| @News @Wiki Editor | |
| **Name:** 4x-UniScaleV2_Soft/Moderate/Sharp | |
| **Author:** Kim2091 | |
| **License:** CC BY-NC-SA 4.0 | |
| **Link:** <https://mega.nz/folder/zZhA1KoD#ds2nmgDNV4hfFpNSfEqN6Q> | |
| **Model Architecture:** ESRGAN | |
| **Scale:** 4 | |
| **Purpose:** Effectively a UniScale successor. It does nearly everything better, other than dealing with noise or compression. Use UniScale_Restore or UniScale_Iterp for that. | |
| **Iterations:** 111k | |
| **batch_size:** 4 | |
| **HR_size:** 112 | |
| **Epoch:** 8 | |
| **Dataset:** Custom Dataset consisting of lossless 4k frames from Metal Arms: Glitch in the System, Just Cause 3, Dirt 3, Forza Horizon 3, Sleeping Dogs, and self-edited photos from SignatureEdits + ATLA DVD images | |
| **Dataset_size:** 18,909 tiles + 288 ATLA Frames | |
| **OTF Training** Yes (JPEG artifacts, base_blur, bsrgan_resize) | |
| **Pretrained_Model_G:** 4xESRGAN | |
| **Description:** I really don't have a great description for this model set. It just works well on nearly everything (other than Anime sadly). They work best with realistic images. I hope you guys like it :stuck_out_tongue: | |
| The Moderate and Soft models are interpolations between UniScaleV2_Sharp, UniScale-Strong, and UniScaleNR-Balanced. | |
| If your image has compression, the Soft model will work the best. Moderate and Sharp will work on images with compression, but they won't be as clean. | |
| UniScale_Restore or UniScale_Interp (available in the main UniScale folder) will likely work better for images with heavy compression. | |
| Comparisons: | |
| https://cdn.discordapp.com/attachments/884239326471393331/884296266857713704/unknown.png | |
| https://cdn.discordapp.com/attachments/884239326471393331/884294468709257216/unknown.png | |
| https://cdn.discordapp.com/attachments/884239326471393331/884298572894441512/unknown.png | |
| https://cdn.discordapp.com/attachments/884239326471393331/884302370136289323/unknown.png |