+ "description": "A relatively simple model that tries to remove (muddied) dot crawl, color bleed, NTSC color encoding (as in color space), and MPEG-4 (DivX) artifacts from 384p sources. Trained for 30 epochs.\nOriginally trained to de-crustify Azumanga Web Daioh at 384p. Specifically trained for the hardsub, not the ultra-low-quality nicovideo rip.\nMovie was chosen for the dataset due to having a similar art style and color range to the web pilot (minus the high contrast from the ntsc source), compared to the more washed-out TV anime.\n\n**Dataset:** Azumanga Daioh Movie (2001) cropped to 4:3, played through composite via a DVD player, chroma blurred, and encoded to MPEG-4 (DIVX) at 512x384, bitrate ~1400kbps\nVHS sources will perform poorly (see examples)\n\nSlowpoke: slow.pics/s/qf7gnU9O",
0 commit comments