Krea2 Turbo Reference Sigmas (From Latent)

Krea 2 Turbo is the distilled 8-step checkpoint, and the single biggest way to ruin it is feeding it a sigma schedule it wasn't trained for. That's what this node is for: it hands you the exact schedule Krea shipped the model with - fixed timestep shift mu=1.15, Euler - plus a hard check that your latent is on the model's 16-pixel grid. No math, no guessing, no accidentally leaving a stray shift node in the graph.

It lives in ComfyUI-Krea2T-Enhancer, a pack that mostly does prompt-adherence patching on Krea 2. The scheduler node is the odd one out: it doesn't touch conditioning at all. It's a guard rail for the sampling side, and honestly the most obviously useful thing in the pack if you've already got a working Krea 2 setup.

What it actually does

Flow-matching models like Krea 2 sample along a near-straight noise-to-image trajectory, and a knob called shift decides how the denoising effort is spread between composition and detail. Turbo has that value baked into its training at 1.15 - unlike the Raw checkpoint, which uses a resolution-dependent rule. Get the shift wrong and you're denoising at noise levels the model never saw, which reads as "softer/mushier than it should be" rather than an obvious error.

Under the hood the node builds the timesteps linearly from 1 down to 0, applies the exponential shift (σ = e^μ·t / (1 + (e^μ − 1)·t)), and returns the tail of the schedule as a SIGMAS tensor. Before it does, it runs three validations:

• the connected MODEL is actually a Krea2 instance,
• the model's current model_sampling.shift is exactly 1.15 (it raises if a conflicting shift node is upstream),
• the latent's pixel dimensions are divisible by 16 (VAE scale 8 × DiT patch size 2).

The last check is the point of the "(From Latent)" name. It can't build a schedule for a canvas it can't see, so you wire in the same Empty Latent Image the sampler uses.

The inputs and outputs that matter

Only four inputs, and two of them are connect-a-cable-and-forget:

• model - your Krea 2 Turbo diffusion model, straight from the loader.
• latent - the same Empty Latent Image the sampler gets. Validated, then passed through unchanged.
• steps - default 8, and 8 is what the reference Turbo setup uses. Going to 12–16 buys you a little stability on img2img; beyond that you're defeating the point of a distilled model.
• denoise - default 1.0 for the full schedule. Lower values keep the final steps of a longer schedule, which is the right move when you want a partial denoise on a high-res base but want the step count to stay meaningful.

Outputs are sigmas and latent. The sigmas socket goes into a sampler node that takes one - SamplerCustom or SamplerCustomAdvanced in core ComfyUI - and the pass-through latent goes into that sampler's latent input. It does not patch prompt adherence; the pack's other two nodes do that.

Installing it

Via ComfyUI Manager, search ComfyUI-Krea2T-Enhancer. Or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/capitan01R/ComfyUI-Krea2T-Enhancer.git

Restart ComfyUI. No extra Python packages - the README is honest there. What you do need is a working Krea 2 stack underneath: ComfyUI 0.25.0+ (that's where native Krea 2 support landed), the Turbo checkpoint, the Qwen3-VL text encoder (~8GB), and the Qwen-Image VAE (~250MB). People routinely miss the last two and then blame the node.

Where people get burned

• The "requires a Krea 2 Turbo MODEL" error. You wired the Raw checkpoint in, or a patched/LoRA'd model that no longer types as Krea2. Turbo is the one.
• The shift-mismatch error. You've got a ModelSamplingAuraFlow-style shift node somewhere upstream. Delete it - Turbo's shift is fixed, and the node will tell you exactly.
• The divisibility error. Width or height not divisible by 16. Fix the Empty Latent Image, not the node. A 1024×1024 base and a 1024×1536 img2img pass both sail through; 1152×1152 does not.

It's a small node with a narrow job, but it's the difference between Krea 2 Turbo looking like the model everyone hyped and looking like a blurry approximation of it.