CyberRealistic Krea 2 - ComfyUI Workflow — v1.0 — v2.5 (Prompt Engineer)

Use this ComfyUI workflow to create Krea 2 images from text prompts, optionally rewrite prompts with an LLM or create prompts from a reference image, and save the resulting image.

Optional Face Detailer and Upscaler steps can refine faces or enlarge the latent image; comparison previews show these results alongside the first pass.

Enter positive and negative prompts, including negative prompt weighting; for image-to-prompt generation, enable Vision and connect a reference image.

The final image is saved with generation metadata.

Use a recent ComfyUI build with subgraph support, a Krea 2 model, qwen_image_vae.safetensors, and either qwen3vl_4b_bf16.safetensors or qwen3vl_4b_fp8_scaled.safetensors.

For Face Detailer, provide bbox/face_yolov8m.pt for face detection and sam_vit_b_01ec64.pth for masking.

Listed ComfyUI add-ons include rgthree-comfy, CyberKrea Sampler, Cyberdelia Z-Engineer, and Cyberdelia Metadata / Save Image With Metadata.

Other listed add-ons are NegPiP for Z-Image / Krea 2, ComfyUI Impact Pack, ComfyUI Impact Subpack, Krea 2 Turbo Enhancer, and SesquiLSR.

Prompt Engineer settings include API URL, model, temperature, max tokens, timeout, retries, output cleanup, and Vision; the documented default API URL is http://localhost:1234/v1.

Face Detailer and Upscaler are off by default; enable either from the module toggles when you want to use it.

Passthrough sends your prompt directly without a local LLM; prompt rewriting and Vision-based prompt creation require a local OpenAI-compatible LLM server.