JanusFlow 1.3B
JanusFlow combines autoregressive language modeling with rectified flow for image generation. Its published checkpoint is derived from DeepSeek LLM 1.3B Base.
Research release
The generation method is the main distinction
JanusFlow uses rectified flow for images while keeping autoregressive language behavior. Its visual understanding still uses SigLIP-L.JanusFlow 1.3B model card
Practical guidance. This makes it useful for comparing generation methods within a combined multimodal design. Keep the comparison focused on the method rather than treating its release name as a simple quality ranking.
Review images at their native output size
Practical guidance. Evaluate text, fine detail and object placement before resizing the image. Enlarging a low-resolution output can make it easier to display, but does not establish that the model generated additional detail.
Keep the release and sampler together
Practical guidance. Record the checkpoint, inference code and sampling settings used for each image. Changes in those components can affect an experiment even when the written prompt stays the same.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Status | Research release |
| Launch date | |
| Access | Downloadable |
| Model identifiers | deepseek-ai/JanusFlow-1.3B |
| Context window | 4,096 tokens Published sequence length. Janus project and downloads |
| Image generation | Rectified flow with an SDXL variational autoencoder; 384 × 384 output.JanusFlow 1.3B model card |
| Checkpoint | Exponential-moving-average checkpoint after pretraining and supervised fine-tuning.JanusFlow 1.3B model card |
| Model terms | DeepSeek Model License, separate from the MIT code license.JanusFlow 1.3B model card |
Related models
- Janus Pro 7B
DeepSeek · Janus Pro
The larger Janus Pro model combines image understanding and image generation through separate visual encoders and a shared language model.
Launch date Research releaseDownloadable - Janus 1.3B
DeepSeek · Janus
Original Janus model with separate image encoding paths; its early release required a tokenizer correction for classifier-free guidance.
Launch date October 2024Research releaseDownloadable