CivArchive
    Penko´s Krea 2 V3.0 - 12GB VRam - 2 Pass - Upscale x2 - v2.0
    NSFW
    Preview 140547663
    Preview 140540801
    Preview 140540599
    Preview 140540685
    Preview 140541017
    Preview 140671548
    Preview 140671677
    Preview 140671811
    Preview 140672059
    Preview 140672640
    Preview 140680517

    About this version

    ### 🚀 [UPDATE v3.0] Ultimate Krea 2 Turbo - Optimized & Enhanced

    #### 🇺🇸 ENGLISH DESCRIPTION & CHANGELOG (v3.0):

    Following my previous setup, here is the optimized Version 3.0! This update completely fixes the common "AI plastic look" and delivers incredibly sharp details (eyes, skin, hair) optimized for 12GB VRAM without slowing down.

    What's New in v3.0:

    * Node Optimization: Many old nodes have been replaced with optimized variants to make the layout cleaner, more organized, and easier to use.

    * Three-Way Prompt System: Implemented a new three-way prompt logic. This helps the workflow better communicate prompt details and aesthetics.

    * Pixaroma Nodes Base: Built directly on top of the Pixaroma custom nodes as the core foundation for this setup.

    Core Features & Setup Logic:

    * RealESRGAN + Bicubic Secret: Replaced standard upscalers with the superior RealESRGAN_x4plus model paired with a custom Bicubic Resize node. This fixes blurred faces and muddy eyes on distant/wide shots perfectly and gives hair a beautifully textured, realistic look. The image is first upscaled to x4 to generate maximum detail and then downscaled back to x2. This supersampling technique keeps the textures incredibly sharp.

    Three-Way Prompt System: Implemented a new three-way prompt logic. This helps the workflow better communicate prompt details and aesthetics.

    * 100% VRAM Safe (Tiled Engine): Utilizes VAE Encode (Tiled) right after the upscale. No more Out-Of-Memory (OOM) crashes on 12GB cards.

    My Personal Setup:

    * 1st Pass: 5 (minimum) Steps (Euler / sgm_uniform or beta57)

    * 2nd Pass: 2 Steps (Euler Ancestral / sgm_uniform or beta57 with Denoise ~0.18 - 0.22)

    * You can try higher steps and other samplers/schedulers for sure!

    #### 🇩🇪 DEUTSCHE BESCHREIBUNG & CHANGELOG (v3.0):

    Basierend auf meinem vorherigen Setup ist hier die optimierte Version 3.0! Dieses Update eliminiert den typischen "KI-Plastik-Look" komplett und liefert unglaublich scharfe Details (Augen, Haut, Haare) – optimiert für 12GB VRAM, ohne das System auszubremsen.

    Was ist neu in v3.0:

    * Nodes komplett optimiert: Viele alte Nodes wurden durch modernere, optimierte Varianten ersetzt. Der Workflow ist jetzt viel übersichtlicher, sauberer und verbraucht weniger Ressourcen.

    * Three-Way Prompt System: Ein neues dreifaches Prompt-System wurde integriert. Dadurch werden Prompt-Details und die gewünschte Ästhetik noch besser übermittelt.

    * Pixaroma-Nodes als Basis: Die Pixaroma-Nodes bilden das feste Fundament, auf dem dieser gesamte optimierte Workflow aufbaut.

    Kern-Features & Logik des Workflows:

    * RealESRGAN + Bicubic Geheimnis: Standard-Upscaler wurden durch das überlegene RealESRGAN_x4plus-Modell in Kombination mit einer benutzerdefinierten Bicubic-Resize-Node ersetzt. Dies korrigiert unscharfe Gesichter und matschige Augen bei Weitwinkelaufnahmen perfekt und verleiht Haaren eine wunderschön strukturierte, realistische Textur, da das Bild auf x4 hochskaliert wird und dann in meinem Fall wieder auf x2 runterskaliert wird. Dadurch bleiben die Texturen perfekt scharf.

    * 100% VRAM-sicher (Tiled Engine): Nutzt VAE Encode (Tiled) direkt nach dem Upscale. Keine Out-Of-Memory (OOM) Abstürze mehr auf 12GB-Karten.

    Meine persönliche Einstellungen:

    * 1. Durchgang (Pass): Mindestens 5 Schritte (Euler / sgm_uniform oder beta57)

    * 2. Durchgang (Pass): 2 Schritte (Euler Ancestral / sgm_uniform oder beta57 mit Denoise ~0.18 - 0.22)

    * Höhere Stepzahlen sowie andere Sampler/Scheduler können natürlich gerne ausprobiert werden!

    ---

    ### 📦 Primary Models I'm Using:

    👉 [Realism by Stable Yogi Krea2]

    (https://civarchive.com/models/2786499/realism-by-stable-yogi-krea2?modelVersionId=3254606)

    👉 [Krea 2 Neutrino]

    (https://civarchive.com/models/2801121/krea-2-neutrino?modelVersionId=3207678)

    👉 [Moody Krea 2 Mix (uncensored)]

    (https://civarchive.com/models/2731187/moody-krea-2-mix-uncensored?modelVersionId=3223899)

    A huge thank you to the creators for providing these amazing models to the community!

    ---

    ### 💻 Info & Hardware

    > "I'm not a professional developer, just someone who really dove deep into the settings to optimize this workflow."

    My System: NVIDIA RTX 4070 Super | 32GB RAM

    Platform: I am using [ComfyUI-Easy-Install] (https://github.com/Tavris1/ComfyUI-Easy-Install) . You can still use the regular standard version (Portable or Installer) with this workflow.

    ---

    ### 🧩 Custom Nodes Used in This Setup

    * ComfyUI-Pixaroma (Core base foundation of the workflow)

    * ComfyUI-Krea2T-Enhancer (Optional – only needed if you are using a censored base model or anything else)

    * ComfyUI-RBG-SmartSeedVariance (Used to achieve more variation in the seeds)

    * ComfyUI-KJNodes

    * rgthree-comfy

    * Comfyui_Essentials (For Resize Images by Longer Edge)

    * ComfyUI-Crystools (For Primitive Float Input)

    * ComfyUI-Prompt-Stash (Optional – I only use it as a prompt storage folder; it is not connected to any active nodes and can be completely removed)

    ***

    A big shout-out and thank you to everyone who shares their Custom Nodes and models with us for free!


    Description

    # 🚀 [UPDATE v2.0] Ultimate Krea 2 Turbo - 12Gb VRam - Ultra Sharp 2-Pass

    ### ✨ What's New in v2.0

    * RealESRGAN + Bicubic Secret: Replaced standard upscalers with a seamless RealESRGAN model paired with a custom Bicubic Resize node. This fixes blurred faces and muddy eyes on distant/wide shots perfectly.

    * Smart Prompt Automation: Built a text concatenator. Your target prompt style (Leica aesthetic, organic skin, zero digital smoothing) is locked in the background. You only need to change your action description in a single text box!

    * 100% VRAM Safe (Tiled Engine): Utilizes VAE Encode (Tiled) right after the upscale. No more Out-Of-Memory (OOM) crashes on 12GB cards.

    ---

    ### 📦 Primary Models I'm Using:

    🌟 [Realism by Stable Yogi Krea2]

    🎬 [Moody Krea 2 Mix (uncensored)

    🚀 Krea 2 Neutrino

    A huge thank you to the creators for providing these amazing models to the community!

    ---

    ### 💻 Info & Hardware

    > "I'm not a professional developer, just someone who really dove deep into the settings to optimize this workflow."

    * My System: NVIDIA RTX 4070 Super | 32GB RAM.

    * Platform: I am using [ComfyUI-Easy-Install] (https://github.com/Tavris1/ComfyUI-Easy-Install) because, in my experience, it runs more stable and effectively than the standard version. Of course, you can still use the regular standard version (Portable or Installer) with this workflow.

    ---

    ### 🧩 Custom Nodes Used in This Setup

    * ComfyUI-Krea2T-Enhancer (Optional – only needed if you are using a censored base model)

    * ComfyUI-Pixaroma (A permanent part of my core setup)

    * ComfyUI-RBG-SmartSeedVariance (Used to achieve more variation in the seeds)

    * ComfyUI-KJNodes

    * rgthree-comfy

    * Comfyui_Essentials (For Resize Images by Longer Edge)

    * ComfyUI-Crystools (For Primitive Float Input)

    * ComfyUI-Prompt-Stash (Optional – I only use it as a prompt storage folder; it is not connected to any active nodes and can be completely removed)

    ---

    ### 🧬 LoRAs

    LoRAs are rarely used in this workflow, but I have included the Power Lora Loader just in case.

    ***

    A big shout-out and thank you to everyone who shares their Custom Nodes and models with us for free!

    FAQ

    Workflows
    Krea 2

    Details

    Downloads
    157
    Platform
    CivitAI
    Platform Status
    Available
    Created
    8/23/2026
    Updated
    9/12/2026
    Deleted
    -

    Files

    penkoSKrea2V3012GBVram2_v20.json

    Mirrors