Overview
Holodayo XL 2.1 is the latest version of the Yodayo Holodayo XL series, following the previous iteration, Holodayo XL 1.0. This open-source model is built upon Animagine XL V3, a specialized SDXL model designed for generating high-quality anime-style artwork. Holodayo XL 2.1 has undergone additional fine-tuning and optimization to focus specifically on generating images that accurately represent the visual style and aesthetics of the Virtual Youtuber franchise.
Holodayo XL 2.1 was trained to fix everything wrong in Holodayo XL 2.0, such as bad hands, bad anatomy, catastrophic forgetting due to the text encoder being trained during the fine-tuning phase, and an overexposed art style by decreasing the aesthetic datasets.
Model Details
- Developed by: Linaqruf
- Model type: Diffusion-based text-to-image generative model
- Model Description: Holodayo XL 2.1, the latest in the Yodayo Holodayo XL series, is an open-source model built on Animagine XL V3. Fine-tuned for high-quality Virtual Youtuber anime-style art generation.
- License: Fair AI Public License 1.0-SD
- Finetuned from model: Animagine XL 3.1
Supported Platform
- Use this model in our platform:
- Use it in
ComfyUIorStable Diffusion Webui - Use it with 🧨
diffusers
🧨 Diffusers Installation
First install the required libraries:
pip install diffusers transformers accelerate safetensors --upgrade
Then run image generation with the following example code:
import torch
from diffusers import StableDiffusionXLPipeline
pipe = StableDiffusionXLPipeline.from_pretrained(
"yodayo-ai/holodayo-xl-2.1",
torch_dtype=torch.float16,
use_safetensors=True,
custom_pipeline="lpw_stable_diffusion_xl",
add_watermarker=False,
variant="fp16"
)
pipe.to('cuda')
prompt = "1girl, nakiri ayame, nakiri ayame \(1st costume\), hololive, solo, upper body, v, smile, looking at viewer, outdoors, night, masterpiece, best quality, very aesthetic, absurdres"
negative_prompt = "nsfw, (low quality, worst quality:1.2), very displeasing, 3d, watermark, signature, ugly, poorly drawn"
image = pipe(
prompt,
negative_prompt=negative_prompt,
width=832,
height=1216,
guidance_scale=7,
num_inference_steps=28
).images[0]
image.save("./waifu.png")
Usage Guidelines
Tag Ordering
For optimal results, it's recommended to follow the structured prompt template because we train the model like this:
1girl/1boy, character name, from which series, by which artists, everything else in any order.
Special Tags
Holodayo XL 2.1 inherits special tags from Animagine XL 3.1 to enhance image generation by steering results toward quality, rating, creation date, and aesthetic. This inheritance ensures that Holodayo XL 2.1 can produce high-quality, relevant, and aesthetically pleasing images. While the model can generate images without these tags, using them helps achieve better results.
- Quality tags: masterpiece, best quality, great quality, good quality, normal quality, low quality, worst quality
- Rating tags: safe, sensitive, nsfw, explicit
- Year tags: newest, recent, mid, early, oldest
- Aesthetic tags: very aesthetic, aesthetic, displeasing, very displeasing
Recommended Settings
To guide the model towards generating high-aesthetic images, use the following recommended settings:
- Negative prompts:
nsfw, (low quality, worst quality:1.2), very displeasing, 3d, watermark, signature, ugly, poorly drawn
- Positive prompts:
masterpiece, best quality, very aesthetic, absurdres
- Classifier-Free Guidance (CFG) Scale: should be around 5 to 7; 10 is fried, >12 is deep-fried.
- Sampling steps: should be around 25 to 30; 28 is the sweet spot.
- Sampler: Euler Ancestral (Euler a) is highly recommended.
- Supported resolutions:
1024 x 1024, 1152 x 896, 896 x 1152, 1216 x 832, 832 x 1216, 1344 x 768, 768 x 1344, 1536 x 640, 640 x 1536
Training
These are the key hyperparameters used during training:
| Feature | Pretraining | Finetuning |
|---|---|---|
| Hardware | 2x H100 80GB PCIe | 2x A100 80GB PCIe |
| Batch Size | 32 | 48 |
| Gradient Accumulation Steps | 2 | 1 |
| Noise Offset | None | 0.0357 |
| Epochs | 10 | 10 |
| UNet Learning Rate | 5e-6 | 2e-6 |
| Text Encoder Learning Rate | 2.5e-6 | None |
| Optimizer | Adafactor | Adafactor |
| Optimizer Args | Scale Parameter: False, Relative Step: False, Warmup Init: False | Scale Parameter: False, Relative Step: False, Warmup Init: False |
| Scheduler | Constant with Warmups | Constant with Warmups |
| Warmup Steps | 0.05% | 0.05% |
License
Holodayo XL 2.1 falls under Fair AI Public License 1.0-SD license, which is compatible with Stable Diffusion models’ license. Key points:
- Modification Sharing: If you modify Holodayo XL 2.1, you must share both your changes and the original license.
- Source Code Accessibility: If your modified version is network-accessible, provide a way (like a download link) for others to get the source code. This applies to derived models too.
- Distribution Terms: Any distribution must be under this license or another with similar rules.