AnimeBoysXL

Features

• ✔️ Good for inference: AnimeBoysXL 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 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 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 describe2boys, male focus, multiple boys, character name(s), series name, anything else you'd like to describeAppend , best quality, amazing quality, best 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 2023For more detailed documentation, you can visit my article on Ko-fi (available to supporters only).
• Negative prompt: Choose from one of the following two presets.Heavy (recommended): lowres, (bad:1.05), text, error, missing, extra, fewer, cropped, jpeg artifacts, worst quality, bad quality, watermark, bad aesthetic, unfinished, chromatic aberration, scan, scan artifacts, 1girl, breastsLight: lowres, jpeg artifacts, worst quality, watermark, blurry, bad aesthetic, 1girl, breasts(Optional) Add , realistic, lips, nose to the negative prompt if you need a flat anime-like style face.
• 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.

Training Details

AnimeBoysXL is trained from Stable Diffusion XL Base 1.0, 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

• best quality: score >= 150
• amazing quality: score in the range of [100, 150)
• great quality: score in the range of [75, 100)
• normal quality: score in the range of [0, 75)
• bad quality: score in the range of (-5, 0)
• worst quality: score <= -5

Aesthetic tags

• best aesthetic: score >= 6.675
• great aesthetic: score in the range of [6.0, 6.675)
• normal aesthetic: score in the range of [5.0, 6.0)
• bad aesthetic: score < 5.0

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-Net5.2e-6 for text encoder 1 (CLIP ViT-L)4.8e-6 for text encoder 2 (OpenCLIP ViT-bigG)

Changes from v1.0

• Train with tag ordering.
• Add sfw rating tag.
• More epochs on the questionable and explicit rating subset.
• FP16 mixed-precision training for final epochs.