>>99012834Aight long time no update /vtai/, been working on WAN image to video loras to get over my burnout from making loras. If anybody's interested, I made an illust lora for birdman's pixel art style, and a WAN I2V Lora for animating the images gen'd from the illust lora, Safe to say, I think I legitimately cooked
https://files.catbox.moe/ai8d1g.mp4Other news, I finished training the Reimu Endou lora for 4 outfits, and it took me longer than I'd like desu, but you can try it here
https://civitai.com/models/1447966The next vtubers I'm gonna do loras for, are two IdolES girls
1. Taiga Toragami
2. Lalabell Lullaby
>>99023783You dont have to, let me tell you a trick, use image to image, with a low denoising and a pure solid black color image, that way it'd work with any model, or atleast that's how I do it. You might need to mess with the denoising settings until you get it right.
>>99014670Wao good stuff anon
>>99030405I've trained models with a single image, so I guess I can say a few things. The idea is you can crop high quality art into sections to reuse it, so lets say you have the full image at first, crop it to the face next for example, and so on. Then use an autotagger like Eva02-Large Tagger v3, to generate captions for those images, and mess with the tags until they're right, and you want to be as accurate as possible because with little images, any mistake will be amplified. (Basically just follow a normal lora training guide, and come back and use a lower learning rate, over a larger number of epochs, so you can pick a model from a larger range)
