DreamX Wan2.2 VAE Loader

DreamX borrowed Wan 2.2's video VAE the way it borrowed Wan's text encoder. DreamX Wan2.2 VAE Loader loads Wan2.2_VAE.pth from the model root and gives you a standard ComfyUI VAE - the same object type your stock VAE Decode and VAE Encode nodes already accept.

One input, model_root; one output, vae. No dtype widget, which is worth noticing: the file on disk is what it is, and comfy.sd.VAE handles the rest. Community consensus on this lineage is that the Wan VAE is a straight upgrade over the 16-channel alternatives it gets confused with, so there's nothing to tune.

Where it gets used

Twice in a normal generation graph, and one of those uses is easy to miss.

First, DreamX First Frame AV Latent encodes your start image through it. That node hard-checks the result is a 48-channel latent - (1, 48, 1, H/16, W/16) - because a 16-channel Wan 2.1 VAE would produce something plausible-looking and completely wrong. If you mix the two up, you get an explicit error naming the expected shape rather than a grey smear, which is a kindness.

Second, the decoded video latent goes back through it after sampling. The shipped example uses native VAEDecodeTiled at 512 px with 64-frame temporal tiles, per the pack's docs - on long clips, cuDNN can report memory pressure as a hard execution failure rather than a normal OOM, which bypasses ComfyUI's automatic VAE fallback. Tiling up front avoids that.

You also need it for the refiner workflow. The shipped examples/dreamx_refiner_ui.json loads UMT5, this VAE, and the refiner; DreamX Causal Refine Video tiles its own Wan VAE encode and decode internally, with the same conservative defaults the native tiled nodes use, so you don't need to think about tile sizes there.

Not needed if you're using DreamX Creator Complete Loader - that already emits vae.

Channels, in one paragraph

Channel count went up and stayed up: SDXL's VAE carries 4, Flux and Wan 2.1 carry 16, and channel count is depth, not compression - the downscale is the same. DreamX reads a 48-channel video VAE because Wan 2.2's TI2V line does. The practical rule from concepts.md applies: a VAE from one channel count on another model produces noise or flat colour, not a subtly wrong image. Don't get creative here.

Install

ComfyUI Manager → DreamX Creator T8. Manual:

cd ComfyUI/custom_nodes
git clone https://github.com/T8mars/Comfyui-DreamX-Creator-T8.git
cd Comfyui-DreamX-Creator-T8
python -m pip install -r requirements.txt

Weights are separate, ~54 GB total, and this one needs wan2.2_ti2v_5b/Wan2.2_VAE.pth sitting directly inside the model root:

python -m pip install -U huggingface_hub
hf download t8star/DreamX-Creator-Comfy --local-dir ComfyUI/models/dreamx_creator
python scripts\verify_models.py

If you'd rather not pull the whole bundle for the refiner workflow only, the pack notes that wan2.2_ti2v_5b/ can come straight from Wan-AI/Wan2.2-TI2V-5B - but you need all three entries: the VAE, models_t5_umt5-xxl-enc-bf16.pth, and the google/umt5-xxl/ tokenizer folder.

Common issues

DreamX requires the Wan2.2 48-channel video VAE. You pointed model_root at something with a different VAE, or a Wan 2.1 install. That guard is the best error message in the pack.

A stale-looking or corrupted VAE file. This checkpoint is pickle-backed and gets SHA-256 checked against the pack's manifest before it's loaded tensor-only. Truncated downloads fail loudly with a size or hash mismatch. Re-fetch the single file.

"Model root was not found", with a list of paths. auto searches the repo's checkpoints/ and then ComfyUI/models/dreamx_creator/. If your download created a nested DreamX-Creator-Comfy/ directory, remove that level.

Refiner decode fails on a long clip. The refiner rejects outputs above 8 megapixels and tiles internally; if you're hitting cuDNN errors elsewhere, keep using tiled native decode for the generation pass.