Testing DeepSeek Janus-Pro for Image Understanding and Generation

要約

30 秒でわかる

In February 2025, Matthew Berman tests DeepSeek Janus-Pro-7B, which supports image understanding and text-to-image generation. He demonstrates a diagram-to-code task, a spreadsheet transcription, meme interpretation and two generated images, with one meme interpretation failing. Vultr partnered on the video and supplied the cloud-compute context; the results are examples rather than a controlled benchmark.

1Unified multimodal modelJanus-Pro-7B can interpret image inputs and generate images. It is not the first research system to combine these modalities.
2Benchmark scopeDeepSeek’s reported GenEval and DPG-Bench results compare specific models and settings, not every measure of visual quality.
3Sponsored cloud demonstrationVultr partnered on the video; Berman runs a Gradio interface on AMD MI300X cloud hardware.
4Vision tasksThe model generates Python from a flowchart and transcribes a spreadsheet, but misreads a workplace meme.
5Image generationBerman produces astronaut and wine-glass examples and considers them good but not exceptional.

ポイント

Janus-Pro combines multimodal understanding and image generation in one model family; earlier research had already explored unified models.
DeepSeek reported higher GenEval scores than specific DALL-E 3 and Stable Diffusion 3 Medium versions, which does not prove superiority on every image task.
Berman’s diagram-to-code and spreadsheet examples appeared successful, while a second meme was misinterpreted.
The demonstration ran on a Vultr-hosted AMD Instinct MI300X, so its apparent speed does not establish local performance on every computer.
The Janus repository and model have separate license terms; review the model license before deployment or commercial use.
The cloud promotion and pricing described in the video were historical and subject to their original terms.

トピック画像と動画 Generative media Multimodal AI Open-weight models AI cost & performance

全文

チャプターごとに整えた全文を読む

元の文字起こしと対照

文字起こしのみ

Janus-Pro and the cloud setup

Berman introduces the model and reviews DeepSeek’s reported image benchmarks. He discloses the Vultr partnership and uses an AMD cloud GPU.

DeepSeek has released Janus-Pro, a multimodal model that can both answer questions about images and generate images. I test the 7-billion-parameter version in February 2025. Its weights are available, but whether it runs conveniently on a particular local computer depends on memory, software and hardware.

スライド 1 / 3DeepSeek’s Janus-Pro benchmark chartThe report compares selected image understanding and generation benchmarks with specific prior models.

DeepSeek reports that Janus-Pro-7B scored 80% on GenEval, compared with 67% for DALL-E 3 and 74% for Stable Diffusion 3 Medium in its reported evaluation. That is a result on one benchmark under the paper’s conditions, not proof that Janus-Pro creates better images in every setting ([technical report](https://arxiv.org/abs/2501.17811)). Its authors also report a DPG-Bench result.

I am excited to see understanding and generation in a single model, but this was not the first research system to attempt both: prior unified systems included Chameleon and Emu3, as discussed in the [Janus paper](https://arxiv.org/abs/2410.13848). The [Janus repository](https://github.com/deepseek-ai/Janus) makes code available, while use of the model is subject to the separate DeepSeek model license; inspect the applicable terms.

Vultr is a partner on this video. I run Janus-Pro in a Gradio interface on a Vultr-hosted AMD Instinct MI300X and point out that the server can also expose an API. This is a cloud demonstration rather than a local-computer speed test, and it does not isolate how much the model or the hardware contributes to performance.

Testing image understanding

The model interprets memes, converts a flowchart to Python and transcribes a spreadsheet. One meme explanation fails.

First I upload a meme depicting competition between DeepSeek and OpenAI and ask for an explanation. The model recognizes the visual joke as a competitive interaction, though parts of its description are imperfect. There is no web search enabled in the setup I show.

Next I ask the model to convert a number-guessing flowchart into runnable Python. I paste and run its generated code, then make several guesses until it reports a win. That is a successful demonstration for this simple example, not proof that generated programs will always match arbitrary diagrams.

I then provide a spreadsheet screenshot and ask for CSV. The output looks correct to me on a visual inspection, but I do not show a cell-by-cell independent audit. On a second meme contrasting startups and large companies, the model misses the joke and describes the large-company side in terms of efficiency and coordination rather than the intended criticism.

Generating images

Berman generates two images in the same interface and discusses his impression. He closes with a historical Vultr promotion.

Without loading a different model in the displayed interface, I switch to image generation and request an astronaut in a jungle with muted colors, then a glass of red wine on a reflective surface. I find the pictures reasonably good rather than extraordinary. These two examples are subjective demonstrations, not a full image-quality evaluation.

スライド 3 / 3Generated wine-glass imageAn example produced by the model in the cloud demonstration.

I conclude that the small model is promising and again promote Vultr’s hosted AMD hardware. I mention a $300 promotional credit for the first 30 days; that offer was part of the 2025 partnership and its current availability and conditions should be checked with the provider.