ILLUMINATE AI — KREA 2 TURBO

INDEPENDENT CREATOR + EDIT STUDIO v1.0

A complete modular generation and image-editing environment built specifically for Krea 2 Turbo inside ComfyUI.

This major update rebuilds the workflow around a new independent architecture. CREATE, Native Edit Original, Native Edit Custom and Classic Img2Img now operate as separate image pipelines with their own decoding, previews, SeedVR2 processing and final save outputs.

The workflow also introduces the new ILLUMINATE KREA PROMPT DIRECTOR, direct Ollama integration, dedicated prompt profiles and the included ComfyUI-ILLUMINATE-AI-KREA custom-node package.

The result is a clearer, safer and more flexible workflow that supports fast native Krea generation, precise instruction-based editing, custom aspect-ratio conversion and creative denoise-based Img2Img without forcing every mode through one shared image route.

Hotfix v1.0.1: Fixed an issue that could cause colorful or pixelated results when using the Face, Hand and other Detailers. The original Detailer group-control configuration has been restored. Please replace the previous workflow and Custom Node Pack, then restart ComfyUI.

MAJOR UPDATE — INDEPENDENT ARCHITECTURE

The entire workflow has been reorganized into four independent working modes:

• CREATE

• NATIVE EDIT — ORIGINAL

• NATIVE EDIT — CUSTOM

• CLASSIC IMG2IMG

Each mode keeps its own local image pipeline.

There is no shared image router connecting all four outputs. This prevents unused branches from interfering with the active mode and reduces required-input errors caused by inactive Native Edit, SeedVR2 or Img2Img sections.

Each image branch now has its own:

• sampler or edit sampler

• VAE decode

• base-result preview

• optional SeedVR2 pipeline

• final preview

• final save output

• local model and conditioning route where required

Only activate the image branch you currently want to use.

Unused image branches should remain on Bypass.

NEW — ILLUMINATE KREA MODE SELECTOR

The new ILLUMINATE KREA MODE SELECTOR provides a central prompt-profile selection for the workflow.

Available modes:

• CREATE

• NATIVE EDIT — ORIGINAL

• NATIVE EDIT — CUSTOM

• CLASSIC IMG2IMG

The selector changes the prompt behavior used by the Prompt Director.

It intentionally does not activate or connect the image pipelines automatically. The selected prompt mode and the image branch enabled through the corresponding Fast Groups Bypasser must match.

This design keeps the image architecture independent and avoids ComfyUI validating unused image branches.

NEW — ILLUMINATE KREA PROMPT DIRECTOR

The new ILLUMINATE KREA PROMPT DIRECTOR combines the complete Manual and Assisted prompting system into one custom node.

It includes:

• Manual prompting

• Ollama-assisted prompt generation

• dedicated prompt profiles for each workflow mode

• direct local Ollama communication

• automatic Ollama model unloading

• exact final prompt preview

• word and character counter

• copy button

• generation status display

• one shared final text output for the active prompt profile

The preview and copy functions do not trigger a second Ollama request.

MANUAL MODE

Use MANUAL when you already have a complete prompt or editing instruction.

In MANUAL mode:

• the text is passed through unchanged

• Ollama is not required

• no Ollama client is created

• no local network request is made

• no additional prompt interpretation is applied

This is the fastest and most predictable mode.

ASSISTED MODE

Use ASSISTED when you want to enter keywords, a short concept or a rough idea.

The Prompt Director sends the idea once to the selected local Ollama model and creates a complete prompt optimized for the selected workflow mode.

The included profiles are not identical.

CREATE profile

Produces a complete image-generation prompt designed for Krea 2 Turbo.

NATIVE EDIT profile

Produces a focused editing instruction that prioritizes preservation of:

• identity

• pose

• composition

• perspective

• anatomy

• camera position

• relevant source-image structure

CLASSIC IMG2IMG profile

Produces a target-image description intended for the visible denoise-controlled Img2Img sampler.

When unload-after-generation is enabled, the Ollama request uses keep_alive: 0. This releases the language model before Krea image generation begins and prevents Ollama from unnecessarily occupying memory during the image pipeline.

The separate comfyui-ollama prompt nodes are not required.

MAIN CREATOR

The CREATE branch is designed around Krea 2 Turbo’s intended fast inference behavior.

Recommended base configuration:

• 8 steps

• CFG 1.0

• Euler sampler

• Simple scheduler

• native Krea-compatible resolution

• modular optional processing

The Main Creator includes:

• dedicated Krea 2 model loader

• Qwen3-VL text encoder loader

• Qwen Image VAE loader

• Power LoRA Loader with MODEL and CLIP routing

• optional negative prompt switcher

• native base generation

• separate Base Preview

• optional Mild Quality Refinement

• separate Refinement Preview

• optional Concept Batch

• optional Detailer Suite

• independent SeedVR2 finishing stage

• dedicated final preview

• dedicated final save output

BASE AND REFINEMENT ARE NOW CLEARLY SEPARATED

The Main Creator now provides two direct review stages:

BASE PREVIEW

Displays the result directly after the first Krea sampler and VAE decode.

REFINEMENT PREVIEW

Displays the result after the optional Mild Quality Refinement stage.

This makes it possible to compare the original Krea result with the refined result before continuing into the Detailer Suite or SeedVR2.

Refinement is optional.

Krea 2 Turbo can already produce strong results without an additional sampling stage. Enable refinement only when it produces a visible improvement for the current image.

OPTIONAL CONCEPT BATCH

The Concept Batch module can generate four variations of the same concept for comparison.

Use it to:

• explore different seeds

• compare compositions

• test prompt interpretation

• select a stronger base concept before additional processing

The Concept Batch is optional and bypassed by default.

COMPLETE OPTIONAL DETAILER SUITE

The workflow includes five independent Detailer modules:

• Face Detailer

• Hand Detailer

• Penis Detailer

• Vagina Detailer

• Female Breast Detailer

Each Detailer contains:

• its own positive prompt

• its own negative prompt

• a dedicated BBOX detector

• an independent enable or bypass control

• Krea-safe sampling settings

• a separate review preview after processing

Default Detailer configuration:

• 8 steps

• CFG 1.0

• Euler

• Simple

The available Detailer chain is:

Base Generation
→ Optional Quality Refinement
→ Face Detailer
→ Hand Detailer
→ Penis Detailer
→ Vagina Detailer
→ Female Breast Detailer
→ Optional SeedVR2
→ Final Preview
→ Final Save

Every Detailer works independently.

You can enable only Face, only Hands or any combination required for the current image. All optional Detailer groups are bypassed by default.

For SFW use, only the Face and Hand detectors are necessary.

REQUIRED DETAILER DETECTORS

The workflow uses dedicated detector models for the Detailer Suite:

• face_yolov8m.pt

• hand_yolov8s.pt

• penis.pt

• pussy.pt

• female_breast-v4.2.pt

Installation information is included inside the workflow.

Detector models are not included in the workflow ZIP and must be installed separately.

IMAGE EDIT STUDIO

The Image Edit Studio contains three independent editing systems:

• Native Edit — Original Target

• Native Edit — Custom Target

• Classic Img2Img

These modes are designed for different purposes and should not be treated as interchangeable.

NATIVE EDIT — ORIGINAL TARGET

Native Edit Original is intended for precise, instruction-based modifications while preserving the original image format.

It is suitable for:

• identity-preserving edits

• clothing changes

• lighting adjustments

• background changes

• object replacement

• controlled appearance changes

• better pose preservation

• better composition preservation

• targeted corrections

The branch uses:

• Krea2Edit Model Patch

• Grounded Edit Prompt

• Identity Edit LoRA

• Grounding Resolution control

• AuraFlow sampling

• visible denoise control

• optional creative LoRA support

• independent VAE decode

• independent SeedVR2 output

• separate preview and save output

Recommended starting settings:

• Grounding Resolution: 768

• Identity Edit LoRA: 1.0

• Steps: 8

• CFG: 1.0

• Sampler: Euler

• Scheduler: Beta

• Denoise: 1.0

Original Target Mode preserves the normalized resolution and aspect ratio of the source image.

Use this mode when the edited result should remain in the same general format as the original.

NATIVE EDIT — CUSTOM TARGET FORMAT

Native Edit Custom uses the original image as a grounded visual reference but generates the result at a new target width and height.

This allows you to convert an image into another format without simply stretching it.

Example formats:

• 1344 × 768 — 16:9

• 768 × 1344 — 9:16

• 1536 × 1024 — 3:2

• 1024 × 1536 — 2:3

• 1536 × 1536 — Square

The Custom Target branch includes:

• adjustable target width

• adjustable target height

• dedicated target latent

• separate Krea 2 Edit sampler

• independent decode

• base-result preview

• optional dedicated SeedVR2 upscale

• final preview

• separate final save output

The model generates the additional image areas based on the source image and edit instruction.

For best results, begin with moderate aspect-ratio changes.

Extreme changes require Krea 2 to generate significantly more new visual information and may alter framing, pose, composition or background structure.

CLASSIC IMG2IMG

Classic Img2Img is a traditional denoise-based branch intended for creative reinterpretation.

Use it for:

• stronger stylistic changes

• visual transformations

• creative reinterpretation

• checkpoint-driven restyling

• greater freedom from the source image

• denoise-controlled preservation

The branch includes:

• independent Krea 2 model stack

• dedicated positive prompt

• dedicated negative prompt

• optional negative conditioning

• separate Power LoRA Loader

• AuraFlow sampling shift

• normalized source resolution

• visible denoise control

• independent VAE encoding and decoding

• optional dedicated SeedVR2 upscale

• separate preview

• separate final save output

Recommended starting denoise:

• 0.50

Lower values preserve more of the source image.

Higher values allow stronger reinterpretation but may change anatomy, identity, pose and composition.

Use Classic Img2Img when creative freedom is more important than strict source preservation.

Use Native Edit when the result should remain closer to the original image.

INDEPENDENT SEEDVR2 OUTPUTS

SeedVR2 is available independently for:

• CREATE

• NATIVE EDIT — ORIGINAL

• NATIVE EDIT — CUSTOM

• CLASSIC IMG2IMG

Each branch has its own SeedVR2 models, image input, preview and final save path.

SeedVR2 should only be enabled after the base result already looks correct.

It may improve:

• final resolution

• perceived detail

• texture clarity

• edge definition

• overall presentation quality

Always compare the base image with the SeedVR2 result.

Small textures, skin details, eyes, facial features and other fine structures may be slightly reinterpreted during upscaling.

FAST GROUPS AND BRANCH CONTROL

The workflow uses Fast Groups to keep optional modules and image branches independent.

For each run:

1. Select the matching prompt profile.

2. Activate the image branch you want to use.

3. Keep all unrelated image branches bypassed.

4. Enable only the optional modules required for the current result.

5. Use the preview and save output belonging to the active branch.

For Native Edit, activate either:

• Original Target

• Custom Target Format

Avoid running both simultaneously unless you intentionally want both output formats.

Use Bypass for inactive modules.

Do not use Mute / Never for optional branches unless you understand how your current ComfyUI version validates disconnected required inputs.

INCLUDED CUSTOM NODES

The download package includes:

ComfyUI-ILLUMINATE-AI-KREA

The included node pack provides:

• ILLUMINATE KREA MODE SELECTOR

• ILLUMINATE KREA PROMPT DIRECTOR

• direct Ollama communication

• Manual and Assisted prompt modes

• dedicated prompt profiles

• automatic Ollama unloading

• final prompt preview

• word and character count

• prompt copy function

• compatibility support for earlier ILLUMINATE Krea test workflows

No separate Python dependency installation is required for the included ILLUMINATE node pack.

Copy the complete folder into:

ComfyUI/custom_nodes/

Then restart ComfyUI completely.

Do not merge the files into an older installation. Replace the previous ComfyUI-ILLUMINATE-AI-KREA folder completely.

A full English installation guide is included in the download package.

REQUIRED MODELS

The workflow requires compatible versions of:

• Krea 2 Turbo diffusion model

• Qwen3-VL text encoder

• Qwen Image VAE

• Krea 2 Identity Edit LoRA

• SeedVR2 DiT model when SeedVR2 is enabled

• SeedVR2 VAE model when SeedVR2 is enabled

• required BBOX detector models for enabled Detailers

• optional creative LoRAs selected by the user

Large model files are not included in the workflow ZIP.

If a loader contains a filename that does not exist on your system, select the compatible model installed locally.

Optional LoRAs that are not installed must be disabled or replaced before queueing the affected branch.

REQUIRED CUSTOM NODE PACKS

Depending on the enabled modules, the workflow uses nodes from:

• rgthree-comfy

• comfyui-krea2edit

• ComfyUI-SeedVR2_VideoUpscaler

• ComfyUI-Impact-Pack

• ComfyUI-Impact-Subpack

• ComfyUI-Easy-Use

• ComfyUI Essentials

• ComfyUI-Custom-Scripts

• TinyTerraNodes

Use:

ComfyUI Manager → Install Missing Custom Nodes

After installing Python custom nodes, close ComfyUI completely and restart it. Refreshing only the browser is not sufficient.

OLLAMA SETUP

Ollama is required only for ASSISTED prompt generation.

MANUAL mode works without Ollama.

Recommended local model:

fredrezones55/Qwen3.5-Uncensored-HauhauCS-Aggressive:9b

Install it with:

ollama pull fredrezones55/Qwen3.5-Uncensored-HauhauCS-Aggressive:9b

Default Ollama address:

http://127.0.0.1:11434

The Prompt Director communicates directly with the local Ollama API.

The separate comfyui-ollama custom nodes are not required.

HARDWARE AND COMPATIBILITY

Created and tested with Windows Portable ComfyUI on:

• NVIDIA RTX 5090

• 32 GB VRAM

• 64 GB system RAM

The workflow is not locked to this hardware.

Actual compatibility, generation speed and usable output resolution depend on:

• available VRAM

• system RAM

• selected model precision

• selected image dimensions

• active Detailers

• active LoRAs

• SeedVR2 settings

• number of simultaneously enabled branches

• local ComfyUI configuration

Update ComfyUI before loading the workflow because Krea 2 and related custom nodes may require recent ComfyUI functionality.

WORKFLOW PHILOSOPHY

The purpose of this workflow is not to force every available module into every image.

Krea 2 Turbo is already designed for fast, high-quality generation.

The workflow provides optional tools, but they should only be activated when they create a visible improvement.

The workflow prioritizes:

• fast Krea 2 generation

• independent and stable image branches

• clear mode separation

• Manual and Assisted prompting

• direct local Ollama integration

• predictable branch control

• strong LoRA compatibility

• separate Base and Refinement review

• creative Img2Img

• precise instruction-based Native Edit

• identity and composition preservation

• custom output formats

• modular Detailer processing

• independent SeedVR2 finishing

• separate previews and final save outputs

VERSION

ILLUMINATE AI — KREA 2 TURBO

INDEPENDENT CREATOR + EDIT STUDIO v3.0

Major Update Highlights

• Complete independent-interface rebuild

• Four separate workflow modes

• New ILLUMINATE KREA MODE SELECTOR

• New ILLUMINATE KREA PROMPT DIRECTOR

• Manual mode with zero Ollama requests

• Assisted mode with one Ollama request

• Automatic Ollama model unloading

• Dedicated CREATE prompt profile

• Dedicated NATIVE EDIT prompt profile

• Dedicated CLASSIC IMG2IMG prompt profile

• Exact final prompt preview and copy function

• Independent CREATE image pipeline

• Independent Native Edit Original pipeline

• Independent Native Edit Custom pipeline

• Independent Classic Img2Img pipeline

• No shared final image router

• Separate SeedVR2 path for every image branch

• Separate previews and save outputs

• Base and Refinement previews clearly separated

• Optional four-image Concept Batch

• Complete five-stage Detailer Suite

• Custom target width and height control

• Native aspect-ratio conversion

• Improved inactive-branch stability

• Included ILLUMINATE Krea custom-node package

• Complete installation guide included