Krea 2 VAE (CRT AutoDL)

Krea 2 was one of the bigger open-model moments of mid-2026 - a 12B diffusion transformer that Krea trained from scratch and released in June under a genuinely commercial-friendly license, and it took the "best base to train on" title away from Ideogram 4 within about a week, mostly on the strength of how easy it is to LoRA-train. If you're setting up a Krea 2 workflow, this node handles one specific, easy-to-get-wrong piece of it: the VAE.

What it actually does

Zero inputs, one output: VAE. That's the entire node signature, and it tells you exactly what's going on - this isn't a loader that points at a file you already have, it's a node from the pack's "AutoDL" family, meaning it fetches the model weights itself the first time you run it. The pack's LTX 2.3 AutoDownload node does the same thing explicitly (built-in HTTP download plus API routes for status/progress), and this node's naming and empty input list strongly suggest the same pattern here - you drop it into your graph, hit run, and it handles getting the right VAE file into place rather than you hunting down a HuggingFace repo and dropping a .safetensors into models/vae yourself.

Why it matters which VAE: Krea 2 doesn't use Flux's VAE. Krea deliberately chose the Qwen-Image VAE instead, because it suits stylized and illustrative work better than photorealism - Krea's own CEO has said on record he'd use the Flux VAE if photorealism were the priority. That's a real architectural choice with a real consequence for output, so pairing a Krea 2 checkpoint with the wrong VAE isn't a small mismatch, it's a different model's assumptions leaking into your decode step.

The output that matters

There's exactly one: VAE, which you wire straight into your VAEDecode (and VAEEncode, if you're doing image-to-image or inpainting) alongside a Krea 2 UNETLoader and the Qwen3-VL-based CLIPLoader that Krea 2 also requires. ComfyUI 0.25.0 shipped native krea2 loader support around the same time as the model's release, so this node isn't replacing anything broken in stock ComfyUI - it's just saving you the manual step of locating and placing the correct VAE file before your first run.

How to install it

This is one node inside the much larger CRT-Nodes pack. Through ComfyUI Manager, search CRT-Nodes and install. Manually:

cd ComfyUI/custom_nodes
git clone https://github.com/PGCRT/CRT-Nodes.git
pip install -r CRT-Nodes/requirements.txt

then restart. You're pulling in the whole pack's dependency list (OpenCV, Ultralytics, spandrel, and more) to get this one auto-downloader - normal for CRT-Nodes, since it ships as a single suite rather than granular installs.

Common issues & troubleshooting

First run is slow, and that's expected. An auto-download node has to actually fetch the weights the first time it executes, so don't panic if your first Krea 2 run stalls for a while on this node specifically - that's the download, not a hang. Subsequent runs should be fast since the file is now cached locally.

Mismatched VAE elsewhere in the graph. If you're getting muddy, low-detail, or oddly color-shifted decodes on an otherwise-working Krea 2 workflow, check that you haven't accidentally got a Flux or SDXL VAE wired in from an older template - Krea 2 specifically wants the Qwen-Image VAE, and this node exists precisely so you don't have to remember that.

Krea 2's open weights behave differently from the hosted version. Not this node's fault, but worth knowing if your local output looks softer or less expressive than what you've seen from Krea's own website: Krea has confirmed the released weights went through an alignment pass the hosted API version didn't, so a gap in prompt adherence and emotional range between local and hosted is a known, acknowledged difference - not something a VAE swap will fix.