ERNIE VAE (CRT AutoDL)

The simplest node in the AutoDL family: no widgets at all, just run it and it fetches ERNIE-Image's VAE and hands you a ready VAE output. If you're building a from-scratch ERNIE workflow, this is the third piece alongside the ERNIE Model and CLIP/text-encoder loaders - the component that turns the model's latent output back into pixels.

Why ERNIE needs its own VAE

ERNIE-Image runs in latent space (unlike the pixel-space wave of models that dropped the VAE entirely around the same era), and per the model's own architecture it uses a 128-channel VAE - a notably wider latent than the 16 channels most contemporary architectures use. That's the property the model's trainability advocates point to when arguing ERNIE converges fast as a LoRA target, though it's worth saying plainly: that's a community claim with limited first-hand testing behind it, not something Baidu itself has made a big case for. Practically, what matters for you is simpler - ERNIE's VAE isn't interchangeable with Flux's, Klein's, or Z-Image's. Load the wrong one and you'll get garbage or an outright shape-mismatch error out of the VAE Decode step.

The inputs and outputs that matter

There's genuinely nothing to configure:

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

That's the entire schema. This node exists purely to save you the trip to HuggingFace and the models-folder placement ritual for one specific, correctly-matched file.

Installing it

It ships with the rest of CRT-Nodes - nothing to install separately for this node specifically:

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. As with the rest of the AutoDL family, the actual VAE weights download the first time you run the node in your graph, not at pack-install time - and this specific node isn't broken out in the CRT-Nodes README's own catalog table, so there's no dedicated documentation beyond what the class name and output type tell you.

Common issues

• First run pauses with no obvious feedback - that's the download happening. VAEs are much smaller than the diffusion model itself, so this should be quick, but check your console log if it seems to hang.
• Decoded images come out wrong (color noise, garbled output) - the classic VAE-mismatch symptom. Make sure this node's VAE output is actually wired to the decode step in an all-ERNIE pipeline, and that you haven't accidentally left a different model's VAE loaded elsewhere in the graph and forgotten to swap the wire.
• Wondering whether you even need this node - you do, if you're building an ERNIE workflow from individual loader nodes rather than a pre-built template. Some workflow templates bundle model+VAE+CLIP loading differently; if you're using one of those, you may not need this node at all. Use it when you're wiring ERNIE up from scratch via CRT-Nodes' AutoDL loaders specifically.
• Repeated re-downloads - if the node seems to fetch the VAE again on every run instead of caching it, check where CRT-Nodes is placing the downloaded file and confirm ComfyUI has write access to that location; a download that can't persist to disk will re-fetch every time.