Eric Krea2 Latent Resize

Somewhere between "generate a latent" and "decode it," you'll want it a different size. The naive move is decode → resize → re-encode, which is slow and, worse, softens the image - every VAE round-trip costs you detail and introduces its own grain. Eric Krea2 Latent Resize does the resize in latent space with bislerp, no VAE trip, no decode. It's the same idea as ComfyUI's own latent upscale, but native to the pack's packed KREA2_LATENT format.

The node takes a KREA2_LATENT and a mode - that's the only required choice - and offers three ways to think about the target:

• scale - multiply both dimensions by scale. Default 2.0, which is 2x per side = 4x the area. Values below 1 downscale. This is the one you'll use 90% of the time.
• dimensions - target width/height in pixels (rounded to /16). Put 0 on one axis and it derives that side from the source aspect. The tooltip suggests wiring the Eric Krea2 Resolution node here, which is exactly right for img2img setups where you want a stage to match a source image.
• megapixels - target area in MP (0.05–16), keeping the source aspect. Handy when you think "this should be a 4-megapixel image" rather than "this should be 2048 wide."

Outputs are latent (the resized packed latent), plus width, height, and a dims string - the last three exist so you can read the resulting size onward without guessing. Wire them into anything that wants to know what it's about to decode.

Where this actually sits in a workflow: Multi-Stage Ultra already handles inter-stage upscaling internally with its own upscale_to_stage2/upscale_to_stage3 factors, so you usually don't need this node in the middle of a multi-stage run. Where it earns its keep is the edges: prepping a Generate latent for a one-off img2img pass at a different resolution, resizing a reference latent before a decode, or any custom graph where the Ultra node's built-in chain doesn't fit. The tooltip's honesty is worth repeating - the latent-space resize is a lowpass; every hop shrinks latent variance slightly, which is contrast/saturation once decoded. One resize is invisible; the Ultra node has an upscale_renorm option to repair that drain across multi-hop runs, and if you're doing several manual hops through this node, keep that mental model.

How to install

Part of EricRollei/Krea2_ComfyUI_Advanced - ComfyUI Manager → search "Krea2_ComfyUI_Advanced", or:

cd ComfyUI/custom_nodes
git clone https://github.com/EricRollei/Krea2_ComfyUI_Advanced

Restart, and satisfy the pack's shared requirements (git-installed diffusers with Krea2Pipeline, transformers>=4.57.0, Krea 2 weights).

Common issues

• Result looks softer than expected - latent interpolation is a lowpass by nature. One hop is fine; stacked hops compound. If you're chaining resizes, that's what the Ultra node's upscale_renorm is for.
• Size isn't what you asked for - dimensions round to /16, so an exact pixel target may land a step off. Read width/height outputs to confirm.
• Nothing to resize - it needs a KREA2_LATENT in, so it sits after Generate or Ultra, not before.