Documentation
¶
Overview ¶
Package vae provides shared utilities for VAE (Variational Autoencoder) operations.
Index ¶
Constants ¶
This section is empty.
Variables ¶
This section is empty.
Functions ¶
func DecodeTiled ¶
func DecodeTiled(latents *mlx.Array, cfg *TilingConfig, decoder func(*mlx.Array) *mlx.Array) *mlx.Array
DecodeTiled decodes latents using tiled processing with overlap blending. This reduces memory usage for large images by processing in overlapping tiles.
Parameters:
- latents: [1, H, W, C] latent tensor in NHWC format
- cfg: tiling configuration (tile size and overlap)
- decoder: function to decode a single tile [1, H, W, C] -> [1, H*scale, W*scale, 3]
Returns: [1, 3, H*scale, W*scale] decoded image in NCHW format
Types ¶
type TilingConfig ¶
type TilingConfig struct {
TileSize int32 // Tile size in latent space (e.g., 64 latent → 512 pixels for 8x VAE)
Overlap int32 // Overlap in latent space (e.g., 16 latent = 25% of 64)
}
TilingConfig holds configuration for tiled VAE decoding. This is a general technique to reduce memory usage when decoding large latents.
func DefaultTilingConfig ¶
func DefaultTilingConfig() *TilingConfig
DefaultTilingConfig returns reasonable defaults matching diffusers. tile_latent_min_size=64, tile_overlap_factor=0.25
Click to show internal directories.
Click to hide internal directories.