DeepSeek / Image generation
Can DeepSeek generate images? Chat, Janus-Pro and the API
Yes, but the route matters. DeepSeek publishes the Janus-Pro image-generation model and a separate demonstration on Hugging Face. In our September 23, 2026 test, that demo returned four actual images. On the same date, one request in DeepSeek’s main web chat, with Search and DeepThink visibly selected, returned a written image prompt rather than an image. DeepSeek’s current Vision API documentation explains how to send an image in for analysis; it does not document image generation through that endpoint.
ChinaAI is independent of DeepSeek. The Janus-Pro Space is published by DeepSeek’s deepseek-ai account on Hugging Face; it is separate from DeepSeek Chat. This page does not generate images on ChinaAI.
By China AI Chat Team · Official sources reviewed and two first-use tests run September 23, 2026 · How we test
Which “DeepSeek image generator” do you mean?
“DeepSeek” can refer to the main assistant, a hosted API or downloadable models. A claim about one does not establish a feature in the others. Choose the route that matches the job you need to finish:
| Your job | Start here | What to expect |
|---|---|---|
| Create a picture from words using DeepSeek’s work | Janus-Pro-7B demo, published by deepseek-ai on Hugging Face | The separate Text-to-Image Generation tab produced four pictures from one prompt in our test. The public demo’s availability and queue can change. |
| Ask the main DeepSeek Chat to draw a picture | DeepSeek web chat | Our one first answer supplied a reusable text prompt, but no image file. Check your own interface; we did not test every account or release. |
| Have an API analyze a picture | DeepSeek Vision API guide | The documented deepseek-flash workflow sends an image as input and reads a text answer. It is a different task from generating a new picture. |
| Create an image in a full web assistant | Our tested Qwen Create Image guide | Qwen’s assistant offered an image mode, aspect-ratio options and a download control during our separate test. It is a Qwen product, not a DeepSeek feature. |
How to generate an image with the Janus-Pro demo
The DeepSeek Janus repository supplies text-to-image inference code and links to its own online Janus-Pro-7B Space. The Space was reachable in our browser on September 23, 2026. It is a research-model demo hosted on Hugging Face, with its own runtime and settings; it is not a switch inside the main DeepSeek Chat.
- Open the Space and choose Text-to-Image Generation. Do not stay on its separate Multimodal Understanding tab; that tab is for questions about an image.
- Write the subject and constraints in Prompt. For example, specify the number of objects, important colors, background and any text that must be absent.
- Press Generate Images once. If the Space queues or shows an error, read its status or error message before trying again. Do not assume that the main chat or Vision API will provide the same output.
- Check the results against your brief. Compare object count, colors, lettering and composition. Open a thumbnail to see the image more closely; small previews can hide unwanted details.
- Download the image you want to keep. Open its preview and use the Download arrow. Then open the saved file and check its size and format. The saving steps below show the control we used.
Our test left the visible advanced controls at their defaults: CFG Weight 5, temperature 1 and optional seed 1234. Those are interface observations, not recommended settings for every subject. The model card identifies Janus-Pro as a separate multimodal understanding and generation model. If you plan to run its weights yourself, use the repository instructions and read its model license. The repository code is MIT-licensed; the model weights are subject to DeepSeek’s separate model license. We did not install or benchmark the weights here.
Save the image to your device
Once the gallery appears, choose the picture you want to keep. Downloading saves that selected picture; you do not need to generate another set.
- Open a thumbnail in Generated Images. This opens the larger preview.
- Choose Download. In the desktop interface we checked, the downward-arrow button sits at the top right of the preview, beside the full-screen control.
- Open the file from your browser’s downloads. Check that it opens as an image, then check its pixel dimensions and file type in your image viewer or file properties. Rename it so you can identify it later.
If no file appears, check the browser’s download list for a pending or blocked download and keep the gallery open while you investigate. Use the image preview’s Download control again if needed; pressing Generate Images creates another set rather than saving the one already on screen.
What we saved: in that check, one new request using the mug prompt reproduced below and visible defaults returned four images. We opened the first thumbnail by position and saved it as a 768 × 768 WebP file (5,984 bytes), then opened and inspected the file. The demo’s published code includes an upscaling step; these are the delivered file’s dimensions, not the model’s native generation resolution.
WebP may not be accepted by every editing or publishing tool. If yours requires PNG or JPEG, keep the original and export a separate copy in the required format. Changing the format or enlarging the image does not restore missing detail.
This check verifies saving one file in Chrome. It does not establish the dimensions of the older September 23 images or general image quality. Download the saved WebP, screenshot and check record (ZIP).
What happened in our first-answer tests?
We fixed each prompt and its success checks before submitting it. Both tests used public product interfaces on September 23, 2026, with no API key or paid request. The prompts were worded for their respective interfaces, but asked for the same single mug, cream tabletop, light blue wall and no extra objects or lettering. We submitted each once and kept the first result. The tests do not measure reliability, image quality across prompts or access in every region and plan.
Main DeepSeek Chat: a written prompt, not an image
With Search and DeepThink selected, our first request at chat.deepseek.com returned a written image prompt and suggestions for other tools. We had asked for an actual image in the chat; no viewable generated image appeared. The interface did not expose the serving model or account plan. This is one session’s result, not a claim about every account, mode or future release.
Read the Chat request and test conditions
Recorded chat conditions: we used a fresh signed-in conversation with no attached file. The saved pre-send screen shows Search and DeepThink highlighted, and a later interface check reported both selected. We had intended to leave them off, but did not verify their persisted state before sending. The answer displayed “Thought for 2 seconds.” This was not a confirmed plain-chat test. The first answer said, “I can’t generate or return actual image files in this chat,” then supplied a written prompt and suggestions for other tools.
Create an original square image showing exactly one red ceramic mug on a plain cream tabletop against a light blue wall. No text, people or extra objects. Please return the actual image in this chat, not a written description or a prompt for another tool.
DeepSeek’s Janus-Pro Space: four actual images
In the Space published by deepseek-ai, we selected Text-to-Image Generation and submitted one brief. Its first run displayed four image thumbnails. Each visibly contained one mug on a pale tabletop against a light blue background, without text or people. The mugs ranged from red to orange-brown, and the tabletop framing varied; treat the brief as a starting point, not an exact-rendering guarantee. In that September 23 test, we did not regenerate, choose a preferred seed, download the image files or inspect their pixel dimensions. The separate September 27 saving check above does not change that original record.
Read the exact Janus-Pro prompt
A square image showing exactly one red ceramic mug on a plain cream tabletop against a light blue wall. No text, people or extra objects.
The fixed criteria, interface captures and first-answer text are in our downloadable test records (ZIP). Screenshots show the outputs actually observed; the archive does not contain private chats, credentials or an image-quality score.
Why image input is not an image generator
DeepSeek’s current Vision API documentation shows deepseek-flash receiving an image alongside a question such as “What is in this image?” and printing the response’s text content. It documents image input through a data URL, an external URL or a Files API reference. This supports tasks such as describing a picture, reading a screenshot or analyzing a chart. It does not document a text-to-image output method for that API endpoint. This is a documentation boundary, not a claim that DeepSeek has no image-generation research models.
Some websites call themselves a “DeepSeek image generator.” Their branding alone does not establish that they are operated by DeepSeek or use the Janus-Pro checkpoint they name. If a site asks you to upload a private image, sign in or pay, check its operator, model disclosure, terms and privacy policy first. For a first-party research route, start from the DeepSeek Janus repository or its linked Space rather than a similarly named site.
For a tested image-input request, two first responses and token usage, see our DeepSeek Vision API guide. For a broader map of releases, hosted API names and license boundaries, see our DeepSeek models guide. For consumer-chat access and other tasks, use the DeepSeek overview.
Common questions
Is Janus-Pro inside DeepSeek Chat?
The official Janus repository presents it as a separate model with its own code and demo. Our consumer-chat interface did not offer a Janus-Pro image mode, and our one image request did not return a picture. A downloadable model release does not by itself establish a feature in the main chat app.
Can DeepSeek describe or read an image?
DeepSeek’s Flash Vision API guide documents image input for text-based analysis. That is distinct from the Janus-Pro text-to-image demo. Features in the consumer chat can differ from API documentation; check the exact interface you plan to use.
What should I use if I need an image now?
Try the DeepSeek-published Janus-Pro Space if you specifically want to explore Janus-Pro and the demo is available. If you need an image-making workflow in a full assistant, our Qwen Create Image test walks through its official mode and shows the first outputs and limits. Check any generated file before using it.
We will revisit this guide if DeepSeek adds an image-generation control to its consumer chat, documents image output in its API, or changes the Janus-Pro demo. Send a correction with the current source and date you checked.