[SK] VAE

Most of the time you don't need this node at all - modern checkpoints bake their VAE in, and attaching an external one is more likely to hurt output than fix it. But there are real reasons to reach for a standalone VAE loader: an SD 1.5 checkpoint that still expects the classic external 840k decoder, a workflow built around one of the newer pixel-space models that has no VAE and needs a stand-in, or you just want the fast low-fidelity preview decoders instead of the real thing while you iterate. This node covers all three, and it does it with the same folder-tree file picker every loader in this pack uses instead of ComfyUI's flat dropdown.

It's one of eight loaders in SK-ComfyUI-FolderingLoader, a small, largely undocumented pack from GitHub author SKFRMSEHF - no README shipped, no discussion of it anywhere I could find. What makes this particular node worth a closer look is what's baked into its picker: alongside your own VAE files, it ships a builtins folder with pixel_space, taesd, taesdxl, taesd3, and taef1 pre-populated - the same short list ComfyUI's own stock VAE loader exposes.

How it works

A VAE is the codec between pixels and the compressed latent space diffusion actually runs in - it decodes the model's output latent into a viewable image (and encodes images the other way, for img2img). This node just reads a .safetensors VAE file and hands it downstream as a VAE object your sampler's decode step consumes; nothing fancier than that. The built-in options are the exception: taesd/taesdxl/taesd3/taef1 are ComfyUI's tiny approximate autoencoders, one per architecture generation (SD1.x, SDXL, SD3, Flux) - drastically faster than the real VAE and noticeably lower fidelity, good for a live preview while a batch runs, bad for your final output. pixel_space is the option for models that don't have a VAE at all: the 2026 wave of pixel-space diffusion models (HiDream-O1 and similar) generate pixels directly, and a workflow built around that architecture needs something in the VAE socket that's effectively a no-op passthrough rather than a real decoder.

The input that matters

Just vae - the folder-tree combo. Your own external .safetensors VAE files show up under whatever folder structure you've organized them into; the five built-ins sit under their own builtins branch at the top. Output is a single VAE, wired into your KSampler's downstream VAE Decode.

Installing it

No install section came with the pack - the README is empty - so it's the standard two routes: ComfyUI Manager → search SK-ComfyUI-FolderingLoader or the class name → install → restart, or cd ComfyUI/custom_nodes && git clone https://github.com/SKFRMSEHF/comfyui_SK_Loader then restart manually. Nothing extra to download for the node itself - the built-in options are already part of ComfyUI's own runtime, and any external VAE you want just goes in your vae/ folder like normal.

Common issues

• The instinct to "fix colors" with a VAE is a habit from one era. On SD 1.5, missing the external 840k VAE really did cause washed-out, desaturated output. On basically everything since, that instinct actively hurts - most SDXL fine-tunes and everything Flux-onward ship the VAE baked in, and loading a mismatched external one is what causes the grey, faded look people are usually trying to avoid.
• Wrong VAE for the architecture entirely. A VAE only decodes correctly for the latent format it was trained against. An SD 1.5 VAE on an SDXL model, or vice versa, produces noise or flat color, not a subtly-wrong image - it's not a shade of "close enough."
• taesd-family options are for previews, not finals. If your output looks noticeably softer or lower quality than expected and you picked one of the builtins entries, that's expected behavior, not a bug - swap to the model's real VAE for anything you're keeping.
• Empty custom-VAE branch. The tree only lists files already present in your VAE folder at ComfyUI startup; add the file and restart rather than expecting a live refresh.