Summary
Fahd Mirza installs and systematically tests the Krea 2 Turbo, a 12-billion-parameter open-weights diffusion transformer from Krea AI, running the FP8-quantized variant through ComfyUI on hardware consuming over 30GB of VRAM. The video serves as both an installation guide and an honest capability assessment across a wide range of image generation challenges.
The setup walkthrough covers three required downloads: the main FP8 safetensors model weights (placed in the ComfyUI unet directory), a CLIP text encoder (4B parameter version recommended), and a VAE for converting latent space outputs to pixels. Mirza also shares his modified ComfyUI workflow file on GitHub. The ComfyUI graph structure is explained clearly for newcomers: model, text encoder, and VAE feed into a KSampler node, with output piped to a preview node.
Test prompts span portrait photography (strong lighting and depth, but three-finger hand artifacts consistent with diffusion model weaknesses at turbo speed), extreme close-up macro photography of a dew-covered spider web (impressive bead reflections and detail), cinematic coastal road landscape, anime-style character illustration (clean line work, expressive eyes), 3D reflective liquid metal rendering, and a challenging multilingual text rendering task using an Indonesian warung restaurant sign with multiple text elements — which the model handles surprisingly well, correctly rendering “warung makan” and menu items. The overall verdict is favorable: fast generation, consistent lighting, and stronger text rendering than expected, with hand anatomy as the main recurring weakness attributable to the turbo speed tradeoff.
📺 Source: Fahd Mirza · Published June 28, 2026
🏷️ Format: Tutorial Demo
