For instructions on how to use this lora, check the enclosed instructions booklet*.

This is a lora trained on 257 screencaps from both Legend of Zelda CD-i games! I found clear footage online that doesn't have any considerable artifacts or compression, and used that to make a completely new dataset. This new version has far cleaner results, and can be tweaked with quality tags (more importantly, the absence of them! See notes). It was trained on vpred 1.0 noobai for 6000 steps. This was really fun to make 🍻

Notes*:

• I made sure to tag a number of important characters, namely: Link, Zelda, King Harkinian, Ganon, and Gwonam. HOWEVER... I tagged more characters, but they had so few images that they have only partial results: Duke Onkled (accidentally mistagged him as ganon... oops...), Impa, Lord Kiro, Morshu, and Militron (the blue knight... I just thought he looked cool :P).

• Quality Tags Matter! If you use the usual "masterpiece, best quality", the resolution will increase, if you use "worst quality, bad quality" in negative, the general composition can change. These can change the image considerably, so consider playing with and without them! If you want quality, use them, but if you want that 2007 YTP-tier quality, then consider removing them and not upscaling the pic at all lol 🔥

• "oekaki" and "flat colors" aren't necessary, but can help keep the jaggedy linework in check.

• All the images trained are technically landscape (1920x1080), so those sideways resolutions might make better images

Older Description:

This is a lora trained on 121 images taken from both the wonderful licensed Legend of Zelda games on the Phillip's CD-i. It was trained for 10,000 steps at 1024x1024 resolution. I upscaled the dataset (nearest neighbor) to account for how old and low res these games are lol.

It was trained on a bunch of screenshots from the games, and has tag data for each character, although it really only understands characters that get a lot of screentime (namely Zelda, Link, King Harkinian, Ganon, and Gwonam). It can do characters, although it was really meant more as a style lora.

NOTES*:

• It can do non-zelda characters as shown in the example images, but the style 'accuracy' is a bit reduced

• I found adding "2D, flat shading, flat color, jaggy lines" can help a decent bit with keeping the style on track. Adding tags like "oekaki" and "pixel art" might make it worse!

• Due to an oversight, there is no "trigger" tag, so simply use the lora and it should activate.

MAJOR SHOUTOUT to tsu, who made the original CD-i lora over a year ago, and graciously uploaded all his meticulous training data which made making this a breeze. All I did was upscale them all. To pass on the torch, I'll upload this training data as well!

Also, see DrIvoRobotnik's CD-i model, which is also quite good!

This peace is what all true warriors strive for!