HunyuanVideo 1.5 Leo VAE Model Loader

In the complete HunyuanVideo 1.5 workflow, this is the boring node - and that's the compliment. HyVideo15VaeLoader hands you the model's 3D causal VAE, which is what compresses video into the latent space the transformer works in (16x spatial, 4x temporal compression) and decompresses it back to pixels on the way out. It's the "Leo VAE Model Loader" in the menu, a leftover display name from the diffusers wrapper lineage this pack is built on; functionally it's just this plugin's own VAE loader.

What it does

One input that matters, one that occasionally matters:

• path - a dropdown of folders under ComfyUI/models/vae. Leave it on None and the loader auto-downloads vae from the tencent/HunyuanVideo-1.5 HuggingFace repo into models/vae/hyvideo15 the first time you run, then never again. Set it to a real folder if you've already placed the VAE yourself.
• enable_tile_parallelism - on by default. The 1.5 VAE can be a memory hog when decoding full-resolution video, and this tiles the encode/decode work so it fits on consumer cards. Keep it on unless you're chasing every last bit of speed and have VRAM to spare.

Under the hood the loader does one nice thing you don't have to think about: it sniffs your GPU memory and picks VAE settings to match. Under 28GB it drops to fp16 with a 128px sample tile; at 28GB or above it uses fp32 with a 256px tile. That auto-config is why it "just works" on both a 12GB card and a 5090.

The single output, model (HYVID15VAE), feeds the encode/decode and latent-prep nodes: HyVideo15VaeEncode, HyVideo15VaeDecode, HyVideo15LatentsPrepare, and the SR-side nodes.

Installing it

It's part of comfyui_hunyuanvideo_1.5_plugin - ComfyUI Manager, search "HunyuanVideo-1.5 nodes", or:

cd ComfyUI/custom_nodes
git clone https://github.com/yuanyuan-spec/comfyui_hunyuanvideo_1.5_plugin
cd comfyui_hunyuanvideo_1.5_plugin
pip install -r requirements.txt

Then restart. The pinned requirements (diffusers==0.35.0, transformers==4.57.1, torch>=2.6.0) should go into ComfyUI's own Python environment, and FlashAttention is recommended pack-wide.

Common issues

The VAE download itself is the thing most likely to fail, usually because the auto-download shells out to the hf CLI and a partial or interrupted download leaves a half-populated models/vae/hyvideo15. The plugin checks whether the folder exists, not whether it's complete, so a broken folder can silently defeat future auto-downloads. If you see weird decode artifacts or a crash, delete models/vae/hyvideo15 and let it re-download, or grab the vae folder from tencent/HunyuanVideo-1.5 manually. Network-speed related, the README suggests HF_ENDPOINT=https://hf-mirror.com if you're in a region where HuggingFace is slow.

One more thing worth knowing before you build a workflow around this stack: HunyuanVideo 1.5 is a strong but niche model in the 2026 video landscape - its I2V quality is genuinely good, but the ecosystem momentum has moved to Wan and LTX. If you're here for the complete workflow, it's because you specifically want 1.5's first-frame fidelity or its 1080p SR pass, and this loader is simply the piece that makes that possible.