GGUF quantized weights have been released for the high-performance Krea 2 (v1.0) model, built on the Diffusion Transformer (DiT) architecture with 12–13 billion parameters. Now, running image generation of this caliber is possible on consumer hardware with limited VRAM via ComfyUI.
What Happened
Both Krea 2 Base and the optimized Krea 2 Turbo are available in GGUF format. To work fully via ComfyUI, specific custom nodes (ComfyUI-GGUF_KREA-2) are required, along with auxiliary components: the Qwen3-VL-4B model and a VAE from Qwen.
Context
The Diffusion Transformer (DiT) architecture provides high generation quality but requires significant computational resources. Using the GGUF format allows heavy models to be moved from server-grade GPUs to home PCs, substantially reducing VRAM requirements without critical loss of quality.
Why It Matters for the Industry
The transition of large DiT models to the GGUF format expands the possibilities for local use of Krea.ai-level neural networks. This lowers the barrier to entry for developers and researchers, undermining the monopoly of cloud APIs and allowing high-quality generation to be integrated into local applications and autonomous pipelines without dependence on paid services.
Why It Matters for Users
Users with home PCs gain access to generation quality comparable to professional cloud services. Quantized versions allow for VRAM savings, making the use of powerful neural networks more accessible and private, while also reducing the need for expensive subscriptions.
What Is Not Yet Known / Limitations
Operation requires the installation of additional custom nodes and auxiliary models, which complicates the initial setup of the working environment.
Sources
Author
Look at AI, Editorial Team