Trigger words:
ANIMSTY, SHW_YN.
I took my 90's anime lora at 53K steps and trained on top of it 231 clips from Makoto Shinkai's "Your Name" an additional 11.5k steps. But all the dataset was only environmental shots. Landscapes, rooms, cityscapes, etc. basically anything not human focused. There are some crowd shots and some scenes do have people though. As a result the people are more in the 90s anime style design but the animation look and feel is more modern. The downside is that the model started to trend back towards a more "3D" look and motion on the people.
This was mostly a test to see if a detailed work like Shinkai's could act a detailer for environments without too much negative effect on other aspects. And I think for the most part it worked, even the people's faces have more details. I think this dataset will work well mixed into a larger pool for a modern anime lora. Give it a try and let me know what you think.
Description
FAQ
Comments (6)
💕👍🏻
You posted this lora on Huggingface the other day. Is this an improved version or the same one?
I dont think I posted it on HF. Are you talking about the 90s anime lora?
@tazmannner379 I think I've confused you with someone else. This person is on the Sulhpur model's Discord server, and he made a hentai model who has similar trigger words.
Awesome work. Have you tried training from ltx 2.3 sulphur base model? Even though Sulphur was finetuned without using 2D data, from my testing I got much better anime loras when training on top of it.
Beautiful 👍