Krea2 Size Preset

What it is

The plain one in this pack. Everything else here is doing something clever with attention and reference latents; Krea2 Size Preset is just a nicer Empty Latent Image - pick a resolution tier and an aspect ratio from dropdowns instead of typing pixel dimensions by hand, and get a batch of empty Krea2-shaped latents out the other end. It exists because Krea 2 has its own set of sane working resolutions, and typing 1216x832 from memory every time you want to try a different aspect ratio gets old fast.

Why it matters for this pack specifically

On its own this node is a convenience. Inside the Style Transfer pack it's load-bearing. The main style-transfer flow depends on your reference latent and your target latent being built to the exact same size - the pack's own README is explicit that the reference gets adapted to match target_latent, and a mismatch there is one of the documented failure modes. The cleanest way to guarantee that match is to generate one latent here and wire its output into both Krea2 Style Reference's target_latent input and your KSampler's latent input, rather than building two separate empty latents that happen to look like the same resolution. One source of truth for the size, used twice.

Inputs and outputs that matter

Three inputs, all simple choices. resolution picks a rough scale tier - 0.5K, 1K (the default), 1.5K, or 2K - which is Krea's own shorthand for total pixel budget rather than a literal side length. aspect_ratio covers the usual spread: square (1:1), portrait and landscape phone-ish ratios (9:16, 16:9), photo ratios (3:2, 2:3), classic frame ratios (4:3, 3:4), a couple of social-crop ratios (4:5, 5:4), and ultrawide (21:9) - default is 16:9. batch_size is how many identical empty latents you get back, from 1 up to 64, useful if you're generating a batch of variations against the same reference in one go.

The single output is latent - a LATENT you can plug straight into a KSampler, or, as above, split off to also feed Krea2 Style Reference's target_latent input so the two stay in lockstep.

Installing it

Comes bundled with the rest of the pack - there's nothing to install separately. ComfyUI Manager, search "ComfyUI Krea2 Style Transfer," or cd ComfyUI/custom_nodes && git clone https://github.com/jieg9341-lab/ComfyUI-Krea2-StyleTransfer and restart ComfyUI. This particular node doesn't touch the model, VAE, or text encoder at all, so it'll show up and work even before your Krea2 checkpoints are sorted out - it just won't be much use until the rest of your graph can consume the latent it produces.

Where people get burned

Honestly, not much to trip on here - it's a dropdown-driven convenience node. The one real gotcha isn't in this node, it's in how people wire it: if you use Krea2 Size Preset for your sampler's latent but build a second, separately-configured empty latent for Krea2 Style Reference's target_latent, you can end up with two latents that look the same on paper but aren't actually identical objects - and if you ever go back and tweak one without the other, you've silently broken the size match the whole style-transfer pipeline depends on. Route one latent output to both places and you avoid the whole class of problem.