card-style is the token for versions 1 and 2. The new version 3 was trained from the ground up using the same dataset and has a new token, cardstyleV3. All of the v2 and v3 images are using the v2-1_768-nonema-pruned.ckpt checkpoint. Trained on 140+ images (260+ for v2) from a variety of modern digital artists for card and videos games such as Gwent and Magic the Gathering. All of the artists have a similar style which has let me get something from the 2.1 checkpoint I was previously unable to achieve. I also recommend using the 2.1 trained checkpoint Ultra Cmodel. I suspect it used similar training images or it's just this modern style which everyone seems to be using. I also tested it with the PRMJ checkpoint, the final image sample was made using it. Update: A 1.5 SD version was requested. You can find it here: https://civitai.com/models/9181/cardstyle15-card-style-for-the-15-checkpoint |