Type | |
Stats | 335 0 |
Reviews | (64) |
Published | Mar 5, 2023 |
Base Model | |
Training | Steps: 1,560 Epochs: 20 |
Hash | AutoV2 91931F1F04 |
Hifu, it's just the name I gave to a personal Lora. There are no real characters, all of them are personal.
She's simply a cat-eared girl with black hair and red eyes, and always has a flower pinned to her hair.
Why am I posting it here?
Simply put, I am using cloud computing systems (Because my computer is from the time of the dinosaurs) and this is one of my character styles, so using this platform to store my models is not a bad idea, whenever I need to use them, I can just request them here. Besides, I also want everyone to have a fresh look at how to use Lora for their own style.
Holy cow! How did you do it, I want to know too!
It's simple, I create images on a topic I want. Then I use Dreambooth to train a Lora on those images of mine.
Why is it so small in size?
I'm not sure why, as I used to export Lora from a fine-tuned model, the result is a tiny Lora file. Also, I don't know how to train a Lora correctly, as mentioned above, I used Dreambooth to train a new model and then exported Lora from the resulting model.
I'm interested in your Lora, please tell me how to use it!
Of course, I'm happy to share. First, download the model using the download button or if you using Colab:
!wget https://civitai.com/api/download/models/18961 -O /content/stable-diffusion-webui/models/Lora/hifu.safetensors
Or:
import requests url = https://civitai.com/api/download/models/18961 save_path = /content/stable-diffusion-webui/models/Lora/hifu.safetensors response = requests.get(url) with open(save_path, 'wb') as f: f.write(response.content)
You need to add the Lora tag "lora:hifu:1", and add some tags like "cat ears, black hair, red eyes, hair flower", finally, you can have a character similar to the picture.
The images above were created using my Lora combined with the CetusMix model.
What model did you train it on?
It is a secret... But if it's you, you can do it on any model you want.
Have nsfw?
Well, duh! =)). You can try it.
Training:
Dataset: You can get the dataset on my Huggingface page. https://huggingface.co/datasets/Ojimi/Hifu-Lora-Example/tree/main . This dataset I did not take this from any author. All created by me.
Schedule: DDIM.
Optimizer: AdamW.
Precision: FP32.
Hardware: Google Colaboratory Pro - NVIDIA A100 40GB VRAM.