Struggling with LoRA dataset prep for realistic virtual influencer: balancing overfitting and consistency
Capable-Swim318 · reddit · 2026-09-01
A user training a LoRA for a realistic virtual influencer in ComfyUI seeks advice on dataset preparation. The core conflict lies in balancing consistency and diversity: keeping clothing and backgrounds constant risks the model overfitting to these features as part of the identity, while using AI editing to diversify them results in a loss of realism and character consistency. The user used Qwen-Image-Edit to generate variations but encountered a 'plastic' AI look. They are looking for workflows and strategies on how much variation to include and how to maintain character fidelity during training.
More from Multimodal
- User Questions Minimax H3 Hybrid Models: Weak Reference Adherence — Tablaski · 2026-09-01
- ABot-Recon lands on Hugging Face: turns dashcam clips into 3D scans in seconds — petewoodbridge · 2026-09-01
- KATok: adaptive video tokenizer drops uninformative tokens for compact representation — kakaocorp · 2026-09-01
- DICS: Exploring Data Intrinsic Consistency for Visual Instruction Selection — SpatialAxiom · 2026-09-01
- Demo: Real-time markerless mocap runs on any camera with <100ms latency — petewoodbridge · 2026-09-01
- AI Music Video Generation Still Rough: Seeking Vibe-Aware Tools — _laundpana · 2026-09-01