Sign In

Project for Improving and Finalizing the Pony Model

Project for Improving and Finalizing the Pony Model

From the moment the Pony model was introduced, I saw great potential in it, especially when controversies surrounding SAI and SD3 arose, causing me to worry about the open-source community, where everyone could freely express their creativity, being destabilized and potentially becoming abandoned in this era. Then came a slew of laws enacted in various countries around the world to restrict AI-generated NSFW content, which gradually left me feeling hopeless.

But at this moment, Pony emerged like a small and rugged path, making me feel that things weren't over yet, and we still had a chance.

Our freedom might be restricted, and most of the world might not understand us. But we still have a peaceful place to retreat to, somewhere we can rely on, freely do what we love, and share our interests with others.

Thanks to AstraliteHeart, and thanks to Civitai for maintaining their stance and not banning NSFW. A path does not exist naturally, but many determined people moving forward make it a path. It might just be a small trail now, but if we have brothers fighting the same battle, it will eventually become a wide boulevard.

LOL, okay, that's enough of my speech. Now, let's get to the main topic of this article.


For those who are familiar with Pony, you already know about its shortcomings and the issues frequently encountered during use, with all kinds of unwanted errors occurring.

In truth, we can't blame AstraliteHeart; after all, he only has one person, probably, that's him telling me ha ha, but he investment is certainly much, much, much lower than that of SAI. Therefore, it's inevitable that Pony has many errors. Nothing is perfect, even the SAI models based on SD that have not been fine-tuned by the community are terrible.

I’ve tried to list the issues with Pony v6, which are as follows:

1. The fragmented ecosystem leading to poor reasoning ability.

2. Being trained on too many keywords, making it rigid and poor in understanding natural language, descriptions, and artistic or emotional abstract art.

3. Some keywords, characters, and phrases cause image errors, such as increasing noise or creating unwanted spots, etc.

4. Lack of a realistic image style.

5. No proper nouns or specific names. While this might be an advantage in avoiding AI-related controversies, but it makes some people who are used to using nouns or proper names to have images that properly express their desires now unable to do so anymore.

6. The tag_score issue. If you want the best and most accurate image, you have to understand how Pony's tag_score works when included in a prompt, which annoys some people.

These are the issues I think Pony v6 is currently facing. If anyone finds any other issues, please let me know.

And as the title of the article "Project for Improving and Finalizing the Pony Model" suggests, we will delve into each issue.

I started this project when my first fine-tuned version of Pony was released. It’s a long and costly process, and it can't be completed quickly, especially since I mostly have to do it alone and invest my own money into this bottomless pit.

Therefore, I’ve divided it into multiple Pony models, each with several versions that improve over time. But all of them revolve around the "project for improving and finalizing the Pony model."

As you can see, Pony is very different from SDXL, which is why combining them faces significant resistance and cannot achieve the best results. Even when training, it differs from training the SDXL model.

Especially when you want to improve it while retaining the essence of Pony v6. Adding something new always disrupts the original structure, and how can you choose exactly what you want to change in the model? That is my difficult problem.

DucHaiten-GameArt is my experiment with how cubist image data like 3D will affect the training of the Pony model. If it has enough diverse NSFW data, it will be the precursor for my future Pony realistic model.

DucHaiten-Pony-XL (no-score) is an experiment to limit the significant influence of tag_score in the Pony model, as well as improve image quality, reduce dependency on tag_score in prompts, improve natural language comprehension, enhance abstract art capability, supplement the model's ecosystem, improve reasoning ability, and reduce rigidity due to Waifu Diffusion's keywords.

I must say I'm genuinely surprised and delighted with the warm reception for the DucHaiten-Pony-XL (no-score) model. Each new version has drawn more attention than the previous one, and it seems I'm on the right track. By the way, while writing this article, I'm also working on DucHaiten-Pony-XL (no-score)_v4.

To express my sincere thanks to everyone, I'll release the v4 beta version first. You can try it out and provide feedback and suggestions for it. Note that it's only a beta version, so many changes will occur in the future.

[Link to the model](https://civitai.com/models/376450?modelVersionId=491406)

Besides, it's unfortunate that I'm struggling with issues 3 and 5. For issue 3, I can only make it slightly better, but a broken image is still a broken image, and it's unusable. For issue 5, I don't think I have enough authority to do anything about it. Perhaps these two issues can only be addressed by AstraliteHeart and his team. Hopefully, Pony v7 will resolve most of the issues I've highlighted.

Finally, here’s a little lamentation from me. If you don't want to hear it, you can skip it, LOL...

As mentioned above, I mostly do the work alone, and almost all of it is open-source, and the money comes from my own pocket. I often have to work 17 hours a day, sometimes even 20 hours, but since it's a hobby, I don't feel like I'm suffering, haha.

It’s also great that I've recently received help with finding and processing data from deathless, but after all, he's just volunteering for free, so I can't ask for too much or impose too many demands. He will only find the image data he likes, but for me, that’s more than enough. I can't ask for more.

Perhaps you don't know that all the money I earn comes from odd freelance jobs related to training AI models, and I don't earn much. Luckily, I don't spend much on personal life (though recently I've had to spend more due to some issues), but most of my spending goes into personal projects like this one and the future Pony realistic project. All of them will be freely available on Civitai, I promise.

Previously, everything was fine, but recently things have changed. My income is decreasing, and I have to spend more on AI projects to improve already good models further. With income decreasing and expenses increasing, I'm starting to struggle a bit.

I once thought about setting up a Patreon and actually made one, but I don't have the time or skills to manage it, so I left it unattended for a long time. I initially planned to let my subscribers use beta models first, but later decided not to do that. I fear someone might upload the beta model to platforms like Civitai before me, or someone might remix it into a new model before I can publicly claim it as mine. But it’s still there for anyone who wants to support my current and future projects, haha...

[Link to Patreon](https://www.patreon.com/duchaitenreal)

Additionally, you can donate directly via PayPal:

[Link to PayPal](http://paypal.me/duchaiten)

Listen, this is not mandatory. No one will force you to do something you don't want to do. I'll continue working on my projects even if I don't receive any donations, as long as my AI training income can sustain it. I'll do my best not to go bankrupt.

But if you really want to help and are willing to make a small contribution, I would be very grateful. It will help me progress faster and do things better. Thank you.

9

Comments