ChronoEdit VAE (CRT AutoDL)

The most forgettable node in a ChronoEdit workflow, and the one that breaks everything if you skip it. ChronoEdit is NVIDIA's image editor built on Wan's I2V architecture, and Wan models use Wan's VAE. Grabbing the SDXL VAE that's already sitting in your vae folder won't cut it - the latent space doesn't match, and you'll decode garbage. This node makes sure the right VAE exists and hands it to you.

What it loads

Zero inputs, one output: VAE. On first run it checks ComfyUI/models/vae for wan_2.1_vae.safetensors, and if it's absent it downloads it from Comfy-Org's Wan 2.1 repackaged split files on HuggingFace. Wan's VAE is a 3D video VAE rather than a 2D image one, which is exactly why it's a separate download and why a "normal" image VAE won't work. Once the file exists the node is instant; the download is a one-time thing and it's small compared to the 14GB model next to it.

Where it sits in the workflow

ChronoEdit VAE → VAE Encode (start image → latent)
ChronoEdit VAE → VAE Decode (edited latent → image)

That's the whole job. If you're running the image-edit flavor of ChronoEdit you encode your input image to latent with this VAE, run the sampler on the patched ChronoEdit model, and decode the last frame back out. The last frame is your edit - that's the trick of the whole architecture.

The pack's other ChronoEdit AutoDL nodes each pull their own matching file: ChronoEdit Model (the fp8 14B), ChronoEdit CLIP - WAN (the umt5-xxl text encoder), ChronoEdit CLIP Vision for reference-image work, and the two LoRA nodes for distilled steps and upscaling. Because every one of them downloads independently on first use, your first run of a full ChronoEdit graph is a series of progress bars. Set it going, make a coffee, and never think about the VAE again.

Install

It's the pack, nothing VAE-specific. ComfyUI Manager → search CRT-Nodes, or:

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

Restart ComfyUI. The pack's requirements are heavy overall (opencv, transformers, audio tooling for the non-AutoDL nodes), but this loader just needs huggingface_hub, which ComfyUI bundles. The model file itself downloads through the node on first run, not through any manual step.