ComfyUI Krea 2 setup and text-to-image workflows

This tutorial covers Krea 2 text-to-image generation in ComfyUI, starting with updates to Easy Install, ComfyUI and Pixaroma nodes. The setup uses the FP8 diffusion model, a Qwen text encoder and Qwen image VAE. The speaker describes Krea 2 as a generation model rather than an image editor and prefers FP8 to GGUF after finding GGUF slower in his tests.

The basic workflow uses eight sampling steps and CFG 1, without negative prompts. Further examples add LoRA styles and use the text encoder to expand short prompts. The speaker shows how changing the image seed produces another image with the same generated prompt, while changing the text-generation seed requests a new prompt. Prompt expansion also fails during some examples, so the generated text needs checking.

For local AI users with limited GPU memory, the tutorial includes a tiled VAE decode workflow. The speaker reports running the roughly 12 GB model on a 6 GB RTX 2060, with 64 GB of system RAM, and recommends trying smaller image sizes first. Those results describe his setup rather than a guaranteed minimum requirement.

The extra-pass workflow applies a 1.5× latent upscale, followed by four sampling steps at roughly 0.4 to 0.5 denoise. A separate workflow targets 2K output. Style tests include photography and watercolor, but also expose lettering errors, plastic-looking details and mistakes in object placement.