HFRemoteVAEDecode

HFRemoteVAEDecode is the direct version of its sibling. Where HFRemoteVAE(Decode Only) hands you a VAE object to plug into a standard VAEDecode node, this one cuts the middleman out: you feed it a latent tensor and it returns actual pixels, one node, done. If your workflow doesn't already have a VAE slot wired up, this is the one you'll actually reach for.

Same premise as the whole pack: instead of loading a local VAE into VRAM and decoding on your GPU, it sends the latents as raw binary bytes to a free Hugging Face-hosted decode endpoint and gets the result back. Your machine never loads a VAE, and you don't even need to download one. That's the entire point - no weights, no model files, just a network call that returns an image.

What you give it

Two inputs, both worth knowing:

• samples - a LATENT, straight from your KSampler (or from a model that produces latent video). This is the thing being decoded.
• VAE_type - the dropdown with Flux, SDXL, SD, HunyuanVideo. This is the whole config, and it's the one place you can screw up: each architecture has its own incompatible autoencoder, so pick what you sampled with. A wrong choice gives you noise or flat colour, not a subtly worse image.

The output is images, an IMAGE tensor - the standard thing you'd wire into a preview, a save node, or any further image-processing node. Nothing fancy, which is the appeal.

When to use this over HFRemoteVAE

They share a brain, so it's about your graph's shape. If you have an existing workflow with a VAE slot feeding VAEDecode - the usual layout for text-to-image - HFRemoteVAE is the drop-in replacement for the loader. But if you just want latents → pixels in one hop, or you're building a workflow from scratch, use this and skip the extra node. It's also the cleaner pick for video: feed it HunyuanVideo latents and you get frames back without ever touching the (comparatively heavy) video VAE locally.

Installing it

Identical to the other node in the pack - it's the same repo. ComfyUI Manager → search "HFRemoteVae" → install, or manually:

cd ComfyUI/custom_nodes
git clone https://github.com/kijai/ComfyUI-HFRemoteVae

Restart, and both nodes appear under the HFRemoteVae category. The only dependency is diffusers, which ComfyUI already has, and there are no model downloads at all. This is about as painless as a Kijai pack gets.

The honest caveats

Same ones as the pack: these are free public endpoints with no SLA, so a decode can be slow or rate-limited, and if it fails it's almost always the network rather than your settings. Every decode also ships your latent bytes to a Hugging Face server, so this is the wrong tool if you're generating anything sensitive. And a little perspective: for SD and SDXL the local VAE is tiny, so you're trading a modest VRAM saving for a network round trip. The interesting cases are video decode and the genuinely VRAM-starved machine. If that's you, this node quietly makes a step that used to cost you memory cost you nothing local at all.