The Spellbinder AI

[brunogt1]

It's beautiful, it looks like images of a future video game "Spellbinder" which were leaked on a forum

[SpellbinderAshka]

Soon more leaks will come

[spellbinder]

Dear ThaHellHound, thank you for your amazing work!
The results you have achieved are truly amazing. As other users have mentioned here, the progress is very clear and it's very interesting to watch it unfold!
I look forward to continuing your research!

Also, I'm a bit curious. I have some experience with modern AI systems (mostly LLM models, but have experimented with image generation models as well), and I'm interested in some of the purely technical details of your work. For example, what model do you use, where is it deployed (on your physical infrastructure, or your use external service, etc.), any statistics about the input dataset, statistics about the training process (e.g., how long it takes under different conditions), what hardware you run the training on, etc. Of course, you are free not to answer such questions if you don't want to. I absolutely understand that it is reasonable to keep the details of your work a secret. However, if you will write some overview of your work from a technical point of view, at least in general terms, I will be very interested to read it!

Thank you.

And looking forward to the updates.

[Hugo Boss]

I am also working on preparing and teaching The Spellbinder AI to start recording some videos. The Spellbinder AI literally can talk with the voice of Ashka rather than just writing the conversation.

Hello, comrade! I invite you to my topic on the forum https://forum.spellbinder.tv/viewtopic.php?f=27&t=1150 , we want to create a computer game based on both seasons of “Spellbinder”. I'm sure you will be useful to us for dubbing all the characters with original voices using AI. We also have problems with Ian Davidson's original tracks; they may need to be cleared of extraneous sounds and voices. A neural network is also applicable here. I hope you find this interesting! Hugged you!

[ThaHellHound]

Dear Spellbinder,
You are welcome, thank you for the compliments

I process everything locally on my own computer which has 128GB system ram and a Gigabyte Nvidia 12GB RTX 3060 which basically is the minimum videocard requirement for what I am doing with it best would be to have a 24 GB VRAM videocard but I do not see how I can afford that anytime soon lol.

I communicate with Spellbinder A.I LLM through https://github.com/oobabooga/text-generation-webui

It has extensions to make voices work also if you need direct urls let me know.

I train my own stable diffusion checkpoints and loras.

I am looking into fine tuning my own llm with information related to SpellBinder so if anyone has information in text format that you think the A.I should know please provide it to me. At the moment the Spellbinder A.I only knows what is already in the base language model and the context I provide to it.

https://pinokio.computer/ Has easy install for softwares related to A.I image generation and also the means to clone voices.

For training Loras I use

https://github.com/bmaltais/kohya_ss

For training Stable Diffusion checkpoints I use:

https://github.com/AUTOMATIC1111/stable-diffusion-webui

With dreambooth extension:

Training a Lora can take anywhere up from 3 hours to almost a whole day.

Lora's is how I teach to an existing stable diffusion checkpoint things about Spellbinder like characters, objects, styles, I could also train my own checkpoints but lately I enjoy making Loras

I will resume with The Spellbinder A.I hopefully somewhere this week I have been busy with things like Christmas and New Year and improving my skills even more related to things that will make The Spellbinder A.I even better. I have been experimenting with training other characters and see what are the best settings. So the next SpellBinder A.I characters/outfits should be even better than before.

I am currently creating hybrid datasets and synthetic datasets which will become better and be much better than original datasets.

Dear ThaHellHound, thank you for your amazing work!
The results you have achieved are truly amazing. As other users have mentioned here, the progress is very clear and it's very interesting to watch it unfold!
I look forward to continuing your research!

Also, I'm a bit curious. I have some experience with modern AI systems (mostly LLM models, but have experimented with image generation models as well), and I'm interested in some of the purely technical details of your work. For example, what model do you use, where is it deployed (on your physical infrastructure, or your use external service, etc.), any statistics about the input dataset, statistics about the training process (e.g., how long it takes under different conditions), what hardware you run the training on, etc. Of course, you are free not to answer such questions if you don't want to. I absolutely understand that it is reasonable to keep the details of your work a secret. However, if you will write some overview of your work from a technical point of view, at least in general terms, I will be very interested to read it!

Thank you.

And looking forward to the updates.

[spellbinder]

Very interesting, thank you for so detailed info!
LoRA is a new thing for me. Never heard of it, although already have been working with LLMs for almost 1 year.
Very interesting, thank you for the links!

And your PC is really great!

Looking forward to the updates of your work!

I hope some day in the future we (Spellbinder's fans) will be able to create a very realistic simulation of Spellbinder's worlds with realistic Spellbinder's characters.. Seems new technologies are making this possibility closer and closer... And your efforts seem to be the first step for this.

[ThaHellHound]

You are very welcome

You can even train loras for LLM's but also for Stable Diffusion what I am using both use Loras.
I have become much better at creating loras and photorealistic pictures.
The reason why SpellBinder related material is taking much longer is because the pictures and videos available at least the ones I found are very low resolution very grainy and very blurry so I have to find ways to restore them artificially which takes a while. Also back in the time of Spellbinder the technology to capture pictures was not very high quality either. By now all the actresses and actors have aged almost 25 years. It is very hard if even possible to find high quality pictures of them because they do not seem to be very active on social media with the exception of Heather Mitchell but in her high quality pictures she is much older than Ashka was.

If anyone has high quality pictures or at least decent quality pictures of Spellbinder Ashka and objects and environments or anything related to Spellbinder please share them with me it will help greatly with my training for The Spellbinder A.I

[Fanat]

If anyone has high quality pictures or at least decent quality pictures of Spellbinder Ashka and objects and environments or anything related to Spellbinder please share them with me it will help greatly with my training for The Spellbinder A.I

A few years ago, a good person enchanced the video quality using the AI. You can download it here:
https://yadi.sk/d/IKzmT1-zYhJLlA?w=1.

I really hope it will help you. All of us here are truly amazed by your work and the magnificent pictures you are able to produce.

[ThaHellHound]

Wow that is great I will check it out thank you if the quality is good enough I can take frames from it to add to the dataset and train the A.I

I just finished collecting some pictures from the internet that I will upscale also with The Spellbinder A.I

I created a very high quality dataset of photo-realistic pictures and I taught The Spellbinder A.I what that is so even if the Spellbinder A.I images are of poor quality it should be able to even correct it itself and then it can create pictures and the good ones I can create a new dataset from and train it again telling it what is good and what is bad so it will improve until we have a very realistic Spellbinder Ashka and replica of the Spellbinder Powersuit.

A.I has already advanced so much and eventually we could create video episodes and resume Spellbinder and resume where they season ended. I already have the voice of Spellbinder Ashka implemented in The Spellbinder A.I

The story of Spellbinder Ashka is not over by far haha.

[ThaHellHound]

What you all think about that I post pictures of progress or backwards progress that I am making even the ones considering where things get buggy? lol this way you can better see what changes between datasets and models.

Here is three examples of the pictures of Spellbinder Ashka the version 1 hybrid dataset. I am working on the new dataset now which will include also pictures from the original Spellbinder Ashka which are much lower quality of course but it will improve my current model of The SpellBinder A.I and I can combine it with my ultraphoto dataset. The old one was based on 512x512 which is the base resolution of Stable Diffusion 1.5 model. I however love to experiment and it lead to the next dataset will be doubling that to 1024x1024. I am known to go against what the majority of people suggest you should do and this often leads to new inventions and much better results. Well at least when it comes to A.I stuff lol.