Can DeepSeek Generate Images? Janus-Pro Guide & Example

Documentation and interface checked September 4, 2026. We reviewed the official model documentation and opened the demo’s generation interface. The illustrated result below comes from the researchers’ published paper, not a new Seek-Chat.com generation test. We did not measure generation speed, reliability or comparative quality.

Choose the correct DeepSeek route

Route	Input and output	What to use it for
Hosted V4 Flash / Pro	Text in, text out.	Writing a description or image prompt; not generating an image file.
Hosted V4 Flash Vision Exp	Supported images and text in; text answers out.	Describing or extracting information from an existing image; not returning a new or edited image.
Janus-Pro, through its model code or a supporting demo	Text-to-image generation, plus a separate image-understanding workflow.	Generating an image from a written prompt. Confirm which model and settings the interface actually uses.

Sources: the current hosted model matrix, the Vision API guide, and the official Janus-Pro model card. These are separate products, not a hidden image-generation switch in the hosted text API.

Try the official Janus-Pro demo

The DeepSeek-AI Janus repository links directly to its Janus-Pro-7B Hugging Face Space. Use that link rather than choosing an unrelated site solely because its name includes DeepSeek.

1. Open the Space and wait for its app to load. Choose Text-to-Image Generation, not Multimodal Understanding.
2. Enter your description in Prompt. For a simple starting point, use the coffee prompt below.
3. Leave the controls at their displayed defaults initially. If you change CFG Weight, temperature or Seed (Optional), record the values.
4. Select Generate Images and inspect the Generated Images gallery. Save the actual result together with the prompt, settings, date and model label.

The labels and workflow above are visible in the demo and its published application source. Access, queues and GPU quotas can change. If it asks you to sign in, upgrade or wait, those are hosting conditions—not proof of a model failure. We have not verified a free-use quota or a successful generation in this review.

A published prompt and its actual result

The authors used this short prompt in Figure 2 of their Janus-Pro paper:

A steaming cup of coffee on a wooden table.

In the figure below, the coffee example is the middle pair on the top row; the right-hand image in that pair is labeled Janus-Pro-7B. It shows a cup and saucer on a wooden surface. This is a selected publisher example, not a random sample or a guarantee of what your run will produce.

The paper describes these image outputs as 384 × 384 pixels. A demo may display larger files: the reviewed Space source includes RealESRGAN upscaling after generation. Upscaled dimensions are not the model’s native generation resolution. Reusing the prompt does not reproduce the paper’s image by itself; the prompt alone is not a complete record of its original run.

Developer alternative: run the repository demo

The official Gradio-demo instructions give the local entry point demo/app_januspro.py. Obtain the repository and use a dedicated environment with compatible Python/PyTorch dependencies, a supported GPU and sufficient device memory. We did not test this installation.

Check sharing before you run it: the reviewed application file ends with demo.launch(share=True). Gradio documents that this creates a public sharing link. For access from the same device only, replace that launch line before starting the app with:

demo.launch(share=False, server_name="127.0.0.1")

Then, from the repository directory, install the demo dependencies and start the app:

pip install -e ".[gradio]"
python demo/app_januspro.py

This disables Gradio’s public sharing link and binds its interface to the loopback address. It does not stop model downloads or every network request made by dependencies. “Run locally” is not, by itself, a guarantee that a demo is private or disconnected; review the current code and environment before using sensitive material.

Check the result before using it

• Prompt fidelity: check the subject, object count, spatial relationships, materials and requested exclusions. A plausible picture may still miss the instruction.
• Fine details: inspect faces, hands, small objects and any written text at full size. Do not infer reliable typography from one successful publisher example.
• Editing needs: do not assume the demo supports inpainting, object removal or consistent characters. Confirm the exact workflow if these are requirements.
• Provenance: label synthetic visuals where context requires it. Do not present a generated image as a photograph documenting a real event.
• Comparison: this guide does not establish that Janus-Pro is better than another image model. A useful comparison needs the same prompts, settings, selection rules and evaluation criteria.

Costs, privacy and licensing

Downloading model files does not make hardware, hosted inference or every third-party interface free. Check the host’s quota and pricing. The model card’s license section distinguishes the code license from the linked model terms. Review the actual license files and the host’s terms for your intended use; the CC BY license on the paper’s figure is not a license for every model output.

Do not submit private photographs, personal documents, credentials or confidential prompts to a public demo without the necessary permission and a review of its operator and data practices. An official repository linking a demo establishes its source; it does not eliminate those responsibilities.

Frequently asked questions