12 steps, CFG 1.5, about 16 seconds on a 5090

(Text to Image + Image to Image)

Two ways to run it, chosen with one number. There is a single blue INPUT box near the left:

1 = text to image. Type a prompt, press Run.

2 = image to image. Load your own picture, press Run.

Two samplers sit behind that box with the denoise already correct for whichever you pick, so there

is one knob and not two. The half you did not choose never runs — nothing to mute, nothing wasted.

This is the one to use when you want a negative prompt. At CFG 2 the model listens to it, which the turbo settings at CFG 1 cannot. CFG 1.5+ with 12+ steps is deliberate and it is a pair: push the CFG higher without adding steps and the bright colours pin to maximum.

WHAT IT RUNS ON

Our own Krea 2 build, and it is free — every build on that page opens for any signed-in member, no code and no card. You also need the Krea 2 text encoder and VAE from the official Comfy-Org

repository, and the ComfyUI Impact Pack from ComfyUI Manager (that is the switch, and it is the only extra thing this file needs).

Watch the VAE: it is qwen_image_vae, NOT wan_2.1_vae. Use the wrong one and nothing warns you — no

error, no red border. It runs every step, decodes, saves a picture, and the picture is coloured

static.

THERE ARE THREE OF THESE

Turbo (8 steps, CFG 1, ~6s, no negative prompt),

Extended (16 steps, CFG 2, ~16s, negative works),

and Base (20 steps, CFG 3, ~46s, negative works).

They are genuinely different files, not three presets of one. Start with Turbo if you just want to see something.

Everything — all three files, the model, the folder each file goes in, and the settings: grab here.

The model on its own: grab here.

Workflow, settings and write-up are mine. Krea 2, its text encoder and VAE are Comfy-Org's. The

switch node is from the ComfyUI Impact Pack by ltdrdata.