Aeterna Opus

Overview

Aeterna Opus is a 'premium', high-fidelity anime-themed text-to-image generative model and the latest installment in the series. Built on a structurally sound custom foundation, the model bridges the gap between painterly artistic depth and rigorous character alignment frameworks. The underlying backbone merges the atmospheric lighting of Ars Divina 7.2.4a with the anatomical precision of Indominus Rex XL. Rather than relying on standard fine-tuning, the merged weights underwent continuous, multi-epoch cyclical re-training, constantly rotating optimization passes between OnomaAI’s s1k_Korean dataset for natural language layout logic and CagliostroLab’s 1.2M Ordered Tags for strict Danbooru token compliance. These enhancements make Aeterna Opus highly adaptive, delivering exceptional structural stability and dual-format prompt adherence while maintaining a signature, illustrative edge.

🧬 Architectural & Training Pipeline

Rather than relying on basic fine-tuning, Aeterna Opus utilizes a complex structural and aesthetic convergence process:

The Foundation (Merge Backbone): The core weights are derived from a powerful merge of Ars Divina 7.2.4a (celebrated for the 'random bullshit go' merging process by yours truly; built originally from Illustrious XL 0.1 and Pony Diffusion v6) and Indominus Rex XL (famed for anatomical precision and dynamic structural composition).
The Cyclic Refinement (Re-training): This merged backbone was subjected to continuous, alternating training iterations using Illustrious-XL-v2.0 principles. The cyclical schedule continuously rotated optimization tasks between:
- OnomaAI s1k_Korean Dataset: Embedding natural language prompt comprehension, unique stylistic variety, and robust high-resolution layout logic ($1536 \times 1536$ native capabilities).
- CagliostroLab 1.2M Ordered Tags: Polishing fine-grain detailing, mitigating aesthetic artifacts, and ensuring strict Danbooru tag ordering and consistency.

The result is a highly adaptive model that maintains unparalleled anatomical stability while rendering rich, deeply stylized, illustrative digital art.

Model Details

Developed by: NeverWinter13
Model type: Diffusion-based text-to-image generative model
License: CreativeML Open RAIL++-M
Core Base: SDXL Architecture (Ars Divina v7.2.4a x Indominus Rex XL Backbone)
Training Methodology: Cyclical continuous pre-training
Primary Focus: High-tier illustration, dynamic character staging, advanced multi-concept prompt adherence, and extreme structural stability.

Downstream Use

Use it in ComfyUI or Stable Diffusion Webui
Use it with diffusers

🖍️ Diffusers Installation

1. Install Required Libraries

pip install diffusers transformers accelerate safetensors --upgrade

2. Example code

import torch
from diffusers import StableDiffusionXLPipeline

pipe = StableDiffusionXLPipeline.from_pretrained(
    "NeverWinter13/Aeterna-Opus", 
    torch_dtype=torch.float16, 
    use_safetensors=True
).to("cuda")

prompt = "masterpiece, best quality, newest, 1girl, [clorinde] \\ [genshin impact], dynamic sword stance, dark fantasy background, dramatic cinematic lighting"
negative_prompt = "(lowres:1.2), (worst quality:1.4), (low quality:1.4), (bad anatomy:1.4), bad hands, multiple views, watermark"

image = pipe(
    prompt=prompt, 
    negative_prompt=negative_prompt, 
    width=832, 
    height=1216, 
    num_inference_steps=28, 
    guidance_scale=5.0
).images[0]

image.save("aeterna_opus_output.png")

🎨 Generation Guidelines & Recommendations

Thanks to the Indominus Rex and Ars Divina DNA mixed with Cagliostro/Onoma training cycles, this model responds masterfully to both highly descriptive natural language and structured Danbooru tags.

Generation Guidelines & Recommendations

Prompting Guide

You can utilize standard descriptive natural language sentences or structured tags. For optimal character styling inspired by CagliostroLab's standards, structure your tags as follows:

1girl/1boy, character name \[series name\], [your prompt whether natural language or tokens]

Quality Tags

Add these tags at the start of your prompt:

Positive Modifiers:

masterpiece, best quality, newest

Positive Modifiers2:

masterpiece, best quality, amazing quality, newest

Negatives

You can utilize these negative embeddings or prompts or tags:

Negative Framework:

bad quality, worst quality, lowres, jpeg artifacts, bad anatomy, bad hands, multiple views, signature, watermark, censored, sketch, flat color, ugly, fat, blurry eyes, wrinkled skin

Negative Framework2:

(lowres:1.2), (worst quality:1.4), (low quality:1.4), (bad anatomy:1.4), bad hands, multiple views, comic, jpeg artifacts, patreon logo, patreon username, web address, signature, watermark, text, logo, artist name, censored

Recommended Inference Settings

Sampling Steps: 26 - 30 steps
Sampler: Euler a
CFG Scale: 3.0 - 6.5
Hires.fix is not necessary (Per my case)
ADetailer is highly recommended

Recommended Resolutions

Orientation	Dimensions	Aspect Ratio
Square	1024 x 1024	1:1
Landscape	1152 x 896	9:7
	1216 x 832	3:2
	1344 x 768	7:4
	1536 x 640	12:5
Portrait	896 x 1152	7:9
	832 x 1216	2:3
	768 x 1344	4:7
	640 x 1536	5:12

Final Prompt Structure Example

masterpiece, best quality, newest, 1girl, tewi inaba \(touhou project\), outdoors, floating petals, charming smile, closed eyes, open mouth, volumetric light, lens flare, sunbeam, clouds, trees

💖 Acknowledgments & Thanks

This project would not have been possible without the groundbreaking work, innovative contributions, and comprehensive documentation provided by Stability AI, Novel AI, and the entire generative AI community.

I am especially grateful for those who were with me from the beginning, starting from the Stable Diffusion 1.5 days up through Illustrious-XL, fueling my interest in generative text-to-image models. Particularly:

Thank you. Aeterna Opus is a reflection of this 'random bullshit go' philosophy.

Downloads last month: 131

Model tree for NeverWinter13/aeterna-opus

Base model

OnomaAIResearch/Illustrious-XL-v2.0

Finetuned

(96)

this model

NeverWinter13
/

aeterna-opus