HY OmniWeaving VAE Loader

Every OmniWeaving workflow needs a VAE, and it can't be just any VAE. The model uses a 3D-causal VAE with 16x spatial and 4x temporal compression - the same family as HunyuanVideo 1.5 - but OmniWeaving-style checkpoints sometimes carry an AutoencoderKLConv3D layout whose state dict is organized differently from what stock ComfyUI's VAE loader expects. HY OmniWeaving VAE Loader is the node that figures out which layout you've handed it and loads it correctly.

The way this plays out in practice: the README's blessed file is hunyuanvideo15_vae_fp32.safetensors (from the vafipas663/HY-OmniWeaving_repackaged repo), which is the plain HunyuanVideo 1.5 VAE and loads fine in stock ComfyUI. But the loader is there so you don't have to care. It inspects the state dict for the OmniWeaving-specific signature - decoder.conv_in.conv.weight instead of decoder.conv_in.weight - and, if present, instantiates the pack's local AutoencoderKLConv3D-equivalent model from a tracked config. If it's a normal Hunyuan 1.5 VAE, it falls back to the stock path. Either way you get a working VAE object.

What you actually set

Just one input:

• vae_name - a dropdown over everything in models/vae. Pick the hunyuanvideo15_vae_fp32.safetensors file (or whatever OmniWeaving VAE you downloaded).

The VAE output is what the rest of the pack consumes. HY OmniWeaving Conditioning uses it to encode your reference image or condition video into the latent that gets injected into the first frame; HY OmniWeaving I2V Semantic Images uses it for the semantic roundtrip; and you'll also hang a stock VAE Decode off it at the end to turn latents back into frames.

Installing it

Same drill as every node in this pack. ComfyUI Manager (search "hy_omniweaving") or:

cd ComfyUI/custom_nodes
git clone https://github.com/Shiba-2-shiba/hy_omniweaving_comfyui_unofficial

Restart ComfyUI. No pip dependencies, needs a recent ComfyUI (the pack uses the comfy_entrypoint extension API). The VAE itself is the smallest of the downloads; the fp32 file is a few hundred MB, and the heavy lifting is the diffusion model and the 7B Qwen encoder.

Common issues

• "Failed to load VAE" or an instantiation error - you've pointed it at some other model's VAE (e.g. a Wan or LTX file). It has to be the HunyuanVideo 1.5 / OmniWeaving VAE.
• Decoder warnings about conv_in shapes - a sign the file uses the non-standard OmniWeaving layout, which is exactly what this loader exists to absorb. If you see an actual traceback, double-check you're on the latest pack commit; the config detection has been patched a few times.
• If the stock VAELoader "works" - fine, you don't strictly need this node for the plain hunyuanvideo15_vae_fp32.safetensors file. But if you ever swap in a repackaged checkpoint with the conv_in.conv layout, this is the loader that won't explode.

One thing that trips people up: the VAE here is a plain inference decoder/encoder, and it isn't where OmniWeaving's "semantic first frame" magic lives. That's the I2V Semantic Images node's job. The VAE just needs to be the right one - and now it is.