Finetune Stable Diffusion Models with DDPO via TRL

Finishing Stable Diffusion models with DDPO via TRL is neat, but it's just an incremental tweak to the fine-tuning process. It's a clever way to manage exploration, I'll give it that. But don't get excited about the methodology; focus on the cost. Fine-tuning massive models like SDXL usually burns serious compute time. Make sure your TRL setup doesn't introduce unnecessary overhead on your VRAM.