Janus 1.3B
Janus 1.3B is the original released Janus checkpoint. It separates visual encoding for understanding and generation within a unified transformer design.
Research release
An important correction in the release files
DeepSeek corrected tokenizer_config.json on 20 October 2024. A missing pad token had caused poor visual generation. The project notes that the issue affected classifier-free guidance.Janus 1.3B model cardJanus project and downloads
Practical guidance. If reproducing an early result, check the configuration revision as well as the weight files. A model comparison based on the earlier broken configuration can be misleading.
The visual input has a defined size
The model uses SigLIP-L with 384 × 384 image input. Its base language model was trained on approximately 500 billion text tokens.Janus 1.3B model card
Practical guidance. For small text or dense diagrams, inspect what remains visible after preprocessing. The source image’s resolution is not necessarily the resolution seen by the model.
One transformer receives two visual representations
The original research separates the visual encoders because recognition and image generation need different levels of image detail. Each path can use its own encoding method while the transformer remains shared.Original Janus research publication
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Launch date | October 2024 |
| Access | Downloadable |
| Model identifiers | deepseek-ai/Janus-1.3B |
| Context window | 4,096 tokens Published sequence length. Janus project and downloads |
| Base model | DeepSeek LLM 1.3B Base.Janus 1.3B model card |
| Model terms | DeepSeek Model License; MIT applies to the code.Janus 1.3B model card |
Related models
- Janus Pro 1B
DeepSeek · Janus Pro
Compact Janus Pro model that pairs a SigLIP image encoder with a separate tokenizer for image generation.
Launch date Research releaseDownloadable - JanusFlow 1.3B
DeepSeek · JanusFlow
Multimodal research model that combines autoregressive text generation with rectified flow to create images at 384 by 384 pixels.
Launch date Research releaseDownloadable