Krea 2 Empty Latent Image

Here's a trap the Krea 2 release threads didn't make a big deal of: Krea 2 doesn't use a 4-channel latent like SDXL or Flux's AE. It uses the Qwen-Image VAE, which works in 16 channels and downsamples by 16 instead of 8. Feed it the stock EmptyLatentImage - four channels, /8 dims - and you either get an immediate shape error or a weird, mushy image with zero warning about why. Krea2EmptyLatentImage is the "just give me a correct starting latent" node so you never think about that again.

It's the rgthree-school of thought applied to Krea 2: pick what you want in friendly units, get a latent that's actually correct for the model on the other end.

How it works

The node holds a table of Krea 2's native aspect-ratio buckets - 1:1, 16:9, 4:5, and so on - each with a base 1K size, and builds a correctly-shaped zero latent at batch_size deep. Because the Qwen-Image VAE downsamples by 16, every dimension snaps to a multiple of 16, so the latent you get out is always VAE-decodable at the exact pixel size you asked for.

The inputs that matter

There are really only two knobs you'll touch:

• dimensions - an aspect-ratio dropdown (15 presets, from 21:9 cinematic down to 4:5 portrait). Pick the shape, don't type pixel math.
• megapixels - a scale dial where 1.0 is the listed ~1K size and ~4.0 is 2K-class. Krea 2 was trained for roughly 1K to 2K output, so this keeps you inside the model's comfort zone instead of letting you wander into 3K territory where it gets soft.

clamp_to_2k (on by default) pins the longest side at or below 2048px - Krea 2's native ceiling - scaling wide ratios down proportionally rather than silently rendering off-spec. batch_size does what it says if you're generating a batch.

What comes out

Three outputs, and the first two are the reason to use this over a plain latent node:

• LATENT - the 16-channel latent, sized correctly and ready for the KSampler.
• WIDTH / HEIGHT - ints matching the computed size. These exist to wire straight into the Krea2 Regional Builder's width/height inputs, so your canvas and your latent can never drift out of sync. That's a genuinely useful touch for a workflow where you'll be drawing region boxes at canvas resolution.

Install

Same as the rest of the pack - ComfyUI Manager, search Krea2 Regional, or:

cd ComfyUI/custom_nodes
git clone https://github.com/januspluto/ComfyUI-Krea2-Regional.git

Restart, and you're done. No extra Python packages, no model downloads. Requires a ComfyUI build with native Krea 2 support (v0.26+).

One honest caveat

The preset labels are approximate 1K sizes (that 21:9 bucket is 1568×672, not 1920×810) - Krea 2 renders best when you stay near the pixel counts it was trained on, and these buckets are chosen to live there. If you crank megapixels toward the 4.2 max and leave clamp_to_2k off, you're on your own; the model was not trained for it. Inside the defaults, this is the most boring, reliable node in the pack - and that's exactly what you want from an empty latent.