Trained on anything v3.1 with ~2.3k images from old japanese games (mostly pc-98, but also pc-88, pc engine and X68000). This lora has a heavy female bias due to the training data (meaning your 1boys will probably not look very masculine! You can try tags such as man instead to counterbalance this somewhat.)

The training tags were stripped of extra "quality control" tags like dithering, jaggy lines, oekaki, retro style, etc. You can add them to your prompt or negative prompt, but mostly the result hinges on including the keyword mentioned above. Also works with both "plaintext" and underscored tags (Ex: dutch angle vs dutch_angle - trained on both just in case).

Now with 3 levels!

Strong - Very in line with the pc-98 concept. It has Dithering, Jaggy lines, etc that emulate the style as much as possible. (However, be aware it IS strong so some experimentation might be needed when it comes to weighting it.) Recommended weight: 0.7-1.

Medium - still in line with the concept, albeit it eases up on the artifacts like dithering, jaggy lines, etc. It's style falls roughly inbetween the light and strong versions. Some tinkering might still be required but overall works "easier" and doesn't require overt wrestling to get alright results. Recommended weight: 0.8-1.

Light - Since this is quite a light version of the original, it'd describe it's style more like a "retro look" than specifically the "pc-98 look". It retains much of the retro style, but things like dithering, jaggy lines, etc are lacking from it. should be pretty easy to handle and weight. Recommended weight: 0.9-1.

描述:

This version is the strongest (lighter versions coming soon). There may be some issues with hairstyles (very large hair intakes, oversized bangs, etc) - but most are the result of the strong retro styling these games had.

In addition, there are 2 things you should also keep in mind.

1. Be vary that the stronger the cfg, the stronger it will emulate the older style. That may potentially cause your image to become overtly "low res". If you encounter such an issue, I'd recommend running some xzy to determine the Lora weight and cfg you will need for your desired result.

2. Since much of the training data was fed with old/retro images, a sort of "soft censor" was produced as a result (ergo, even if you prompt "nude", you still might need to add additional tokens regarding genitalia and areolae to get the desired result).

训练词语: pc98