CyberRealistic Krea 2 - ComfyUI Workflow

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

CyberRealistic Krea 2 Workflow V2.5

Version 2.5 builds on the existing 2.0 workflow and adds the Cyberdelia Prompt Engineer node for optional LLM prompt enhancement, raw prompt passthrough, and Vision-based image-to-prompt generation.

For installation, setup, and a complete explanation of all options, see the Cyberdelia Prompt Engineer Complete Guide.

CyberRealistic Krea 2 Workflow V2.0

A clean and practical ComfyUI workflow built specifically for CyberRealistic Krea 2.

V2.0 is a major rebuild of the sampling core. The layout, module toggles and general feel of V1 are still there, but the main generation path is now shorter, cleaner and built around a single sampler.

The separate second pass is gone, RES4LYF is no longer required, resolution selection is easier, and the new sampler gives you simple presets while still allowing full manual control.

Workflow Files

The download includes both the JSON workflow and a PNG workflow image.

Use the PNG version if you want ComfyUI to restore the original node layout, group positions and visual formatting.

What Changed in V2.0

New Sampler: CyberKrea Sampler

ClownsharKSampler_Beta and the RES4LYF sampling chain are gone.

Sampling now runs through CyberKrea Sampler, a compact sampler designed specifically around Krea 2 Turbo.

It includes three calibrated presets:

• fast

• balanced

• quality

Selecting a preset automatically fills in the sampler settings. After that, you can still adjust any individual value manually.

The sampler exposes controls for:

• restart_frac and sigma_r — internal restart/detail recovery

• plunge — locks the main composition before the restart focuses more on texture and detail

• detail — adjusts fine-detail emphasis

• eta0 and sigma_gate — gated ancestral noise

• contraction — pulls the starting noise toward the photo manifold

• selectable live preview methods

The restart stage is handled inside the sampler instead of requiring a completely separate second sampling chain.

The optional VAE input can also decode a preview thumbnail directly on the sampler node without affecting the LATENT output.

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

New Latent Node: CyberKrea Empty Latent

No more manually typing width and height.

Choose a size tier:

• S

• M

• L

• XL

The resolution dropdown then filters down to matching dimensions and aspect ratios.

The workflow ships with:

XL — 1184 × 1776 (2:3)

The 2nd Pass Group Is Gone

V1 used a separate optional second sampler and latent path for additional detail recovery.

V2.0 moves that functionality into the restart stage of CyberKrea Sampler.

This removes the extra sampler, extra latent hop and another set of settings that previously needed to stay synchronized.

The result is a simpler generation path with fewer nodes and fewer opportunities for settings or seeds to drift between passes.

RES4LYF Is No Longer Required

V1 used RES4LYF for ClownsharKSampler_Beta.

V2.0 does not use RES4LYF anywhere in the workflow.

You can leave it installed if you use it for other workflows, but CyberRealistic Krea 2 Workflow V2.0 does not require it.

Negative Prompt Handling

The negative conditioning input on CyberKrea Sampler is intentionally left disconnected.

The workflow already uses NegPiP, which merges the positive and negative prompt strings and converts the negative side into negative weights before sampling.

So nothing is missing when the sampler's separate negative input is empty.

You can still connect it manually if you want to experiment with traditional negative conditioning / sigma-window guidance.

Krea 2 Turbo Enhancer Included

ComfyUI-Krea2T-Enhancer is wired directly between the model loader and the LoRA stack.

It is included but disabled by default:

• enabled: false

• strength: 0.3

Enable it if you want to experiment with stronger prompt adherence.

Leave it disabled for the normal CyberRealistic Krea 2 behaviour.

The node is still considered experimental by its author.

Face Detailer Now Runs Before Upscale

The Face Detailer now works on the image at its original generation resolution.

The corrected result is then passed into SeedVR2, allowing the upscale stage to carry the improved face forward instead of trying to repair an already-upscaled image.

Both comparison nodes remain available so you can inspect:

• before / after Face Detailer

• before / after Upscale

Save Node Renamed

The metadata saver is now:

CyberdeliaSaveImageWithMetaData

Make sure comfyui-cyberdelia-metadata is updated before loading the workflow.

Older versions use the previous node name and may show the save node as missing/red.

Custom Nodes Used

The full workflow references nine custom-node packs.

The core generation workflow only uses part of this list. SeedVR2 and Impact Pack/Subpack belong to the optional Upscale and Face Detailer modules.

ComfyUI-CyberKrea-Sampler

Nodes:

• CyberKreaSampler

• CyberKreaEmptyLatent

Krea 2 Turbo sampler with presets, restart-based detail recovery, manual tuning controls and a resolution-aware empty latent node.

New in V2.0.

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

rgthree-comfy

Nodes used include:

• Context Big

• Seed

• Power Lora Loader

• Image Comparer

• Fast Groups Bypasser

• Label

Handles much of the workflow wiring and UI, including context bundles, LoRA loading, before/after comparisons and module toggles.

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

comfyui-negpip-zimage

Node:

ZImageNegPipPrompt

NegPiP-style negative prompt weighting.

It combines the positive and negative prompt strings and converts the negative side into negative weights before sampling.

https://github.com/cyberdeliaAI/comfyui-negpip-zimage

ComfyUI-Krea2T-Enhancer

Node:

ComfyUI-Krea2T-Enhancer

Patches the Krea 2 text-fusion path during sampling to improve prompt adherence.

One MODEL in, one MODEL out.

Included in the workflow but disabled by default.

https://github.com/capitan01R/ComfyUI-Krea2T-Enhancer

comfyui-cyberdelia-metadata

Node:

CyberdeliaSaveImageWithMetaData

A replacement for the standard Save Image node that writes clean A1111/Forge-compatible metadata into PNG files, including Civitai compatibility fixes.

https://github.com/cyberdeliaAI/comfyui-cyberdelia-metadata

Optional Upscale Module

ComfyUI-SeedVR2_VideoUpscaler

Nodes:

• SeedVR2LoadDiTModel

• SeedVR2LoadVAEModel

• SeedVR2VideoUpscaler

Diffusion-based upscaler for both still images and video.

The required SeedVR2 models can be downloaded automatically to:

ComfyUI/models/SEEDVR2

on first use.

https://github.com/numz/ComfyUI-SeedVR2_VideoUpscaler

comfyui-vrgamedevgirl

Node:

FastUnsharpSharpen

Fast unsharp masking used after SeedVR2 to restore some edge definition after the upscale.

Part of a larger video enhancement node pack.

https://github.com/vrgamegirl19/comfyui-vrgamedevgirl

Optional Face Detailer Module

ComfyUI-Impact-Pack

Nodes:

• FaceDetailer

• SAMLoader

Detection and detailing toolkit.

FaceDetailer detects the face, resamples the detected area and composites the result back into the image.

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

ComfyUI-Impact-Subpack

Node:

UltralyticsDetectorProvider

Separate installation that provides the YOLO detector required by the Face Detailer bbox input.

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

Notes

Impact Pack and Impact Subpack are two separate installations.

Installing only Impact Pack will leave the YOLO face detector unavailable.

The Upscale and Face Detailer groups are bypassed by default.

The main CyberRealistic Krea 2 workflow therefore runs without the SeedVR2, SAM or YOLO model files being present.

Enable the optional modules you want through the Enable/Disable toggle node in the Models & Settings section.

Previous Version — Workflow V1.0

V1.0 is the original CyberRealistic Krea 2 ComfyUI workflow and remains available for users who prefer the previous sampling setup or want to reproduce older results.

The overall workflow concept is similar to V2.0, with NegPiP prompting, LoRA support, optional Face Detailer, SeedVR2 upscale and metadata saving.

The main differences are in the sampling path.

V1.0 Uses RES4LYF

V1.0 uses:

ClownsharKSampler_Beta

from RES4LYF.

RES4LYF is therefore required when using the original V1 workflow.

https://github.com/ClownsharkBatwing/RES4LYF

Separate 2nd Pass

V1 includes an optional 2nd Pass group with its own sampler and latent path.

This was used for additional detail recovery after the initial generation.

V2.0 replaces this with the internal restart system of CyberKrea Sampler, which is why the separate second-pass group is no longer needed.

Older Metadata Save Node

V1 uses the older:

SaveImageWithMetaData

node name.

V2.0 uses:

CyberdeliaSaveImageWithMetaData

If you load V1 with a newer version of comfyui-cyberdelia-metadata, keep this difference in mind.

V1 Workflow Files

Both the original JSON and PNG workflow remain available.

Use the PNG version if you want the original V1 node layout, group positions and visual formatting preserved.

For new setups, I recommend starting with Workflow V2.0.