Krea 2 Turbo in ComfyUI: setup and official LoRA tests

This tutorial sets up Krea 2 Turbo in ComfyUI and tests the base model alongside eight official style LoRAs. The speaker describes Krea 2 as open source and contrasts the raw model's CFG 3 and roughly 52 sampling steps with Turbo's CFG 1 and typical eight steps. The tests focus entirely on Turbo.

The setup covers BF16 and FP8 Safetensors files, plus GGUF quantizations for limited VRAM. It requires the Qwen3-VL 4B text encoder and Qwen Image VAE. The workflow separates model loading, resolution and format controls, positive prompts, and sampling. The speaker says a negative prompt is generally unnecessary for Turbo. The supplied description recommends the latest ComfyUI nightly build and reports testing BF16 on Linux with an RTX 3060 12 GB, 48 GB of system RAM, and a 1920x1080 target resolution. These are the test specifications, rather than stated minimum requirements.

The examples examine anatomy, wide-angle poses and changes in visual style. One rainy-window scene tests readable "US Open" text on a tennis ball; other examples use crayon drawing, retro anime, watercolor and ink wash effects. The speaker rates the results positively, including prompt adherence and detail preservation. Those judgments describe the shown examples, rather than establishing reliability across other prompts or hardware.