Implementation:Deepseek ai Janus AutoencoderKL Decode
| Knowledge Sources | |
|---|---|
| Domains | Image_Generation, Generative_Models |
| Last Updated | 2026-02-10 09:30 GMT |
Overview
HuggingFace Diffusers AutoencoderKL.decode method used to convert denoised latents to pixel images in the JanusFlow pipeline.
Description
The AutoencoderKL from the diffusers library provides the SDXL VAE decoder. In JanusFlow, the denoised latent is divided by the VAE's config.scaling_factor (0.13025) before being passed to decode(). The output is in range [-1, 1].
External Reference
Usage
Call after the ODE denoising loop completes. The VAE must be loaded separately from the main model.
Code Reference
Source Location
- Repository: External — HuggingFace Diffusers
- Usage reference: demo/app_janusflow.py:L134
Signature
# Called via: vae.decode(z / vae.config.scaling_factor).sample
AutoencoderKL.decode(
z: torch.Tensor, # [B, 4, 48, 48] latent divided by scaling_factor
) -> DecoderOutput # .sample gives [B, 3, 384, 384]
Import
from diffusers.models import AutoencoderKL
I/O Contract
Inputs
| Name | Type | Required | Description |
|---|---|---|---|
| z | torch.Tensor [B, 4, 48, 48] | Yes | Denoised latent divided by scaling_factor (0.13025) |
Outputs
| Name | Type | Description |
|---|---|---|
| decoded_image | torch.Tensor [B, 3, 384, 384] | Pixel images in range [-1, 1] |
Usage Examples
VAE Decode After ODE Loop
# z: denoised latent [5, 4, 48, 48] from ODE loop
decoded_image = vae.decode(z / vae.config.scaling_factor).sample
# decoded_image shape: [5, 3, 384, 384], range [-1, 1]
Related Pages
Implements Principle
Requires Environment
- Environment:Deepseek_ai_Janus_CUDA_GPU_Environment
- Environment:Deepseek_ai_Janus_JanusFlow_Diffusers_Environment
Uses Heuristic
Page Connections
Double-click a node to navigate. Hold to expand connections.
Principle
Implementation
Heuristic
Environment