⚡ Gigachad VAE Loader

Gigachad VAE Loader is a VAE loader with three dials the stock one doesn't give you: which device to run it on, what dtype to hold it in, and - the genuinely odd one - a pixel_space option that isn't a file at all. Based on KJNodes' popular VAELoaderKJ, it's the kind of node you install for one feature and end up using for all of them.

The three inputs that matter

• vae_name - the dropdown. It lists everything in your models/vae folder, plus a bunch of auto-detected "approx" VAEs from vae_approx: the tiny TAESD decoders (taesd, taesdxl, taesd3, taef1) and the video TAEs (taehv, lighttaew2, lighttaehy). Those TAEs are the fast preview/quick-decode models - not production quality, but cheap enough for live previews and rough passes. And then there's pixel_space.
• device - main_device (your GPU) or cpu. A VAE is comparatively small, so offloading it to CPU to free VRAM for the diffusion model is a legitimate play on tight cards.
• weight_dtype - bf16, fp16, or fp32. On an 8GB card, bf16 is the usual sweet spot; fp32 is there for when you care about bit-exact output more than memory.

The pixel_space mode nobody explains

Here's the trick. Some modern models - notably the pixel-space diffusion family - don't use a latent VAE at all. Their "latents" are already in pixel space, and forcing a real VAE through encode/decode would corrupt the pipeline. Selecting pixel_space loads a stub VAE that exists solely to signal ComfyUI "skip the VAE math here." You don't pick it for normal checkpoints; you pick it when your model's docs tell you to. If you're on SD1.5/SDXL/Flux, leave it alone.

How it works

Behind the scenes it mirrors comfy.sd.VAE construction with explicit device and dtype args, which is what gives you the per-run control. The video TAE detection means you can drop the LTX or Wan "tiny approximate encoder" into a video pipeline without hunting down the right filename yourself - the node finds taehv/lighttaew2 files and serves them up.

Install

Same as the whole pack:

cd ComfyUI/custom_nodes
git clone https://github.com/Winnougan/comfyui-gigachad.git

or search comfyui-gigachad in ComfyUI Manager, then restart. No requirements.txt - pure ComfyUI, nothing extra to install. The approx VAEs only appear in the list if you've already downloaded them into models/vae_approx, so don't panic when the TAESD entries don't show up on day one.

Where this earns its place: if you've ever run out of VRAM and wished you could push the VAE to CPU, or wished the stock loader exposed dtype, this removes the friction. Just remember that the single output is a vae, and it goes wherever a VAE goes - into the encode/decode nodes or the checkpoint's VAE slot if you're swapping it.