ILLUMINATE AI — KREA 2 TURBO MODULAR WORKFLOW

A clean, modular and quality-focused ComfyUI workflow designed specifically for Krea 2 Turbo.

This workflow does not treat Krea 2 Turbo like SDXL. The main generation pipeline is built around Krea’s intended fast inference settings, with optional modules for concept exploration, refinement, face restoration, LoRA support, image-to-image generation and final SeedVR2 upscaling.

The main branch works independently, while additional modules can be enabled only when needed.

Main Features

Krea 2 Turbo Main Generation

Optimized for fast Krea 2 Turbo inference:

• 8 sampling steps

• CFG 1.0

• Euler sampler

• Simple scheduler

• 1024 × 1024 default resolution

• Separate preview and save outputs

• Dedicated Krea 2 model, Qwen3-VL text encoder and Qwen Image VAE loaders

Krea 2 Turbo is a distilled model, so the workflow uses Zero Negative Conditioning by default instead of forcing a large SDXL-style negative prompt.

Optional Negative Prompt Switcher

The workflow includes an optional negative text prompt together with a centralized conditioning switcher.

Default priority:

1. Krea Turbo Zero Negative

2. Optional Negative Text Prompt

To use the negative text prompt:

1. Enable the Optional Negative Text Prompt node.

2. Mute the Krea Turbo Zero Negative node with Ctrl + M.

3. The switcher automatically sends the text-based negative conditioning to all supported branches.

The negative prompt should remain short and targeted. Large SDXL-style negative prompt lists are not recommended for Krea 2 Turbo.

Power LoRA Loader

The integrated Power Lora Loader from rgthree supports multiple LoRAs and passes both:

• MODEL

• CLIP

This allows compatible Krea 2 LoRAs to influence the model and text-encoder paths correctly.

Place your LoRAs inside:

ComfyUI/models/loras/

Some LoRAs may require their trigger word to be added manually to the positive prompt.

Start with moderate LoRA strengths around 0.6–0.9 and adjust according to the individual LoRA.

Optional Concept Batch

Generate four quick variations before committing to a final image.

• Batch size: 4

• Fast concept generation

• Separate preview for all four images

• Select one latent with Latent From Batch

Batch index:

0 = Image 1
1 = Image 2
2 = Image 3
3 = Image 4

Keep the batch length at 1.

Optional Quality Refinement

A mild secondary refinement stage designed to improve texture and small details without behaving like a heavy SDXL HiRes pass.

Recommended settings:

• Low denoise

• Approximate range: 0.12–0.20

• Identity, composition and colors should remain mostly intact

Higher denoise values may introduce unnecessary changes.

Optional Face Detailer

Designed for person-focused generations.

The Face Detailer uses:

• A dedicated face detector

• Gentle Krea-compatible settings

• Low denoise

• CFG 1.0

• Euler and Simple scheduler

The Face Detailer input automatically prefers the refined image when Quality Refinement is active.

Input priority:

1. Refined image

2. Main generated image

This prevents the Face Detailer from ignoring the refinement stage.

Final SeedVR2 Upscale

SeedVR2 is positioned at the actual end of the image pipeline.

The final image selector automatically chooses the most advanced available image before upscaling.

Final output priority:

1. Face Detailer result

2. Quality Refinement result

3. Main Krea 2 Turbo result

This ensures SeedVR2 does not upscale the original image before the optional detail stages have been completed.

The normal ESRGAN-style upscale branch has been removed in favor of SeedVR2.

Torch Compile remains disabled by default for better stability with Windows Portable ComfyUI.

Optional Image-to-Image

A separate Krea 2 Turbo image-to-image branch is included.

Recommended denoise range:

0.25–0.40

Lower values preserve more of the source image. Higher values introduce stronger changes to identity, composition and style.

Required Models

Place the files in the corresponding ComfyUI folders.

Diffusion Model

ComfyUI/models/diffusion_models/

Example:

krea2_turbo_fp8_scaled.safetensors

Compatible Krea 2 Turbo checkpoints can also be selected manually.

Text Encoder

ComfyUI/models/text_encoders/

Example:

qwen3vl_4b_fp8_scaled.safetensors

or:

qwen3vl_4b_bf16.safetensors

The CLIP loader type must support:

krea2

VAE

ComfyUI/models/vae/

Required:

qwen_image_vae.safetensors

SeedVR2 Models

Install the required SeedVR2 DiT and VAE models according to your SeedVR2 node package.

Example model files used in the workflow:

seedvr2_ema_7b_sharp_fp16.safetensors
ema_vae_fp16.safetensors

Required Custom Nodes

Depending on which optional modules you use:

• rgthree-comfy

• ComfyUI Impact Pack

• Impact Subpack

• SeedVR2 nodes

The main Krea 2 Turbo generation branch does not require every optional module to be enabled.

If the Power Lora Loader or Any Switch nodes display incorrectly, disable the experimental ComfyUI Nodes 2.0 interface and restart ComfyUI.

Created and tested on Windows Portable ComfyUI using an NVIDIA RTX 5090 with 64 GB VRAM.

The workflow is not limited to this GPU. Performance, supported resolutions and optional modules depend on the available VRAM, system configuration and installed models.

Important Notes

• Update ComfyUI before loading the workflow.

• Older ComfyUI versions may not recognize the krea2 CLIP type.

• Model filenames may differ depending on your installation.

• Optional modules can increase generation time and VRAM usage.

• SeedVR2 should only be enabled after the final image path has been selected.

• This workflow is modular, so you do not need to use every module for every generation.

Workflow Philosophy

The goal of this workflow is not to stack as many processing stages as possible.

Krea 2 Turbo is already designed for fast, high-quality generation. Optional modules should only be enabled when they provide a visible improvement.

The workflow prioritizes:

• Fast generation

• Clear organization

• Modular control

• LoRA compatibility

• Stable conditioning

• Mild refinement

• High-quality final upscaling

• Easy preview and save outputs

Version

ILLUMINATE AI — KREA 2 TURBO MODULAR WORKFLOW v1.6

Version Highlights

• Native Krea 2 Turbo generation pipeline

• Power LoRA Loader with MODEL and CLIP routing

• Four-image Concept Batch

• Optional negative prompt switcher

• Mild Quality Refinement

• Face Detailer with refined-image fallback

• Optional image-to-image branch

• Final-image selector

• SeedVR2 positioned as the true final upscale stage

• Notes and usage instructions included directly inside the workflow