Inference
For inference I used ComfyUI.
Trigger words: sh1nka1 style (if the effect is weak, add also the word "animation" to the prompt)
Strength: 0.8-1.0
Trainig details:
Trained only on videos.
dataset: 143 videos 640x352 81,65,33 frame videos
steps: 20020
LR: 2e-5
optimizer: AdamW8bitKahan
rank: 32
batch size: 1
gradient accumulation steps: 1
For training I used the diffusion-pipe repo.
Important Notes:
This LoRA is created as part of a fan project for research purposes only and is not intended for commercial use. It is based on the movies called "Tenki no Ko" and "Kimi no Na wa" that are protected by copyright. Users utilize the model at their own risk. Users are obligated to comply with copyright laws and applicable regulations. The model has been developed for non-commercial purposes, and it is not my intention to infringe on any copyright. I assume no responsibility for any damages or legal consequences arising from the use of the model.
Description
FAQ
Comments (3)
👍🏻👍🏻👍🏻
Can it be used for Tusheng video,It turns out that it takes so many steps to train a LoRa,I feel that my 30 videos and 3720 steps out of the video are still very bad.
I don't think you need so much steps like I used for this. It is purposedly overtrained so you can use it with lower strength. It makes it easier to merge it with other loras. Moreover I found some "cursed" prompt that doesn't work with the lora so I wanted to see if I could make them work with more training. Spoiler: I couldn't :) but I get the style already at much lower steps I just decided to share this one. Also the more videos you have the more training you need. So with 30 videos you'll get results much sooner.
Details
Files
Available On (1 platform)
Same model published on other platforms. May have additional downloads or version variants.