Death note manga style [Anima]

https://civitai.com/models/2697669/death-note-manga-style

Trigger Word: Death_note_manga
Model: anima-base-v1.0
Diffusion-pipe was used for training

Background

Subscribers who follow my account (Wow! There are already over 100 of you! Where did you come from? But you're definitely cool :)) I think you've noticed that I've previously posted lora exclusively for video models. I'm not a big fan of training lora for static images. Previously, I only trained SD 1.5 on small cats for training purposes. And the generation quality wasn't very good. But I kept an eye on the new models that were popping up like mushrooms after the rain. At some point, I came across a Reddit post called multiple_characters_anima_generations_are_so_good and was a bit stunned by how Anima could generate art with multiple anime characters.

Dataset

Inspired, I decided to train the stylistic lora using a selection of color art from Death Note by Takeshi Obata. I'd had this dataset for quite some time, and I experimented with it on the ltx and wan models. But, as expected, lora trained on static images didn't work correctly in video models, so the experiments failed. The dataset itself consists of 57 images with natural language descriptions. I considered changing the dataset description to tags, but judging by the official readme, Anima trained using more than just tags, so I left it as is. The dataset itself is very complex and has a huge amount of detail and a large number of characters in each image. It didn't include black-and-white frames from the manga. Although the manga itself is certainly gorgeous.

Lora Training

Since this was my first time training the Anima model, I decided to experiment with the training parameters. Below is an example of the 100th epoch trained with different parameters.

Unfortunately, I made a typo and trained with 5e-04 instead of 5e-05... But I was too lazy to retrain, so I decided to use these experiments as a basis. I also tested visualizations on earlier epoch, but I didn't really like them. I tried to choose a lora that maintained a balance of style and wasn't overly overtrained on specific characters. Essentially, I was looking for something that would give a stylish effect without distorting the original prompt. Therefore, I ran tests on a prompt that didn't contain characters from the original Death Note. Training was done at 1536 resolution, as this is the maximum recommended resolution for anima. Rank=32, lr=1e-04.

P.S. Big thanks to Smexlo for sharing the amazing prompts