AI Influencer LoRA Training + Dataset Generation Workflow In ComfyUI
Master Character Consistency: The Ultimate LoRA Training Guide for AI Influencers
Character consistency is the "final boss" of building a successful AI influencer. In this video, I break down my entire professional workflow for creating a rock-solid character LoRA using Qwen Image 2.5 and ComfyUI.
We move past simple generations and dive deep into dataset preparation, multi-step upscaling for skin realism, and cloud-based training on RunPod. Whether you’re an indie hacker or a developer looking to level up your GenAI game, this is the blueprint for high-fidelity, consistent results.
🚀 What You’ll Learn:
Dataset Strategy: Why you only need 3 base images to start.
The Qwen Stack: Using Qwen Image 2512 for generation and Image Edit 2511 for perfect variations.
Advanced Upscaling: Adding hyper-realistic skin textures with Z-ImageTurbo and SeedVR2.
Control Nets: Using DW Pose to fix "short/stocky" character issues.
Automated Captioning: Using Qwen3-VL to generate precise training captions.
Cloud Training: Step-by-step Ostris AI Toolkit setup on RunPod (RTX 6000 Pro).
🛠️ Workflows & Resources
ComfyUI Workflow:
drive.google.com/file/d/1zRKRqYqJa9PCawt_pSPFFYnZo4Hot1Ka/view?usp=sharing
Recommended GPU: RTX 6000 Pro (or any 24GB+ VRAM card)
📑 Video Timestamps
0:00 The Problem: Character Consistency
0:53 Dataset Generation process
1:38 Model Breakdown (Qwen 2.5 Image & Edit)
3:35 Stage 1: Generating the Base Character
4:30 Stage 2: Multi-step Upscaling for Realism
6:00 Stage 3: Generating Dataset Variations (Poses & Outfits)
7:25 Fixing Proportions with DW Pose & ControlNet
14:55 Stage 4: Batch Captioning with Qwen3-VL
18:04 Stage 5: Training on RunPod (Settings & Optimization)
25:50 Analyzing Training Samples (Step 0 to 3000)
28:40 Final Results & Testing the LoRA
#AIinfluencer #ComfyUI #Qwen2 #LoRATraining #GenerativeAI #AIModelTraining #IndieHacker