Claude Opus breaks down the task into subtasks, then deploys instances of Haiku to solve each task. Once these tasks are solved, the output goes back to Opus for review. Pretty incredible.
When I tried to build things like this with GPT4, it was very hard for the model to know when to "stop."
Ever wondered how to fine-tune SDXL for any style you desire?
I've got the simplest guide for you using @replicatehq.
Let's dive in!
Why fine-tune?
Fine-tuning lets you specialize a model to specific styles, people, or objects, making it perfect for generating high-quality, consistent images. SDXL's capabilities are game-changing.
Here an example from one of my model trained on Hiroshi Nagai art.
First, decide on your style and gather images. I usually use 30-50 high-quality pics. Remember, it's not about quantity but QUALITY. Include variety if you're aiming to emulate a particular style. Sometime even 10 good pics are enough.
Many people have asked me for a tutorial on how I built Zing (Bing AI clone).
Follow me in this step-by-step and how with a bit of nudging and clever prompting, I got GPT-4 to do that for me in around 12 messages.
🧵👇
First GPT-4's knowledge still stops at 2021, which means it has no idea of how to implement @OpenAI's new Chat API. So, I just provided it with the entire API documentation in the first prompt.
It worked! It seems to understand how this new API functions. We're in for some fun now. 🔥