VAE Loader With Name

The pack's own README describes this one as "as it says on the tin," and that's fair. VAE Loader With Name is the core Load VAE node plus one extra output: a string with the VAE's name. Tiny change, genuinely useful when you're tracking what your workflow actually used.

Worth backing up a step first, because a lot of confusion around VAEs starts here. The VAE is the component that converts between the latent space the diffusion model actually works in and the pixels you see. Load the wrong one (or none), and nothing about the prompt changes - you just get grey, washed-out, desaturated images, and people chase the sampler for an hour before realizing it was never the sampler. Some checkpoints bake their VAE into the file; others expect you to load one separately. For SD 1.5 models that means the classic vae-ft-mse-840000-ema-pruned. SDXL and Flux ship with their VAE included, so on those this node is mostly a way to override rather than a necessity.

How it works

Mechanically it's identical to core: it looks up the file in ComfyUI/models/vae, loads the torch state dict, wraps it in comfy.sd.VAE, and hands it out. The one twist is that before returning the name it strips .safetensors and .pt, so you get vae-ft-mse-840000-ema-pruned instead of vae-ft-mse-840000-ema-pruned.safetensors.

Inputs and outputs

One input, two outputs:

• vae_name - dropdown of everything in ComfyUI/models/vae.
• VAE - wire this into VAE Decode, or anywhere you'd normally send a VAE.
• VAE Name (STRING) - the extension-stripped filename.

That string is the whole point. Wire it into the same pack's Image Info node's optional vae_name input, and every generated caption will record which VAE decoded the image. If you're doing side-by-side VAE comparisons and dumping results with their settings, you don't want to be guessing from filenames later.

When to actually reach for it

Three situations. You're running an SD 1.5 checkpoint with no baked VAE and your colors are flat - this fixes it. You're deliberately swapping a VAE to change the look (the swap-a-VAE-for-better-texture game has been a real thing since the Qwen/Wan VAE compatibility debate, though that applies to much newer models than this node's era). Or you're building the metadata workflow this pack is clearly oriented around. If none of those apply - you just need to decode a latent and don't care about names - the core Load VAE is byte-for-byte the same and you should use it instead.

The one thing to know

Loading a separate VAE here replaces the checkpoint's baked-in one for the whole downstream path. That's usually the point - it's how you fix the washed-out look - but it means you're overriding the model author's choice, so if a model's page says "use its included VAE," this node is how you'd ignore that advice. Keep the "baked vs separate" question in mind when a VAE-swapped image looks subtly off: newer architectures use different channel counts, and a VAE from the wrong family produces noise or flat colour, not a gentle degradation.

Installation is the standard two-liner shared by the whole pack (or ComfyUI Manager, search "ComfyUI_Dragos_Nodes"):

cd ComfyUI/custom_nodes
git clone https://github.com/drago87/ComfyUI_Dragos_Nodes

Then restart ComfyUI. No extra dependencies, no downloads beyond the VAE files you already manage yourself. It's a small, barely-maintained utility pack - the one person who found this node in search apparently clicked it, and the honest take is that its entire value proposition is that little name string. When you're pasting generation info onto your images, that string is worth having.