Implementation:Deepseek ai Janus Noise Initialization

Knowledge Sources	Janus
Domains	Image_Generation, Diffusion_Models
Last Updated	2026-02-10 09:30 GMT

Overview

Pattern for sampling Gaussian noise in the SDXL VAE latent space as the starting point for rectified flow ODE solving.

Description

This user-defined pattern initializes the latent noise tensor and computes the Euler step size for the ODE loop. The latent space matches the SDXL VAE's configuration: 4 channels at 48×48 spatial resolution.

Usage

Implement this pattern after CFG input preparation and before the ODE denoising loop.

Code Reference

Source Location

Repository: Janus
File: demo/app_janusflow.py
Lines: L86-89

Pattern Implementation

# Initialize Gaussian noise in SDXL VAE latent space
# 4 channels, 48x48 spatial (for 384px images, 8x downscale)
z = torch.randn((parallel_size, 4, 48, 48), dtype=torch.bfloat16).cuda()

# Euler ODE step size
num_inference_steps = 30
dt = 1.0 / num_inference_steps
dt = torch.zeros_like(z).cuda().to(torch.bfloat16) + dt

Import

import torch

I/O Contract

Inputs

Name	Type	Required	Description
parallel_size	int	Yes	Number of images to generate
num_inference_steps	int	No	ODE steps (default 30)

Outputs

Name	Type	Description
z	torch.Tensor [parallel_size, 4, 48, 48]	Gaussian noise in VAE latent space (bfloat16, CUDA)
dt	torch.Tensor [parallel_size, 4, 48, 48]	Euler step size broadcast to latent shape

Usage Examples

Standard Initialization

parallel_size = 5
num_inference_steps = 30

z = torch.randn((parallel_size, 4, 48, 48), dtype=torch.bfloat16).cuda()
dt = 1.0 / num_inference_steps
dt = torch.zeros_like(z).cuda().to(torch.bfloat16) + dt

# z: initial noise, shape [5, 4, 48, 48]
# dt: step size, shape [5, 4, 48, 48] (broadcast-compatible scalar)

Related Pages

Implements Principle

Principle:Deepseek_ai_Janus_Noise_Initialization

Requires Environment

Page Connections

Double-click a node to navigate. Hold to expand connections.

Principle

Implementation

Heuristic

Environment