VAE
The part that turns a latent into a picture and a picture into a latent.
In plain words
A VAE (variational autoencoder) is a small model with two halves. The encoder squeezes a picture into a latent. The decoder unfolds a latent back into pixels. The sampler only ever sees latents, so every graph that shows a result ends with a VAE decoding it, and every graph that starts from a picture begins with a VAE encoding it.
Each family has its own latent "language", so each family has its own VAE. SD 1.5 and SDXL VAEs differ; Flux.1, Chroma and Z-Image share ae.safetensors; Wan 2.1 and the Wan 2.2 14B models share wan_2.1_vae, while Wan 2.2 5B needs wan2.2_vae; Qwen-Image has qwen_image_vae.
Where you meet it in a graph
- Load VAE (
VAELoader) reads a file frommodels/vae; Load Checkpoint also has aVAEoutput. - VAE Decode (
VAEDecode): inputssamples(LATENT) andvae; outputIMAGE. - VAE Encode (
VAEEncode): inputspixels(IMAGE) andvae; outputLATENT. - VAE Encode (for Inpainting), InpaintModelConditioning, WanImageToVideo and other start nodes take a
vaeinput too. - Tiled versions (VAE Decode (Tiled), VAE Encode (Tiled)) work in pieces to save memory on large pictures and long videos.
Values and what they change
There are no strengths to set. What matters is the right file. Tiled nodes have tile_size and overlap fields: smaller tiles use less memory and take longer; too little overlap shows seams.
Typical mistakes
- VAE of another family: washed-out or wrong colours, a grid pattern, noise, or an error about channels.
- A broken VAE in an old checkpoint: some early SDXL and SD 1.5 checkpoints have a VAE that gives black pictures in half precision (fp16). A separate fixed VAE in Load VAE cures it.
- Out of memory at the very end: decoding a big picture or a long clip needs a lot of memory. Use the tiled decode.
- Decoding the latent of one model with the VAE of another in a two-model graph: each picture is decoded by the VAE of the model that drew it.
How the app helps
The descriptions of VAE nodes say what the VAE does and that it must match the family. Check names models of different families in one path. The lessons tell exactly which VAE file each video graph needs.