Krea 2 One Node Is Here (Patreon)

WATCH THE YOUTUBE VIDEO

I built a single ComfyUI node that replaces the entire Krea 2 workflow. No KSampler, no CLIPTextEncode, no LoadImage, no separate VAEDecode chain sitting around your canvas. Prompt in, image out.

What it does

Both Krea 2 Turbo and Krea 2 Raw are built into one node, switchable with a tab strip at the top. Switching tabs doesn't just change a label, it swaps the sampling settings underneath and recolors the whole node so you always know which model you're on.

Turbo runs 8 steps, guidance floored at 1.0, euler and simple. There's no negative prompt box on Turbo, and that's intentional. The official template zeroes out the negative branch instead of encoding into it, so a negative prompt there does nothing. Rather than pretend it works, the node just hides the box.

Raw runs 52 steps, guidance 4.5, also euler and simple. Worth knowing, there's no official Comfy-Org template for Raw. These numbers are inferred from a third party conversion thread, not confirmed by the developers, and I haven't personally tested Raw yet. Treat it as a starting point, not settled fact.

Both t2i and i2i modes are built in, and each keeps its own prompt and negative prompt independently. Your t2i prompt doesn't get touched when you flip to i2i and write something completely different, they're saved separately.

i2i adds a source image panel with drag and drop, a file picker, and a browse gallery that reads your whole input folder with search and paging built right into the node. Denoise defaults to 0.4 and for most i2i work you shouldn't be pushing that around much, higher values just drift further from your source.

For sizing, t2i gets the full aspect chip grid.

For i2i, use the longest edge toggle instead of the presets, it scales to your source image's own shape so the longer side hits whatever number you type. The presets will crop or letterbox against your input's shape.

LoRA stack goes up to 20 slots, each with its own strength slider and a trigger word field that autofills from the LoRA's metadata the moment you pick a file.

Why one node

Most custom node work in this space ships as bundles, a handful of specialized nodes you still have to wire together yourself. This is different, it's one node covering the full pipeline for a specific model family, built with the right defaults baked in instead of a wall of generic widgets you have to configure from scratch every time.

Install instructions

Download the zipped node folder attached to this post.

Extract it. You should end up with a single folder, don't extract into a folder that already has files in it, just unzip it fresh.

Move that extracted folder into your ComfyUI custom_nodes directory. This is inside wherever you have ComfyUI installed, the folder is always named custom_nodes regardless of your setup:

ComfyUI/custom_nodes/

Restart ComfyUI completely, don't just refresh the browser tab.

In the node search, look for Krea 2 One Node.

Before your first run, set the three model dropdowns manually, Turbo diffusion model, Raw diffusion model, and text encoder. They default to whatever's first alphabetically in your folder, which is almost never what you want.

Make sure you have the four required files sitting in the right folders first:

Krea 2 Turbo diffusion model → models/diffusion_models

OR

Krea 2 Raw diffusion model → models/diffusion_models

qwen3vl_4b_fp8_scaled.safetensors → models/text_encoders

qwen_image_vae.safetensors → models/vae

If a dropdown comes up empty, it means that model file isn't in the right folder yet, drop it in and restart.