Important Krea Usage Tips – Raw Model with Turbo LoRA, VAE, and Uncensor LoRA
Apply the raw Krea 2 model with turbo LoRA at 0.6 weight, 12 steps, Euler scheduler, Wan‑2.1 VAE, uncensor LoRA, and reference image support for optimal output.
Implement the recommended settings in your ComfyUI pipeline and test output quality and speed.
Summary
A Reddit user shares a set of practical tips for using Krea 2 effectively. The first tip recommends using the raw Krea 2 model with the turbo LoRA applied at 0.6 weight, 12 steps, and an Euler or beta/beta57 scheduler for better quality than turbo alone. The second tip suggests using the Wan‑2.1 VAE for higher quality, requiring either Kijai’s Diffusers WF or the ComfyUI‑VAE‑Utils plugin. The third tip introduces an uncensor LoRA that improves expression and prompt responsiveness, especially for content that may be censored. The fourth tip highlights the ability to use reference images via a dedicated text encoder. Additional tips advise being specific with prompts, using “In the style of” or “Painted by” for artist names, and employing “Cosplaying as” for character mixing. The user also recommends a realism enhancer LoRA and an uncensored realism engine LoRA for better realism.
These tips provide a comprehensive workflow for maximizing Krea 2’s capabilities in ComfyUI.
Key changes
- Use raw Krea 2 with turbo LoRA at 0.6 weight, 12 steps, Euler or beta/beta57 scheduler
- Employ Wan‑2.1 VAE for better quality, requiring Diffusers WF or ComfyUI‑VAE‑Utils
- Add uncensor LoRA to improve expression and prompt responsiveness
- Enable reference image usage via dedicated text encoder
- Specify prompts with artist names or “Cosplaying as” for better character mixing
- Use realism enhancer and uncensored realism engine LoRAs for higher realism