manually using the tkinter library. For the upscale process itself, it uses the OpenCV library with a model obtained from open source. Checked using vermin. Minimum required versions: 3.6 Incompatible versions: 2.

Convert models from GoldSource engine to the Source engine with AI. This utility converts GoldSource engine models to Source engine models. A feature of this utility is the ability to improve the quality of textures of models using Upscaling, deblurring, and normal map generating. All operations to improve the quality of textures are performed by neural networks. To improve the quality of the texture, it is first Upscaled using RealESRGAN. The user can select scaling factor: x2, x4 or x8. After the Upscaling procedure, the texture is deblured using the Scale-recurrent Network for Deep Image Deblurring. An example of a processed texture is shown in the following image (parameters used: scaling-factor = 4 and deblur iterations = 4) besides upscaling and debluring the utility also generates normal maps for each texture. This is implemented using the DeepBump by HugoTiny model. Examples of normal maps are shown in the following images.

This project generates hand-written digits from the MNIST dataset using WGAN architecture. While debugged, a few lingering issues may remain, if you encounter any please submit them and they will be resolved. The sources and features of the project can be found in the wiki. Before use, one should verify that the network architecture is as-desired. This may be done with the GUI to the right of the number display. The "Default" button resets the network to hard-coded values which I have verified function. The "Reset" button sets the ACTIVE network to whatever architecture is displayed. In order to change which network is displayed, use the "Critic [1] or Generator [0]" checkbox. To begin training the network, press the "Train" button, after which you may use the "Clear" button to reset the average error and average percent correct value textboxes. In order to use the project on an alternative dataset, one must replace the MNIST files, then rewrite IO.FindNextNumber.

Super-resolution using a CNN, based on the work of the DGtal team. First of all, an Nvidia graphics card (neither AMD nor Intel integrated) is highly recommended to parallelize the CNN. You will then need to install CUDA. No CUDA = dozens of times slower. This program will generate "model_epoch_ .pth" files corresponding to the model at epoch n, in a folder saved_model_u t_bs bs_tbs tbs_lr lr, where corresponds to the scale factor, bsthe size of the training batch, tbsthe size of the test batch and lrto the learning rate. Low res images should be located in a "dataset/input" folder, and high res targets in a "dataset/target" folder, where each different quality image has the same name in both folders.

This program is upscaling any image by a factor 2 using an algorithm of cubic interpolation. You may need to install the following libraries to run the program, tqdm, itertools, and OpenCV.

Single Image Super-Resolution Generative Adversarial Network (GAN) which uses the VapourSynth processing framework to handle input and output image data. Transform, Filter, or Enhance your input video, or the VSGAN result with VapourSynth, a Script-based NLE. You can chain models or re-run the model twice-over (or more). Have low VRAM? Don’t worry! The Network will be applied in quadrants of the image to reduce up-front VRAM usage. You can use any RGB video input, including float32 (e.g., RGBS) inputs. Using VapourSynth you can pass a Video directly to VSGAN, without any frame extraction needed. Any edit you make in the VapourSynth script with or without VSGAN can be re-used for any other video. VSGAN is released under the MIT License, ensuring it will stay free, with the ability to be used commercially.

A set of tools for slicing a texture atlas to individual components and merging back. These tools simplify the approach for upscaling a texture atlas. Manually select some tiles in the texture atlas (that are placed in difficult positions) Add them to a json file following this pattern: json file; you can use any image editing tool such as GIMP to manually select and inspect the location, width and height of each selection. Run the tool with the split; this will automatically create a folder with separate files from the selections, as well a [basefilename]_sliced.png file containing the texture without the sliced files. Manually upscale all the resulting files with the correct transparency / seamless mode, including the [basefilename]_sliced.png from the previous step. it is important to use the same scale factor for all the files. you can leave some of the files unscaled (including the basefile).