metadata

license: openrail++
tags:
  - text-to-image
  - stable-diffusion
  - diffusers

AnimeBoysXL v3.0

It takes substantial time and efforts to bake models. If you appreciate my models, I would be grateful if you could support me on Ko-fi ☕.

Features

✔️ Good for inference: AnimeBoysXL v3.0 is a flexible model which is good at generating images of anime boys and males-only content in a wide range of styles.
✔️ Good for training: AnimeBoysXL v3.0 is suitable for further training, thanks to its neutral style and ability to recognize a great deal of concepts. Feel free to train your own anime boy model/LoRA from AnimeBoysXL.
❌ AnimeBoysXL v3.0 is not optimized for creating anime girls. Please consider using other models for that purpose.

Inference Guide

Prompt: Use tag-based prompts to describe your subject.
- Tag ordering matters. It is highly recommended to structure your prompt with the following templates:
```
1boy, male focus, character name, series name, anything else you'd like to describe
```
```
2boys, male focus, multiple boys, character name(s), series name, anything else you'd like to describe
```
- Append
```
, best quality, amazing quality, best aesthetic, amazing aesthetic, absurdres
```
  to the prompt to improve image quality.
- (Optional) Append
```
, year YYYY
```
  to the prompt to shift the output toward the prevalent style of that year. YYYY is a 4 digit year, e.g. , year 2023

Negative prompt: Choose from one of the following two presets.

Heavy (recommended):

lowres, bad, text, error, missing, extra, fewer, cropped, jpeg artifacts, worst quality, bad quality, watermark, bad aesthetic, unfinished, chromatic aberration, scan, scan artifacts

Light:

lowres, jpeg artifacts, worst quality, watermark, blurry, bad aesthetic

VAE: Make sure you're using SDXL VAE.
Sampling method, sampling steps and CFG scale: I find (Euler a, 28, 5) good. You are encouraged to experiment with other settings.
Width and height: 832*1216 for portrait, 1024*1024 for square, and 1216*832 for landscape.

🧨Diffusers Example Usage

import torch
from diffusers import DiffusionPipeline

pipe = DiffusionPipeline.from_pretrained("Koolchh/AnimeBoysXL-v3.0", torch_dtype=torch.float16, use_safetensors=True, variant="fp16")
pipe.to("cuda")

prompt = "1boy, male focus, shirt, solo, looking at viewer, smile, black hair, brown eyes, short hair, best quality, amazing quality, best aesthetic, amazing aesthetic, absurdres"
negative_prompt = "lowres, bad, text, error, missing, extra, fewer, cropped, jpeg artifacts, worst quality, bad quality, watermark, bad aesthetic, unfinished, chromatic aberration, scan, scan artifacts"

image = pipe(
    prompt=prompt, 
    negative_prompt=negative_prompt, 
    width=1024,
    height=1024,
    guidance_scale=5,
    num_inference_steps=28
).images[0]

Training Details

AnimeBoysXL v3.0 is trained from Pony Diffusion V6 XL, on ~516k images.

The following tags are attached to the training data to make it easier to steer toward either more aesthetic or more flexible results.

Quality tags

tag	score
`best quality`	>= 150
`amazing quality`	[75, 150)
`great quality`	[25, 75)
`normal quality`	[0, 25)
`bad quality`	(-5, 0)
`worst quality`	<= -5

Aesthetic tags

tag
`best aesthetic`
`amazing aesthetic`
`great aesthetic`
`normal aesthetic`
`bad aesthetic`

Rating tags

tag	rating
`sfw`	general
`slightly nsfw`	sensitive
`fairly nsfw`	questionable
`very nsfw`	explicit

Year tags

year YYYY where YYYY is in the range of [2005, 2023].

Training configurations

Hardware: 4 * Nvidia A100 80GB GPUs
Optimizer: AdaFactor
Gradient accumulation steps: 8
Batch size: 4 * 8 * 4 = 128
Learning rates:
- 8e-6 for U-Net
- 5.2e-6 for text encoder 1 (CLIP ViT-L)
- 4.8e-6 for text encoder 2 (OpenCLIP ViT-bigG)
Learning rate schedule: constant with 250 warmup steps
Mixed precision training type: FP16
Epochs: 40

Changes from v2.0

Change the base model from Stable Diffusion XL Base 1.0 to Pony Diffusion V6 XL.
Revamp the dataset's aesthetic tag based on the developer's preference.
Update quality score and aesthetic score criteria.
Use FP16 mixed-precision training.
Train for more epochs.