Nodes/Eric_Krea2/Eric Krea2 Latent -> ComfyUI LATENT
ComfyUI Node

Eric Krea2 Latent -> ComfyUI LATENT

The escape hatch that hands Krea 2 latents to the rest of ComfyUI

By EricRollei·Created 2 months ago·Updated about a month ago· 11
Eric Krea2 Latent -> ComfyUI LATENT
  • latent
  • latent

A KREA2_LATENT is a packed, flow-format thing - the pack's own representation, not what the rest of ComfyUI expects. Any tool that wants a stock LATENT - VAEDecode, latent upscalers, masking math, whatever you've built your habits around - will refuse to wire to it. This node is the adapter that fixes that.

Eric Krea2 Latent -> ComfyUI LATENT takes one KREA2_LATENT in and emits a standard LATENT out: the packed flow latent unpacked to a spatial [B, 16, 1, H_lat, W_lat] tensor, normalized the way ComfyUI keeps LATENT['samples']. That means everything downstream is stock. ComfyUI's own VAE applies its de-normalization at decode, which is the part you'd otherwise get subtly wrong by hand.

The intended pairing, per the source, is a ComfyUI-loaded Wan 2.1 (or Qwen-Image) VAE plus the stock VAEDecode node - because Krea 2, Qwen-Image, and Wan 2.1 all share identical latents_mean/std, the native decode just works. And that pairing is not a curiosity, it's the node's primary purpose: A/B the pack's diffusers decode path (its own tiling and cosine blending) against ComfyUI's native tiled decode of the same latent, to isolate whether a bottom band or grid artifact is coming from the decode side or is baked into the generation. If the band vanishes through native VAEDecode, it's the pack's decode path; if it survives, it's in the latent.

There are no knobs. One input (latent), one output (latent). It's a two-second wire.

When would a normal person reach for it? Three honest cases. One: you like the stock VAEDecode + a ComfyUI VAE of your choice for the final decode, and want to stop worrying about which pack node does what. Two: you want to pipe a Krea 2 latent into any ComfyUI-native latent pipeline (resize, blend, mask combine) that only speaks LATENT. Three: debugging, per above - and the pack even ships a Save Latent (debug) node to complement exactly this workflow. The reverse direction (ComfyUI LATENT → Krea 2) doesn't exist in this pack, so it's a one-way door: take a packed latent out of the ecosystem, and you're in stock-land from there.

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. The pack's shared requirements apply: git-installed diffusers with Krea2Pipeline, transformers>=4.57.0, and Krea 2 weights for the nodes upstream of this one.

Common issues

  • "can't connect LATENT to KREA2_LATENT" - that's the whole reason this node exists; run it between the two worlds.
  • Colors look off after native decode - you paired it with a non-Wan/Qwen VAE. The VAE must be Wan-2.1-family or Qwen-Image, since those share Krea 2's latent statistics.
  • Nothing shows up - it's an intermediate node; wire its latent output onward to VAEDecode or wherever you're headed.
CategoryEric/Krea2

Inputs (1)

NameTypeDefaultDescription
latentKREA2_LATENT

Outputs (1)

NameTypeDescription
latentLATENT