Principle:Deepseek ai Janus Image Post Processing

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

Overview

A procedure for converting decoded image tensors from normalized float values to displayable uint8 images and saving them to disk.

Description

After the VQ-VAE decoder produces pixel tensors in the range [-1, 1], post-processing converts them to standard image format:

1. Convert from NCHW to NHWC layout
2. Rescale from [-1, 1] to [0, 255]
3. Clip values and cast to uint8
4. Create PIL Image objects and optionally save to disk

This is a straightforward numerical transformation that bridges the gap between the model's output representation and standard image file formats.

Usage

Use this principle as the final step in the autoregressive image generation pipeline, after VQ-VAE decoding.

Theoretical Basis

The rescaling formula:

${\text{pixel}}_{uint8}=\text{clip}(\frac{{\text{pixel}}_{float}+1}{2}\times 255,0,255)$

This maps the model's output range [-1, 1] to the standard image range [0, 255].

Related Pages

Implemented By

• Implementation:Deepseek_ai_Janus_Image_Post_Processing_AR