✨ Krea 2 The Ultimate v1

Modular Minimalist Pipeline for Krea 2

A clean, modular pipeline built for clarity, flexibility, and efficiency. Each node has a single responsibility, and all logic is organized into independent, interchangeable blocks — no clutter, no unnecessary wiring.
This workflow is fully functional out of the box. A general overview is provided below, followed by a dedicated "For Beginners" section at the end with clarifications on specific aspects that may not be immediately obvious to new users.

🧠 Prompt Engineering

Google Translate Text Node — write prompts in any language, automatically translated to English.
WD14 Tagger (optional) — extracts tags from a reference image.
Merge Strings [Eclipse] — combines translated prompt + WD14 tags.
Qwen3-VL-4B-Abliterated ("Prompt Enhancer" optional) — generates an enhanced, artistic prompt from the reference image.
When both WD14 and Prompt Enhancer are active, the enhancer uses WD14 tags as context — noticeably improving results.
Note: Qwen3-VL-4B-Abliterated can either pass the prompt through unchanged or actively enhance it when the Prompt Enhancer signal is enabled. Merge Strings [Eclipse] controls which signals are routed to Qwen, offering multiple flexible combinations:
- Your own prompt only
- Your prompt + WD14 tags
- WD14 tags + Prompt Enhancer
- Your prompt + WD14 tags + Prompt Enhancer
This adds significant flexibility, allowing you to choose the ideal prompt composition for each use case. If you need even more flexibility, you can add additional text sources directly into the Merge Strings [Eclipse] node.

🧩 LoRA & Trigger Text

The workflow includes built‑in support for LoRA models, with a dedicated Trigger Text system.

Trigger Text — a text field where you can enter any phrase (e.g., a LoRA trigger word, character name, or style cue).
This text is automatically inserted at the very beginning of the final prompt, followed by a comma, before all other prompt content.

How it works:

You write your trigger text (e.g., "krea2style").
The workflow places it at the start of the prompt:
"krea2style, [the rest of your prompt...]"
This ensures LoRA‑specific triggers are applied first, maximizing their effectiveness.

Use cases:

Activate a LoRA trigger word without manually editing the prompt each time.
Add a consistent style prefix (e.g., "photorealistic, masterpiece") to every generation.
Quickly switch between characters, styles, or themes by changing just the trigger text.

🖼️ References: Single Image or Batch Processing

The same bypasser manages reference inputs:

Reference — for processing a single image.
Batch Reference — for batch processing: simply specify a folder, and the workflow will automatically process all images in sequence.

⚙️ Generation

Fast Groups Bypasser (rgthree) — selects active sampler.
KSampler (Advanced) — 8-step base + 5-step refinement.
Detail Daemon Sampler — alternative sampler for extra variety.

🔀Two Seeds in KSampler Advanced

Base pass (8 steps) — controls composition and structure.
Refinement pass (5 steps) — introduces fresh noise for micro‑details and texture.
Using two different seeds prevents over‑fitting and allows for richer detail while preserving overall structure.

🎨 Post-Production (Subgraph)

ProPostApplyLUT — applies a cinematic color look using LUT files (Look-Up Tables). A LUT is a preset color transformation that shifts colors, contrast, and tone to achieve a specific filmic or stylized look in one click. The node is disabled by default — to use it, you must place a .cube LUT file into the following folder after installing the custom node:
Comfy-Desktop\ComfyUI-Installs\ComfyUI\ComfyUI\models\luts
Orion4D · Channel Mixer APP — a flexible color channel mixer that comes pre-configured for monochrome (black-and-white) output by default, but can be fully customized to adjust RGB channel mixing, allowing you to create custom color toning, stylistic looks, or precise color grading as needed.
Orion4D · Curves APP — global contrast control with a built‑in Soft Contrast preset, or you can manually adjust the RGB curve to fine‑tune contrast, brightness, and color balance exactly the way you want.
Ultimate SD Upscale — upscale + detail enhancement.
— Portraits (default): 2 steps / 0.10 denoise
— Landscapes / abstract (manual override): 3 steps / 0.20 denoise
Face Detailer (DetailerForEach) — automatic face enhancement. The node crops the face area from the image, then uses Qwen3-VL-4B-Abliterated and WD14 Tagger to generate a detailed positive prompt for regeneration, preserving original character and skin quality — no manual input required.
This module is disabled by default, as the result after Ultimate SD Upscale is already strong, but it can be enabled when needed.
It also works well with face-specific LoRAs, or you can disable auto‑prompting and manually adjust the prompt.
— For improving the existing face (texture, clarity, detail) — set denoise to 0.10-0.15
— For replacing the face with a different one (e.g., using a LoRA) — set denoise to 0.40–0.50.
Contrast Skin — subtle local contrast for skin depth.
Seed VR2 — final upscale to 4K with texture (recommended: 3B model).
Orion4D · SharpnessMax APP — final sharpening (Micro Texture preset).

🧩 SUBGRAPH NOTE

The entire Post-Production block is inside a subgraph for visual clarity.

⚠️ Due to a ComfyUI limitation, Orion4D nodes inside a subgraph may not show image previews when clicking "Open APP". Simply expand the subgraph to view previews — processing remains unaffected.

💾 Save & Metadata

Image Saver — saves images with full CivitAI‑compatible metadata (prompts, seeds, models, LoRAs, CFG, steps, denoise, etc.).
Metadata Controls (Group: 1️⃣Save image) — contains manual inputs for Denoise, Steps, Detailer Steps, CFG, Sampler Name, and Scheduler.
These values exist only for metadata logging and do not affect the actual generation or processing.
Prompt, LoRAs, model name, and seed are automatically pulled from the workflow. The seed is taken from the first generation pass.

🎯 ColorMatch (KJNodes)

ColorMatch (KJNodes) — used to prevent unwanted color shifts introduced by processing steps like Ultimate SD Upscale.
For example, if you apply a monochrome preset and then run the image through Ultimate SD Upscale, the upscale process may reintroduce color casts or alter the black‑and‑white tonality. ColorMatch locks in the correct color palette at key stages, ensuring that:
— Monochrome remains truly monochrome.
— The final output stays consistent with your intended color grading.
Strength:
0 — disabled
1 — full color transfer from the reference
Used:
— Before Ultimate SD Upscale — to freeze the current color state.
— After monochrome conversion — to prevent the upscaler from adding unwanted color back into the image.

🎈 VRAM Cleanup

After resource‑intensive nodes (such as generation, upscaling, and Face Detailer), the workflow includes VRAM‑Cleanup nodes to free up memory and prevent slowdowns.

This ensures:

Stable long‑running sessions (batch processing, multiple generations).
Lower risk of out‑of‑memory errors, especially on GPUs with limited VRAM.
Smoother performance when switching between heavy modules.

These cleanup nodes are placed strategically after key processing steps, so memory is released automatically — no manual intervention required.

RAM‑Cleanup — optional, disabled by default. If you experience high RAM usage, simply connect the output of VRAM‑Cleanup to the input of RAM‑Cleanup to clean system memory as well.

This keeps long sessions (batch processing, multiple generations) stable and smooth.

⚠️ Node ID Note

Node 5551 (WidgetToString) — extracts LoRA strength from the Power Lora Loader for metadata logging.

Node 5518 (Workflow Input Value (Image Saver)) — retrieves the model name for metadata.

When copying this workflow, node IDs change. You must manually update the referenced node IDs in both nodes for metadata logging to work correctly.

📦 Models

UNET: krea2_turbo_fp8_scaled.safetensors
VAE: qwen_image_vae.safetensors
CLIP: qwen3VL4BAbliteratedComfyui_v10.safetensors
Upscale: 4x-ClearRealityV1_Soft.safetensors
Skin Contrast: 1xSkinContrast-High-SuperUltraCompact.pth
SeedVR2 (3B): seedvr2_ema_3b_fp8_e4m3fn.safetensors

📦 Required Custom Nodes

ComfyUI-KJNodes
ComfyUI-Detail-Daemon
Orion4D_FXMax
comfyui-ultimatesdupscale
rgthree-comfy
ComfyUI-WD14-Tagger
comfyui-easy-use
ComfyUI-Eclipse
Use Everywhere (UE Nodes)
ComfyUI-SeedVR2_VideoUpscaler
ComfyUI-Manager
ComfyUI-Image-Saver
ComfyUI-Impact-Pack

⚙️ Default Configuration

Out of the box, the workflow is pre‑configured for a clean, fast start:

Prompt: Your own manual prompt is used.
Reference: Disabled.
Text Enhancers (WD14, Qwen3‑VL‑4B‑Abliterated): Disabled.

Post‑Production enabled by default:

Ultimate SD Upscale
Contrast Skin
Seed VR2
SharpnessMax

All other post‑production nodes (LUT, Channel Mixer, Curves, Face Detailer, ColorMatch) are disabled by default — you can enable them as needed.

This gives you a lightweight, fast baseline that you can expand with additional modules depending on your requirements.

⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️⚠️

🐣 For Beginners: How to Work with Subgraphs

Some parts of this workflow are grouped into Subgraph nodes — visual containers that help keep the main canvas clean and organized.

To open a Subgraph and see the nodes inside:

Locate the Subgraph block (e.g., labeled post production).
Double‑click on it with your left mouse button.
The internal nodes will open in a new view — you can now edit settings, add, or disable individual nodes as needed.

Why use Subgraphs?

Keeps the workflow clean and easy to navigate.
Hides complex logic behind a single block.
Makes it easier to reuse entire processing chains in other workflows.

⚠️ Important: If a Subgraph contains Orion4D nodes with an Open APP button, the preview may not appear when you click it — this is a known ComfyUI limitation. To see the preview, simply expand the Subgraph back into the main workflow.

⚠️ Important Note: Ultimate SD Upscale

Ultimate SD Upscale is useful for increasing resolution and adding detail, but it is not always beneficial.

If your source image is already highly detailed — especially male portraits with visible wrinkles, skin folds, and age‑related textures — using this node can produce the opposite effect:

Instead of sharpening details, it may create a "soap opera" effect — skin becomes smooth, wrinkles blur, and fine textures get washed out.
This is especially noticeable when using texture‑heavy LoRAs, which already add plenty of detail on their own.

Solution:
If you're using LoRAs that already deliver strong texture and detail, simply disable Ultimate SD Upscale using the Fast Groups Bypasser (rgthree). The workflow will continue without this node, preserving all the fine detail already generated by the model and LoRAs.

All Subgraph blocks in this workflow are fully editable. If you want to tweak or fine‑tune the internal settings:

Double‑click on any Subgraph to open it and reveal its internal nodes.
Adjust parameters inside as you wish — the workflow is designed to be modular and flexible.

⚠️ Before changing any values, it's strongly recommended to:

Copy the node name (or the specific parameter name) and search for it in an AI assistant or documentation to understand what it does.
Many parameters (e.g., guide_size, denoise, tile_width) affect processing speed, quality, or memory usage — knowing what they control will help you avoid unexpected results.

This way, you can safely experiment and adapt the workflow to your own needs without breaking it.