Civitai #144970991: Korean Steam Sauna Portrait - BloomGirls & Krea 2 Turbo
Full Resolution
Generation Specs RTX 4070
Sampler euler
Scheduler simple
Steps 10
CFG Scale 1
Seed 88392104
Target VRAM 12GB
VRAM 8GB - 12GB
GPU Compatibility
Optimal Performance

Runs entirely in VRAM with no memory swapping.

Krea 2 Turbo (Official Comfy-Org) Civitai #144970991 Min 8GB VRAM 832×1248 RTX 4070 Verified

Korean Steam Sauna Portrait - BloomGirls & Krea 2 Turbo

Candid Korean spa portrait featuring dewy glass skin, relaxed cedar sauna aesthetics, and soft golden rim lighting on Krea 2 Turbo.

Blueprint Summary RTX 4070 Verified

Reproducible ComfyUI workflow for Civitai #144970991: Korean Steam Sauna Portrait - BloomGirls & Krea 2 Turbo using Krea 2 Turbo (Official Comfy-Org) at 832×1248 resolution. Requires minimum 8GB VRAM with sampler euler and scheduler simple (10 steps). Includes 1-click terminal model sync and canvas JSON graph.

Pipeline Flow 11 Nodes
Open in Resolver
01 Load Models
UNETLoader
5 loaders (DiT, CLIP, VAE)
02 Adapters
LoraLoader
2 adapters
03 Conditioning
CLIPTextEncode
2 prompt encodings
04 KSampler
KSampler
DiT latent denoising
05 Decode & Save
VAEDecode
Latent to pixel space

Execution DAG Topology Interactive Visualizer

Drag to pan · Scroll to zoom · Hover wires
Node Graph 0 nodes
MODEL CLIP LATENT VAE IMAGE

Model & Asset Setup Setup Script

Run in your ComfyUI root:

curl -fsSL https://decomfy.com/api/scripts/civitai-steam-lounge-portrait-krea2.sh | bash

Positive Prompt

bl00m, candid portrait of a gorgeous 21-year-old Korean girl with wet dark silky hair in a loose bun with damp strands framing her delicate face. Dewy translucent glass skin, soft rosy cheeks, subtle lip gloss, sweet gentle smile looking directly at camera. She wears a soft minimalist white waffle-cotton spa lounge wrap robe, sitting comfortably on cedar wood bench inside a luxury steam sauna room. Swirling warm vapor mist, warm ambient golden lighting glowing behind frosted panels, glistening water droplets on damp cedar planks, soft cinematic rim light, 35mm documentary photography, intimate editorial spa lookbook.

Negative Prompt

unclothed, disrobed, revealing clothes, lowres, blurry, bad anatomy, deformed hands, extra fingers, missing fingers, low quality, artifacts, distorted proportions, text, watermark, cartoon, 3d render, plastic skin
Prompt Customization: This workflow is pre-wired and calibrated. Swap this prompt with your own character, subject, or style without breaking the pipeline.

LoRA Adapter Stack 2 Adapters

RTX 4070 Calibrated Weights
01 BloomGirls UltraRealism LoRA +0.8
02 Famegrid Natural Realism LoRA +0.7

Required Models 5 Models

Required Storage: 19.6 GB (5 models · DiT/Base: 12.6 GB · Text Encoder: 4.9 GB · LoRAs: 1.9 GB · VAE: 253 MB)
models/diffusion_models/ 12.56 GB HuggingFace
krea2_turbo_int8_convrot.safetensors Official Krea 2 Turbo INT8 ConvRot Diffusion Transformer core
models/text_encoders/ 4.88 GB HuggingFace
qwen3vl_4b_fp8_scaled.safetensors Official Qwen3-VL 4B FP8 Scaled Text Encoder (CLIPLoader type: krea2)
models/vae/ 253 MB HuggingFace
qwen_image_vae.safetensors Official 16-channel Video & Image VAE
models/loras/ 447 MB HuggingFace (Mirror)
krea2-bloomgirls-realism-step00004000.safetensors Weight: 0.8, Trigger: 'bl00m' (BloomGirls UltraRealism LoRA for refined facial geometry and delicate makeup rendering)
models/loras/ 1.49 GB HuggingFace (Mirror)
Famegrid-Natural-V1-Krea-2.safetensors Weight: 0.7 (Famegrid Natural Realism LoRA for natural skin pores, high dynamic range, and candid lighting bounce)

Field Notes RTX 4070 Benchmark

Benchmark: Rendered on NVIDIA GeForce RTX 4070 (12GB VRAM, CUDA 12.4)
1. ARCHITECTURAL REASONING & COMPOSITION GEOMETRY: Deconstructs high-humidity ambient lighting and delicate East Asian facial geometry within diffusion latents. Steamy sauna environments with airborne vapor mist and warm backlight typically lead to lost micro-contrast or distorted skin gloss when rendered naively. Chaining BloomGirls UltraRealism (v1.0, trigger bl00m, weight 0.8) with Famegrid Natural Realism (weight 0.7) stabilizes delicate facial proportions and translucent glass skin while preventing over-smoothed plastic artifacts. Grounding the subject in a clean white waffle-cotton spa robe provides high-frequency textile micro-geometry, separating foreground latents cleanly from ambient cedar wood background tones. 2. CABLE ROUTING & SIGNAL TOPOLOGY: Sequential two-stage LoRA injection architecture feeding from Krea 2 Turbo INT8 UNET (krea2_turbo_int8_convrot.safetensors) and dedicated Qwen3-VL 4B text encoder (qwen3vl_4b_fp8_scaled.safetensors). The first stage applies BloomGirls to guide facial geometry and East Asian aesthetic conditioning; the second stage injects Famegrid to anchor natural pore detail and specular rim light gradients. Conditioning routes directly into a single deterministic positive conditioning node wired straight to KSampler, enforcing the WYSIWYG single-node invariant with zero standby nodes. Latents are decoded through the native 16-channel qwen_image_vae.safetensors. 3. HARDWARE BENCHMARKS & CONSUMER GPU FOOTPRINT: Synthesized and benchmarked on host NVIDIA GeForce RTX 4070 (12GB VRAM, CUDA 12.4). Peak VRAM consumption maintained at 4.2 GB during the 10-step Euler sampling at 832x1248 resolution. Total generation latency clocked at 111.7 seconds with zero host RAM offloading, confirming fluid viability across 8GB and 12GB consumer GPUs.

Frequently Asked Questions FAQ

What GPU and VRAM are required to run Civitai #144970991: Korean Steam Sauna Portrait - BloomGirls & Krea 2 Turbo?

This workflow requires a minimum of 8GB VRAM (recommended 12GB VRAM). Tested and verified on NVIDIA GeForce RTX 4070 (12GB VRAM) at 832x1248 resolution.

How do I resolve missing custom nodes for this workflow?

You can drop the workflow JSON into our client-side Missing Node Auto-Resolver at https://decomfy.com/resolve/ to detect missing nodes and generate install commands, or run the 1-click terminal setup script provided below.

Which VAE and conditioning text encoders are required for Krea 2 DiT?

Krea 2 DiT single-stream models strictly require the 16-channel Wan2.1 VAE and Qwen3-VL text encoder. Using standard Flux or SDXL VAE causes inverted checkerboard artifacts. Also pair with RescaleCFG for high-contrast portrait clarity.

Related ComfyUI Blueprints

View all 44 workflows
Krea 2 DiT
Fast Turbo / DMD2 #144433221

Artisan Espresso Barista - Natural Realism & Krea 2 Turbo

10st · euler 832×1248 8GB
Krea 2 DiT
Fast Turbo / DMD2 #145008974

E-Girl Flash Hardwood Snapshot - BloomGirls & Krea 2 Turbo

10st · euler 832×1248 8GB
Krea 2 DiT
Fast Turbo / DMD2 #144982792

Candid Flash Loft Portrait - SNOFS & Krea 2 Turbo

10st · euler 832×1248 8GB
Action completed