Janus Pro 7B
Janus Pro 7B combines image understanding and image generation in one model. It uses separate visual-encoding paths for those two tasks while retaining a shared transformer.
Research release
Why the image paths are separate
The design separates visual representations for understanding from those used to generate an image. Janus Pro builds on the DeepSeek LLM base family.Janus Pro 7B model card
Practical guidance. Evaluate the two tasks independently. A useful description of an input image does not demonstrate that the model can generate every requested layout, word or object relationship.
Review instruction following in the image
Practical guidance. Use prompts with explicit object counts, relative positions and required text. Check each requirement against the produced image. Keep visual quality and instruction completion as separate review criteria.
Treat the checkpoint as a research release
Janus Pro adds changes to training strategy, data and scale over the original Janus work.Janus project and downloads
Practical guidance. Use its published setup for reproducible experiments. If you need an operated service, evaluate hosting and reliability separately from the checkpoint’s research results.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Launch date | |
| Access | Downloadable |
| Model identifiers | deepseek-ai/Janus-Pro-7B |
| Context window | 4,096 tokens Published sequence length. Janus project and downloads |
| Visual encoder | SigLIP-L for understanding, with 384 × 384 image input.Janus Pro 7B model card |
| License distinction | MIT code; the model card points to the separate DeepSeek Model License for weights.Janus Pro 7B 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