A new LoRA (Low-Rank Adaptation) model, named SDA Diversity LoRA, has been released for Krea 2 Turbo. This LoRA aims to restore the sampling diversity that was lost during the Turbo distillation process of Krea 2 Turbo. It achieves this by using Semantic Directional Alignment (SDA) and Forward XM candidate exploration, without negatively impacting image quality or prompt adherence. The creator notes that the LoRA is most effective when applied within the first two denoising steps. AI
IMPACT Enhances existing image generation models by restoring lost diversity, potentially improving creative output for users.
RANK_REASON Release of a LoRA, which is a fine-tuning technique for existing models, rather than a new foundational model.
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
