Flux2Klein VAE (CRT AutoDL)

Zero widgets, one job: run this node and it fetches the VAE Flux 2 Klein needs to decode its latents into pixels, and hands you back a ready VAE. It's the third leg of a from-scratch Klein setup alongside the model and CLIP AutoDL loaders - skip the manual HuggingFace-and-folder-placement step entirely.

Why this matters for Klein specifically

Klein ships in four checkpoints - 4B and 9B, each in a base and a distilled flavor - and all four share the same generation-plus-editing architecture. What all four also need is a correctly matched VAE; that part doesn't change across the size/distillation split the way the diffusion weights and text encoder do, which is exactly why an auto-download node built for "Flux2Klein" generally (not tied to one specific size in its name) makes sense here. If you're new to Klein: don't assume you can reuse a VAE from Flux.1 or from a different model family - decode will either error outright on a shape mismatch or, worse, silently produce garbled color and noise if something upstream tolerates the mismatch just enough to run.

The inputs and outputs that matter

• No required or optional inputs.
• Output: a single VAE.

That's the whole node. There's no widget to pick between checkpoint sizes because the VAE doesn't vary by size the way the model and text encoder do.

Installing it

Comes bundled with the rest of CRT-Nodes - no separate install:

ComfyUI Manager - search CRT-Nodes, install, restart.

Manual:

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

Restart ComfyUI. The VAE file itself downloads the first time you run this node in a workflow, not at pack-install time. VAEs are small relative to the diffusion model or the LLM-based text encoder, so this should be one of the quicker AutoDL downloads in the family - still, give it a moment on first use before assuming something's wrong.

Common issues

• Decode output is noisy or clearly wrong - the standard VAE-mismatch symptom. Confirm this node's output is actually wired into the VAE Decode step of your Klein pipeline and that nothing upstream is still pointed at a leftover VAE from a different model in the same graph.
• First run pauses - that's the download; check your ComfyUI console for HTTP activity rather than assuming the node has hung.
• Building a mixed pipeline (say, generating on one model and doing a Klein edit pass on the output) - make sure you're not accidentally feeding a non-Klein latent into a Klein-VAE decode step, or vice versa. The VAE has to match whichever model actually produced the latent you're decoding.
• Repeated downloads on every run - check that CRT-Nodes actually has write access to wherever it's caching the fetched file; if the download can't persist between runs, you'll pay the download cost every time instead of just once.