Small, fast, no magic nodes. Just what you need to run Krea2 clean.
What it does
Text-to-image with Krea2
Optional upscale pass — toggle on/off
Built-in image comparer (base vs upscaled)
Power LoRA slot ready to plug your stacks
Everything runs off the same Krea2 model — no extra checkpoints, no heavy upscalers loading in and out. On my machine (RTX 5070 Ti, 16GB VRAM, 48GB RAM): ~29s base, ~60s with upscale.
Design choice — the upscale
I could have plugged SeedVR or a heavy upscaler here. Better quality, sure — but you'd be loading a second model into VRAM and paying the price on every generation.
Instead, the upscale runs Krea2 itself over an ImageScaleBy 1.5x → low-denoise refine. Same model already in memory, no context switch. It's a deliberate compromise: not the sharpest possible option out there, but the quality-to-speed ratio sits exactly where I wanted it — usable on a home PC, no minutes-per-image waits.
Heads up — I'm likely dropping a follow-up version in the next few days with SeedVR integrated as a second upscale option, not a replacement. You'll be able to pick which method fits the shot.
Requirements
Krea2 — I run
krea2TurboFP8(~12GB, fp8 safetensors)Text encoder — Qwen3-VL 4B (~5GB in the variant I use)
VAE —
qwen_image_vae(~250MB)
Tighter VRAM: the ComfyOrg repo also ships an nvfp4 build (~7GB). The workflow is built for .safetensors, not GGUF — keep that in mind if you're shopping for lighter variants.
Custom nodes:
rgthree-comfy— Power LoRA Loader, Fast Groups Muter, Image Comparereasy-use— set/get nodesComfyUI-ConditioningKrea2Rebalance(nova452) — per-layer conditioning tuning, dialed in for Krea2comfyui_image_metadata_extension(Santodan)
Krea2 clicked for me the second I loaded it. First base model in a while where realism doesn't need me to fight the sampler.
— ShinobiSat