DeepSeek releases Janus-Pro for image understanding and generation

Background

Janus-Pro belongs to DeepSeek's Janus line of models. The release describes it as an advanced version of the original Janus, so its design builds on the ideas that earlier model introduced.

What happened

On 27 January 2025 DeepSeek released Janus-Pro as an advanced version of its Janus model. ^[1]

Janus-Pro is one model for both multimodal understanding, where an image goes in and text comes out, and text-to-image generation. ^[2]

A multimodal model works with more than one kind of input or output. Here the inputs and outputs are images and text. A model that answers questions about a photo and one that draws a picture from a sentence are both multimodal, and Janus-Pro does both jobs.

Key details

Janus separates visual encoding into separate pathways for understanding and generation while keeping one unified Transformer. ^[1]

There are two sizes, Janus-Pro-1B and Janus-Pro-7B, each with a 4,096-token sequence length. ^[1]

Compared with Janus, the model has an optimized training strategy, expanded training data and a larger model size. These changes give better understanding, better text-to-image instruction-following and more stable generation. ^[2]

The code is released under the MIT License, and model use is governed by the DeepSeek Model License, which permits commercial use. ^[1]

Impact

For readers, the notable part of this release is the design: one model that reads images and also writes them. The code is under the MIT License, and model use falls under a license that permits commercial use. Specific claims about quality are best checked against the technical report's own figures.