A specialized UltraReal LoRA has been released for the Krea 2 base model, designed to solve one of the main problems of modern diffusion models — the 'plastic skin' effect. By training on high-quality 4K images, the new adapter adds natural pore texture and fine details, making character generation significantly more natural.
What Happened
The community has gained access to the UltraReal LoRA, developed specifically for the Krea 2 model. The tool utilizes a dataset of high-quality 4K images, allowing the model to reproduce skin texture in great detail, which is particularly effective in close-up and medium shots. Additionally, the training helped reduce visual bias toward Asian faces, ensuring more diverse generation.
Context
Modern base diffusion models often suffer from excessive texture smoothness and ethnic uniformity. The use of highly specialized LoRA adapters is becoming the standard for fine-tuning visual styles and physical material properties without the need to fully retrain massive base models.
Why It Matters for the Industry
The development of such modular approaches allows for the rapid correction of fundamental flaws in base models, transforming them from general-purpose tools into commercially viable solutions for niche sectors such as beauty, fashion, or medicine. This demonstrates the effectiveness of using small adapters to customize neural networks in production-ready environments.
Why It Matters for Users
Krea 2 users can now obtain photorealistic portraits without the 'doll skin' effect, integrating the UltraReal LoRA into their workflows to produce content ready for use in commercial layouts. There is also improved capability for precise control over character ethnicity through text prompts.
What Is Not Yet Known / Limitations
There is a potential conflict between technical progress achieved through specialized datasets and the legal risks associated with copyrights on the 4K images used for training.
Sources
Author
Look at AI, Editorial Team