Struggling with Artifacts in Concept LoRA Training? Parameter Guide
Ton_Phanan · reddit · 2026-08-24
A user training a Concept LoRA encountered significant artifacts and is now curating the dataset for cleaner results. The challenge is that most tutorials focus on Character LoRAs, lacking guidance for concept-specific training. The user asks about the importance of image size, scale, the ideal image count (20-40), and techniques to ensure SFW and uncensored outputs. They are currently using Buzz and Krea 2 on a tight budget.
More from Multimodal
- Report: Video generation consumes 70% of AI tokens in China — c_valenzuelab · 2026-08-24
- Developer uses SeeDance to create Monty Python-style animations for educational video — craigsdennis · 2026-08-24
- Krea 2 with ControlNet + Rebalance node = disaster, user reports — Epictetito · 2026-08-24
- Krea 2 ControlNet LoRA causes plastic disaster, users seek fixes — Epictetito · 2026-08-24
- Gemini 3.7 Flash generates BMW M4 3D model — CodeByPoonam · 2026-08-24
- Gemini 3.7 Flash generates realistic SpaceX Raptor engine — CodeByPoonam · 2026-08-24