In this video I walk you through an all-in-one starter guide to ComfyUI for AI influencer creation—without assuming you know anything yet. I show the main ways to install ComfyUI (terminal, Easy Install, Portable), and I explain when you should skip local installs and just use RunPod if you don’t have a proper GPU. I break down what Sage Attention is and why it matters for speed—especially for newer models and stronger GPUs—then I show which launcher options to use and how to update ComfyUI safely without constantly breaking your setup. After that, I make you run the default ComfyUI workflow on purpose. It’s not sexy, but it proves your install actually works, your GPU is detected, and you’re not troubleshooting a broken custom workflow before you even know your basics are stable. I show you where models come from (Hugging Face), where to place them in your ComfyUI folders, and why refreshing the browser matters after you add anything new. Then I explain custom nodes and why most people get stuck—because they load a “crazy CivitAI workflow” and everything turns red. I show you how to use ComfyUI Manager to install missing nodes and keep your setup functional. I also give a simple explanation of FP16 vs FP8 vs FP4, how to think about speed vs quality, and why I like FP8 as a sweet spot on my 5090. From there I move into LoRAs: what they are, why they’re the key to consistent faces, and how I trained one (Audrey Hepburn as a demo, but I warn you not to use real people’s likeness commercially). I show my dataset workflow: collect images, generate captions for each image using my dataset helper tool, then train in Ostris AI Toolkit with simple settings so you can actually get results without overthinking. Finally, I show the “correct” way to make a brand-new influencer that doesn’t exist—generate a clean base face (Z-Image Turbo), upscale for realism, create multiple angles with an image-edit workflow, then feed that into the same dataset + training pipeline. I close by pointing out that ComfyUI has built-in templates you should explore before you go hunting random workflows, and I warn you that storage becomes a real issue fast—models, outputs, encoders, everything adds up.

I am now migrating all paid fan content to smallzero.io You can join for free or as a paid member. All members get access to the discord and upscaler. I appreciate your support, free or premium.

Hi everyone! Could you provide prompts to create a dataset with which to train my LoRA? Thank you so much!!!