Join The Tinkerer on Whop. Membership gets you early releases, private tools and a bunch of extra stuff.

👉 Join on Whop
💬 Join the community for support, free tools and early news on Discord

What’s New in V3.5

Version 3.5 keeps the proven generation pipeline from V3.0, but introduces the new Cyberdelia Prompt Controls — Presets node.

This preset-aware control node brings the main prompt options together in one place: enter your prompt, choose between engineered (LLM) and passthrough (RAW), enable or disable Vision, and select the system-prompt preset used by Prompt Engineer.

The preset selector automatically follows the selected input type:

• With Vision disabled, it shows the normal prompt-rewriting presets.

• With Vision enabled, it switches to the dedicated image-to-prompt presets.

• The complete selected system prompt is stored inside the workflow, keeping shared workflows reproducible even if an external preset is later changed.

• Passthrough mode continues to work without LM Studio or another local LLM.

This makes the workflow a natural companion to the new Krea 2 prompt-preset library. You can use presets to turn the same starting idea into photographic realism, painterly artwork, editorial illustration, graphic-novel artwork, studio portraits, and many other targeted styles.

🆓 Preset guide and free starter selection:

https://civarchive.com/articles/35214/turn-z-engineer-into-a-prompt-style-library-custom-presets-for-krea-2-and-z-image-turbo

📦 Complete Krea 2 and Z-Image Turbo preset collections (for Whop members):

https://whop.com/cyberdelia-ai-lab/exp_Q7Jc5e2sjBWTae/app/

A clean and practical CyberRealistic Krea 2 workflow for ComfyUI, built around the CyberKrea Sampler and designed to keep the main generation process simple while still offering useful extras such as Face Detailer, latent upscaling, LoRA support, prompt engineering, comparison previews, and improved saved-image metadata.

CyberRealistic Krea 2 - ComfyUI Workflow V3

The workflow keeps the parts you use most close together: model settings, prompt input, generation, previews, and saving. Optional modules stay out of the way until you need them.

Face Detailer and Upscaler can be enabled or disabled from the module toggle panel. Both are disabled by default, so the workflow starts with a straightforward Krea 2 generation.

What's included in V3

• CyberKrea Sampler. The main generation pass uses the CyberKrea Sampler with Fast, Balanced, and Quality presets. Balanced is the default, while the more advanced sampler controls remain available when you want to fine-tune the result.

• Krea 2 resolution presets. The CyberKrea Empty Latent node provides practical aspect ratios and resolutions without having to enter dimensions manually.

• NegPiP prompt encoding. Positive and negative prompting is handled through the dedicated NegPiP node, including support for negative prompt weighting while keeping the Krea 2 conditioning pipeline intact.

• Optional Prompt Engineer. Rewrite simple ideas or existing prompts into natural Krea 2 prompts using a local OpenAI-compatible LLM. Vision mode can also analyze a reference image and turn it into a Krea 2 prompt. Normal generation does not require a local LLM; keep the Prompt Engineer in passthrough (raw) mode to use your prompt directly.

• LoRA loading. An rgthree Power LoRA Loader is included for quickly adding and combining LoRAs without rebuilding the model chain.

• Krea 2 Turbo Enhancer. The optional Krea2T Enhancer is included in the model section and disabled by default.

• Face Detailer. Optional YOLO face detection with SAM-assisted masking and Impact Pack Face Detailer. Useful for repairing or refining smaller faces without affecting the complete image.

• SesquiLSR latent upscale. The optional Upscaler uses SesquiLSR to enlarge the latent, followed by a lightweight second Krea 2 sampling pass. No separate ESRGAN-style upscale model is required.

• Comparison previews. Dedicated comparison nodes make it easy to check the original first pass against the Face Detailer and upscale results.

• Improved image metadata. Final images are saved through Cyberdelia Save Image With Metadata for cleaner generation metadata and easier sharing.

Face Detailer and Upscaler are disabled by default. Enable only the modules you need and the final Save Image node follows the resulting image through the chain.

Prompt Engineer Settings

The Prompt Engineer is kept compact in the main workflow. To access its advanced settings, click the small open subgraph icon in the top-right corner of the 🧠 Prompt Engineer node.

Inside the subgraph you can configure the local LLM connection and generation settings, including API URL, model, temperature, max tokens, timeout, retries, output cleanup, and Vision settings.

The default API endpoint is:

http://localhost:1234/v1

This is intended for a local OpenAI-compatible server such as LM Studio. The model can normally remain set to auto.

You do not need a local LLM to use the workflow. With Prompt Engineer set to passthrough (raw), your prompt is sent directly to Krea 2 without calling the LLM.

For image-to-prompt generation, enable Vision and connect an image to the Prompt Engineer image input.

Requirements

• A recent ComfyUI build with subgraph support

• Krea 2 Model

• qwen3vl_4b_bf16.safetensors or qwen3vl_4b_fp8_scaled.safetensors Krea 2 text encoder

• qwen_image_vae.safetensors

• bbox/face_yolov8m.pt for Face Detailer detection

• sam_vit_b_01ec64.pth for SAM-assisted Face Detailer masking

• A local OpenAI-compatible LLM server only if you want to use the optional Prompt Engineer

Nodes / Custom Nodes Used

• rgthree-comfy
  https://github.com/rgthree/rgthree-comfy

• CyberKrea Sampler
  https://github.com/cyberdeliaAI/ComfyUI-CyberKrea-Sampler

• Cyberdelia Z-Engineer
  https://github.com/cyberdeliaAI/comfyui-cyberdelia-z-engineer

• Cyberdelia Metadata / Save Image With Metadata
  https://github.com/cyberdeliaAI/comfyui-cyberdelia-metadata

• NegPiP for Z-Image / Krea 2
  https://github.com/cyberdeliaAI/comfyui-negpip-zimage

• ComfyUI Impact Pack
  https://github.com/ltdrdata/ComfyUI-Impact-Pack

• ComfyUI Impact Subpack
  https://github.com/ltdrdata/ComfyUI-Impact-Subpack

• Krea 2 Turbo Enhancer
  https://github.com/capitan01R/ComfyUI-Krea2T-Enhancer

• SesquiLSR
  https://github.com/LoganBooker/SesquiLSR

Prompt Engineer note: the workflow works normally with Prompt Engineer in passthrough mode. A local LLM is only required when you enable prompt rewriting or vision-based prompt generation.