Nodes/WAS Affine/Wan VAE Decode
ComfyUI Node

Wan VAE Decode

A ComfyUI node in latent with 11 inputs and 1 output.

By WASasquatch·Created 12 months ago·Updated 4 months ago· 5
Wan VAE Decode
  • vae
  • latents
  • image
horizontal_tiles1
vertical_tiles1
overlap1
last_frame_fixfalse
frames_per_pass0
batch_per_pass0
max_decode_megapixels0.0
accumulate_dtypefp32
output_dtypeauto
Categorylatent

Inputs (11)

NameTypeDefaultDescription
vaeVAEVAE used to decode the latents. Must correspond to the model that produced the latents. Respects the VAE's device/dtype and internal time/space scaling.
latentsLATENTLatent dict with key 'samples' shaped [B,C,F,H,W]. F=frames (F=1 for images). H/W are latent-space, not pixel-space.
horizontal_tilesINT11–12Tiles across WIDTH in latent space.
vertical_tilesINT11–12Tiles across HEIGHT in latent space.
overlapINT10–16Overlap in LATENT pixels. Output overlap = overlap * VAE scale.
last_frame_fixBOOLEANfalseAppend last latent frame before decode, trim after. For video VAEs.
frames_per_passoptINT00–64Limit frames per pass (0 = all).
batch_per_passoptINT00–64Limit batch per pass (0 = all).
max_decode_megapixelsoptFLOAT0.00–400VRAM budget ≈ decH * decW * out_frames * batch / 1e6. 0 = off.
accumulate_dtypeoptCOMBOfp32Accumulator dtype for blending.
output_dtypeoptCOMBOautoFinal NHWC dtype ('auto' = fp16 on CUDA, else fp32).

Outputs (1)

NameTypeDescription
imageIMAGE