Summary
Highlights
The final part touches upon monetization. Once videos gain traction, traffic from TikTok is redirected to Instagram, and then to a 'Bridge Page' created using Link Me, ultimately leading to Funview. The instructor concludes the guide, encouraging viewers to subscribe to their Telegram channel for more information and to like the video if it was helpful.
This video introduces a fundamental course on creating an AI model from scratch. It outlines a six-step process: creating the model's face, building a photo dataset, training Lora, creating video content with ComfyUI, preparing devices for social media, and warming up media accounts. The instructor encourages viewers to follow along and subscribe to their Telegram channel for additional resources and community support.
The first step involves creating a unique face for the AI model. This is achieved by selecting two different female faces from Pinterest, saving them, and then combining them using the 'Wavespeed' website's 'Cdream V4.5' model. The process involves inputting a prompt (available on the instructor's Telegram) and swapping the order of the images if the initial result is too similar to one of the references. The goal is to generate an original appearance for the model.
The second and third steps focus on creating a high-quality dataset and training Lora. The dataset should consist of three types of photos: face, face plus body, and full body, with 10 photos for each category, totaling 30 images. These images are sourced from Pinterest as references, and prompts for generating new images are created using ChatGPT, then processed with Nanobanana Pro. It's crucial to ensure all photos are 'perfect' to avoid compromising the quality of future AI generations. Once the photos are prepared and numbered, text files containing a 'Trigger Word' and descriptions (generated by ChatGPT) are created for each image. Finally, these images and text files are compressed into a ZIP archive and uploaded to the 'Lora Trainer' on Wavespeed, where the Lora is trained for 3,000 steps, typically taking 15-20 minutes.
The fourth step involves creating video content using Runpod and ComfyUI. The instructor explains how to fund a Runpod account (suggesting a bot for those in Russia without foreign cards to avoid high commissions) and select a suitable GPU (like the 2X40 for a more economical option). ComfyUI is then set up by changing the template and adjusting specific overrides. The trained Lora model is uploaded to Runpod via Hugging Face using a direct download link. Additionally, an upscaling model ('1x Skin contrast superultra') is uploaded. A workflow from the instructor's Telegram is imported into ComfyUI. To generate video, a reference TikTok video is downloaded, a screenshot is taken, and a prompt describing the model is generated with ChatGPT. This prompt, along with the 'Trigger Word', is input into ComfyUI. The video's duration is set, and specific nodes are bypassed to allow the model to dance on its own background. After processing, the generated video is saved. A key tip is to segment longer videos into 5-second parts for better quality.
After video generation, it's crucial to clean the video's metadata using video editing software like CapCut. This involves adding filters, slight blur, trimming, and color correction to bypass social media checks for AI-generated content. For device and account preparation, the instructor provides guidelines on using a reliable VPN (like ExpressVPN), buying Instagram and TikTok accounts with a 7-14 day 'rest period' (отлежка) from Dark Store or Bullkc, or creating new accounts with Gmail or Outlook emails for better trust. A detailed 7-14 day warm-up plan for accounts is also provided, emphasizing that if an account doesn't show stable organic growth, it should be changed, and the device should be reset to factory settings to ensure a clean slate, enhancing trust with social networks.