Eric Krea2 Upscale Decode (2x)

Here's the thing most people miss about Krea 2's fast path: the model's VAE is AutoencoderKLQwenImage, a 16-channel f8 encoder with a latent space identical to Wan 2.1. That shared latent space is what makes this node possible, and it's the reason "Turbo → fast 4k" is a real workflow instead of a marketing slide.

Eric Krea2 Upscale Decode (2x) does exactly one thing: it decodes a KREA2_LATENT through the spacepxl Wan 2× upscale VAE - a decoder-only finetune trained to emit a 2x-larger, sharper image - instead of the plain 1× VAE. No re-denoise, no second sampling pass. The upscaler is baked into the decode. On Turbo that turns an 8-step 2048² render into a 4096² output in roughly the time it takes to decode once.

The wiring is almost comically short:

Loader (Krea-2-Turbo) → Generate (2048², auto) → Upscale VAE Loader → Upscale Decode 2x → image

Three required inputs: krea2_pipeline (from the Loader), latent (the KREA2_LATENT output of Generate or Multi-Stage Ultra), and upscale_vae (from the Eric Krea2 Upscale VAE Loader). One output: image. There are no widgets to fiddle with, and that's deliberate - the resolution and any downsample/blur behavior are whatever the loaded UPSCALE_VAE was configured with. The loader owns the settings; this node just executes them.

A few honest caveats. This is a 2x linear decode - 4x the pixels - not a hallucinated detail-factory. It's a sharp upscale of what's already in the latent, which is exactly what you want from a Turbo render that's already fully denoised. If you want a quality pass at native size instead of a literal resolution gain, the Upscale VAE Loader has a downsample_to_1x toggle that decodes at 2× then Lanczos-downsamples back - a supersampling-style pass that's usually cleaner than the Wan 2.1 VAE at the same size. That knob is the "wait, I don't actually want 4x the pixels" escape hatch.

The other thing to know: the upscale VAE auto-downloads from Hugging Face on first use (spacepxl/Wan2.1-VAE-upscale2x, Apache-2.0), so the first run of this node can look like a hang while it fetches ~the model. It's fine. Subsequent runs are instant.

When do you not want this node? When you're on the Raw checkpoint doing multi-stage high-res. Raw keeps a resolution-aware trajectory, and the 2× decode sidesteps that entirely - for Raw, the Multi-Stage Ultra node with its controlled re-denoise is the right tool, and you'd use this only as a final optional boost to 8k-class. And it does nothing without the upscale_vae input connected - if you skip the loader, you get nothing, not a fallback.

How to install

Part of the EricRollei/Krea2_ComfyUI_Advanced pack. ComfyUI Manager → search "Krea2_ComfyUI_Advanced", or clone:

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

Restart. The pack needs a git-installed diffusers with Krea2Pipeline and transformers>=4.57.0, plus your own Krea 2 weights. The upscale VAE is the one model this node fetches for you.

Common issues

• Node errors "upscale_vae not connected" - you need the Upscale VAE Loader wired in; the node has no fallback path.
• First run is slow / looks stuck - it's downloading the spacepxl VAE. Give it a minute, then it's cached.
• Output bigger than you wanted - that's 2× linear = 4x area by design. Use the loader's downsample_to_1x or the Multi-Stage node's downsample modes if the resolution blow-up isn't what you wanted.