---

Ars AI

Explore. Analyze. Understand. Advance.

---

About

Ars AI is an independent AI research laboratory dedicated to the exploration of open-source generative AI, with a primary focus on Stable Diffusion, model merging, cyclical training, and iterative refinement.

What began as simple curiosity about model merging gradually evolved into a long-running research project spanning multiple generations of Stable Diffusion models. Rather than treating every release as a standalone checkpoint, each model became another experiment—testing merge strategies, training methodologies, datasets, and architectural ideas.

Ars AI is the culmination of that journey: a place where every success, every failure, and every "what if?" becomes a stepping stone in an ongoing body of research.

---

Research Evolution

Phase I — The Mix Era (Stable Diffusion 1.5)

Every journey has to begin somewhere. Mine began during the Stable Diffusion 1.5 era, where I explored the fundamentals of model merging through projects such as Ketchup Mix, Mayonnaise Mix, Saki Mix, AnyRGB Mix, AnySushi, BreakMaple, and Funizu.

These early experiments weren't merely attempts to create better models—they were opportunities to understand why certain merges worked, why others failed, and how seemingly unrelated checkpoints could produce entirely new capabilities when combined. This period laid the technical foundation for everything that followed.

Phase II — The Abyss Series

With the fundamentals established, research shifted from isolated experiments toward building cohesive model families. The Abyss lineage—

• AbyssMizuHell
• AbyssLunaMix
• Abyssal Blossom
• Abyss Light
• Dream Abyss

—represented my first serious attempt at iterative model development, where every release directly informed the next. Rather than producing isolated checkpoints, the goal became continuous refinement.

Phase III — The SDXL Transition

The arrival of Stable Diffusion XL marked the beginning of a new architectural era. Research transitioned toward understanding the strengths and limitations of the XL architecture, beginning with NW-VXP XL. It was less about replacing SD1.5 and more about learning an entirely new foundation.

Phase IV — Pony Diffusion

The release of Pony Diffusion v6 introduced another major shift. Projects including NWXL Pony, TelePony, and Ketchup PDXL explored expressive capabilities, merge compatibility, and stylistic diversity within the Pony ecosystem. Many of the techniques explored during this period would later become essential to the Ars research lineage.

Phase V — IPS Mix

When Illustrious XL was released, one question immediately came to mind:

What happens if Pony Diffusion and Animagine XL are merged together?

The answer became IPS Mix. Although ultimately retired, IPS Mix was my first successful merge between the two architectures and served as the proof of concept for everything that followed. Without IPS Mix, the Ars series would never have existed.

Phase VI — The Ars Project

Building upon IPS Mix, research shifted from creating individual models toward establishing a continuous research lineage. The Ars series includes Ars Aeterna, Ars Divina, Ars Caelestis, Ars Domini, Ars Imperia, and Ars Magnificat, alongside numerous intermediate prototypes and experimental branches that were never intended as public releases, but nevertheless influenced subsequent models.

The Ars series marked the point where my work stopped being a collection of independent merges and became an evolving body of research. While several Ars models explored different ideas, Ars Divina ultimately became the primary branch under active development, serving as the foundation for many of the experiments that followed.

Phase VII — Experimental Branches

Research rarely follows a perfectly straight line. While Ars Divina remained the primary focus, new questions continued to emerge. One of them was surprisingly simple:

What would happen if Ars Divina were merged with a mature anime-oriented model?

That question became C-Test. The project opened several unexpected research directions and eventually led to a number of platform-exclusive experimental checkpoints released for communities such as SeaArt, Moescape AI, and PixAI.

Maintaining multiple independent branches eventually became impractical, so development was consolidated back into the primary Ars research lineage. Not every experiment was intended to become a long-term project—but every experiment contributed something valuable.

Phase VIII — Indominus Rex XL

One experiment naturally led to another. The next question was:

What if C-Test were merged with one of those experimental branches and then subjected to cyclical training?

The result was Indominus Rex XL. Built upon years of accumulated experimentation, Indominus Rex XL became my most versatile model to date, capable of handling an exceptionally wide variety of prompts while maintaining impressive consistency.

Unexpectedly, it also developed a remarkable aptitude for generating middle fingers. To this day, I still don't have a satisfactory explanation.

Phase IX — Aeterna Opus

Every research lineage eventually reaches a point where its accumulated knowledge can be brought together into something greater. For Ars AI, that project became Aeterna Opus.

Constructed from the combined foundations of Ars Divina 7.2.4a and Indominus Rex XL, it represents the synthesis of years of experimentation with model merging, cyclical training, iterative refinement, and empirical development. The project's guiding question was straightforward:

Can the successor surpass its own predecessor?

Every iteration of Aeterna Opus is another attempt to answer that question. Its early reception within the community has been both encouraging and motivating, reinforcing the idea that the iterative approach developed throughout the Ars lineage continues to resonate with users.

But like every project before it, Aeterna Opus is not intended to be the final destination. It is simply the next experiment. Because in Ars AI, no model is ever truly finished. Every model eventually becomes training data for the next idea and next generation of researchers.

---

Research Interests

• Stable Diffusion
• Foundation Models
• Model Merging
• Cyclical Training
• Iterative Fine-Tuning
• Image Generation
• Experimental AI Workflows

---

Philosophy

Knowledge is defined not by what is known, but by what remains absent.

• Every model is an experiment.
• Every experiment generates data.
• Every failure teaches something.
• Every successful merge reveals another question worth exploring.
• Even the most ridiculous prompts and the most unexpected results are documented.

---

Open Science

Ars AI is committed to open-source research. Whenever practical, models, workflows, and research findings are shared with the community so others may build upon them, challenge them, or discover entirely new directions.

Progress in AI is rarely the work of a single individual; it grows through collective experimentation.

---

FAQ

How many researchers are in Ars AI?

One, for now.

Do you have a research team?

Not yet. At the moment, Ars AI is powered by one overly curious researcher, one increasingly exhausted GPU, and an unhealthy amount of coffee.

What motivates your research?

Officially:

To advance open-source generative AI through empirical experimentation, reproducible workflows, and transparent research.

Unofficially:

Doing model merges and cyclical training for the shits and giggles.

Somehow... it keeps producing decent models.

---

Acknowledgments & Thanks

1. Someone Very Significant
2. OnomaAIResearch
3. Cagliostro Labs
4. nukeai1106
5. GoofyAI
6. Raelina
7. DaoOwOarts
8. Kohya_ss
9. SeaArt AI
10. PixAI
11. Moescape AI
12. Civitai
13. Comte
14. Minthybasis
15. WAI0731
16. And the entire generative AI community!

"Beyond the Visible."

There is no final model. Only the next experiment.