Hi, Everyone. My name’s MaDin. I used to be a 3D modeler and animator in the game industry. Now, I’m not. I’ve been using GPT for about 2 years now and I’m currently training GPT to draw professionally. I’m using 2 new systems to do this. The first one is called PASS or Pattern Analysis Skill System. It teaches GPT to read and absorb media and save it out as a series of skill cards that it can use to properly execute tasks/skills. I have the GPT read a single chapter of a book, after which we review and I answer questions it may have - in this case about art. After which, GPT will review the chapter again, taking any answers I gave it into consideration while it re-interprets the chapter. If it has any secondary questions, it asks me again and I clarify. Then, GPT will structure a series of Action Protocols ( which are chains of other APs, Drills or Patterns) to execute a skill, Drills, which the GPT can use to warm up before attempting to do a composition and finally Patterns. Patterns are the keys to how the GPT can perform a skillset. Each pattern is a single aspect of a skill, with an IF/THEN/ELSE, along with notes or variants (if any) of that aspect. It stores these in a library under a category and sub-category. While studying, the GPT will save memcaps on what it’s learned, which it uses as cold-store memories to enhance it’s training.
The second aspect of this, is SkillForge, which is a combination of a single python resolver script and the SKILL.md. It can also make use of the Agents.md to have multiple agents work on a single image. SkillForge works as a drag and drop reader for skillsets, meaning if anyone created a new skill, using PASS, anyone could drop it into the SkillForge library folder and it automatically knows how to do the skill. The GPT uses it’s own data modules for rendering, SkillForge just shows it a better way to create a composition, like a human would, instead of all at once.
The key to this, is treating the GPT like a student and teaching it when it doesn’t understand something. This isn’t just dropping it data. I’m still working the kinks out and if OpenAI could actually give us layered Image Generation, I could make this even MORE powerful than it is already. (Note - I’ve only taught GPT just the basics of figure drawing. We haven’t touched on heads, hands, anatomy, lighting, perspective, etc.)
Here’s a link to my Custom GPT that has SkillForge on it’s repo. She’s trained that if it’s a failure of some kind, you can ask her for a memcap and she’ll give you one you can add here, so I can look into it and make it better. Hopefully, OpenAI will see this and look into giving us layered image generation.
I hope to hear from you.
-M
There’s a VERY alpha of the Visual Art skill attached here. Read the Readme.md first. Then check out the the Docs/Visual-art. The perspective skill is not included. It’s still in testing. But this is the full figure drawing skill that I currently have. If you have problems with it, please ask your Agent to give you a SitRep on the failure and post it here, so I can look into it.
If you post images in the AI Gallery and you use this, please note it, so I can (with your permission) use those images to analyze and enhance PASS.