Principle:Deepseek ai Janus Noise Initialization
| Knowledge Sources | |
|---|---|
| Domains | Image_Generation, Diffusion_Models |
| Last Updated | 2026-02-10 09:30 GMT |
Overview
A step that initializes Gaussian noise in the VAE latent space as the starting point for the rectified flow ODE solver.
Description
In rectified flow generation, images are produced by solving an ordinary differential equation (ODE) that transports samples from a Gaussian noise distribution to the data distribution. The noise initialization step samples the starting point for this ODE from a standard Gaussian distribution in the latent space of the SDXL VAE.
The latent space has 4 channels at a spatial resolution of 48×48 (for 384×384 output images with the SDXL VAE's 8× downscaling factor).
Usage
Use this principle after CFG input preparation and before entering the ODE denoising loop.
Theoretical Basis
Rectified flow defines a transport map from noise to data:
Where z_0 is the noise sample (t=0), z_1 is the target image (t=1), and z_t follows a straight-line (rectified) path.
The noise initialization samples:
The Euler ODE step size is: Failed to parse (syntax error): {\displaystyle dt = \frac{1}{\text{num\_inference\_steps}}}