DreamX Audio VAE Loader

If you've been taught "bfloat16 is the default for everything modern," this node is the exception that will cost you an afternoon. DreamX Audio VAE Loader loads the CreatorDACVAE checkpoint that turns DreamX's 128-channel audio latent back into a 48 kHz waveform - and its own author tells you, in the field tooltip, to leave the dtype at float32. That is not laziness. The dtype dropdown here defaults to float32 while every other loader in the pack defaults to bfloat16, and the shipped DreamX Creator Complete Loader hard-pins fp32 for this piece even when you asked for bf16 elsewhere. Match that.

What it loads and from where

Two inputs, one output. model_root defaults to auto, which searches the node-local checkpoints/ directory first and then ComfyUI/models/dreamx_creator/. You can also paste an explicit path. The dtype combo offers float32, bfloat16 and float16 - and again, the tooltip is the author's own note that the released Audio VAE decodes in float32.

The output is audio_vae, typed DREAMX_AUDIO_VAE. That's a custom socket type, so it only plugs into DreamX Audio VAE Decode; don't expect it to accept a normal VAE or to feed a stock decode node. It's a handle with a patcher attached, which is how ComfyUI can offload it separately from the video VAE and the text encoder.

Why fp32 matters here

Internally the loader builds a CoreModelPatcher on the VAE's offload device and the decode path calls load_models_gpu on it. On CUDA and XPU it wraps the decode in autocast only when the weights are not already float32 (enabled = device_type in {"cuda","xpu"} and dtype != torch.float32). So the precision you pick here decides whether the decode runs under autocast at all, and the released decoder path - the one the model was validated on - is the plain fp32 one. bf16 audio decode is the kind of change that sounds free and shows up as a thin, gritty top end, or as a straight-up numerical mess. It's a small model next to the 7B generator; there's no reason to gamble on it.

You also don't need this node if you're using the bundle. DreamX Creator Complete Loader outputs audio_vae already, at fp32, alongside model, clip and vae. Reach for this loader when you're building the graph by hand or swapping the audio VAE independently of everything else.

Install

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

Or search DreamX Creator T8 in ComfyUI Manager and restart.

The weights are a separate ~54 GB download (20 files, from t8star/DreamX-Creator-Comfy). The audio piece is audio_vae/config.json plus audio_vae/diffusion_pytorch_model.safetensors inside the model root:

hf download t8star/DreamX-Creator-Comfy --local-dir ComfyUI/models/dreamx_creator
python scripts/verify_models.py

Run the verifier before you load 50 GB of weights, not after. Note the layout requirement: the root you point at must contain creator/, audio_vae/, refiner/ and wan2.2_ti2v_5b/ directly. If your download nested itself one level deeper as DreamX-Creator-Comfy/, model_root=auto won't find it and you'll get the classic "model root was not found" complaint listing the paths it checked.

Requirements note: the pack deliberately does not install torch - ComfyUI supplies it. What it does add is real, though: diffusers>=0.37.1,<0.38, transformers>=4.57,<5, peft, accelerate and friends. That's the standard node-pack dependency squeeze the audio layer lives in. If you have another pack pinned to an older transformers, sort it out before you blame the VAE.