DeepSeek isn't done yet with OpenAI – image-maker Janus Pro is gunning for DALL-E 3
(2025/01/28)
- Reference: 1738020970
- News link: https://www.theregister.co.uk/2025/01/27/deepseek_image_openai/
- Source link:
Barely a week after DeepSeek's R1 LLM turned Silicon Valley on its head, the Chinese outfit is back with a new release it claims is ready to challenge OpenAI's DALL-E 3.
Released on [1]Hugging Face on Monday amid an ongoing [2]cyberattack , Janus Pro 1B and 7B are a family of multimodal large language models (LLMs) designed to handle both image generation and vision processing tasks. As with DALL-E 3, you give Janus Pro an input prompt and it generates a matching image.
The models are said to improve upon the Chinese lab's first 1.3B Janus model [3]released last year. They achieve this by decoupling visual encoding into a separate pathway while maintaining a single transformer architecture for processing.
[4]
In a research [5]paper [PDF] detailing the model and its architecture, the boffins behind the neural network noted that the original Janus model showed promise, but suffered from "suboptimal performance on short prompts, image generation, and unstable text-to-image generation quality." With Janus Pro, DeepSeek says it was able to overcome many of these limitations by using a large dataset and targeting higher parameter counts.
[6]
[7]
Pitted against a variety of multimodal and task-optimized models, the startup claims Janus Pro 7B narrowly outperforms both Stable Diffusion 3 Medium and OpenAI's DALL-E 3 in the GenEval and DPG-Bench benchmarks. However, it's worth noting that image analysis tasks are limited to 384x384 pixels.
[8]
DeepSeek claims its Janus Pro image models offer higher performance than either OpenAI's DALL-E 3 or Stability AI's SD3-Medium ... Click to enlarge
Much like DeepSeek V3, the model maker claims it was able to achieve these results using only a few hundred GPUs running the HAI-LLM framework on PyTorch. The process, detailed in the paper above, claims the "whole training process took about 7/14 days on a cluster 16/32 nodes for 1.5B/7B model, each equipped with eight Nvidia A100 (40GB) GPUs."
Training times may have been aided by the reuse of earlier models rather than training an entirely new one from scratch. We've reached out to DeepSeek for clarification.
But while competitive with other multimodal LLMs and diffusion models, DeepSeek admits there's still more work to be done. "In terms of multimodal understanding, the input resolution is limited to 384x384, which affects its performance in fine-grained tasks, such as OCR," the researchers explained. Meanwhile, for image generation, they note the limited resolution also results in images that lack fine details.
[9]
The Janus codebase is available under an MIT license, with the use of the Pro models subject to DeekSeek's Model License, which you can find [10]here .
DeepSeek latest at 2315 UTC
[11]China's DeepSeek releases free challenger to OpenAI's o1
[12]How did DeepSeek train its AI model on a lot less – and crippled – hardware?
[13]DeepSeek's R1 tells Register reader: 'My guidelines are set by OpenAI'
[14]Tech stocks tank as US AI dominance no longer a sure bet
[15]DeepSeek limits new accounts amid cyberattack
Meanwhile, Nvidia, whose GPUs were used to train China-based DeepSeek's models, told El Reg the LLMs were "an excellent AI advancement," and used technologies that are "fully export control compliant."
If you're interested in giving either of the Janus Pro models a go, DeekSeek has a pair of quick-start scripts available on their [16]GitHub page for local testing or you can check out their demo running in Hugging Face Spaces [17]here . Note: it took several minutes for the HuggingFace demo to load during our testing.
DeepSeek's model releases caused significant market reactions, sending Silicon Valley stocks [18]sliding precipitously on Monday as US superiority in AI and the need for billions of dollars of infrastructure was called into question. However, it hasn't been without a few hiccups including challenges with censorship.
If that weren't enough, DeepSeek was [19]forced to limit new signups for its AI chatbot on Monday amid an ongoing cyberattack. ®
Get our [20]Tech Resources
[1] https://huggingface.co/deepseek-ai/Janus-Pro-7B
[2] https://status.deepseek.com/incidents/666k4t024szr
[3] https://huggingface.co/deepseek-ai/Janus-1.3B
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2Z5hkfFpb01qdnHHrD3NBLQAAAc0&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[5] https://github.com/deepseek-ai/Janus/blob/main/janus_pro_tech_report.pdf
[6] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Z5hkfFpb01qdnHHrD3NBLQAAAc0&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Z5hkfFpb01qdnHHrD3NBLQAAAc0&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[8] https://regmedia.co.uk/2025/01/27/ds_janus_pro_bench.jpg
[9] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Z5hkfFpb01qdnHHrD3NBLQAAAc0&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[10] https://github.com/deepseek-ai/DeepSeek-LLM/blob/HEAD/LICENSE-MODEL
[11] https://www.theregister.com/2025/01/26/deepseek_r1_ai_cot/
[12] https://www.nextplatform.com/2025/01/27/how-did-deepseek-train-its-ai-model-on-a-lot-less-and-crippled-hardware/
[13] https://www.theregister.com/2025/01/27/deepseek_r1_identity/
[14] https://www.theregister.com/2025/01/27/tech_stocks_tank_as_us/
[15] https://www.theregister.com/2025/01/27/deepseek_suspends_new_registrations_amid/
[16] https://github.com/deepseek-ai/Janus?tab=readme-ov-file
[17] https://huggingface.co/spaces/deepseek-ai/Janus-Pro-7B
[18] https://www.theregister.com/2025/01/27/tech_stocks_tank_as_us/
[19] https://www.theregister.com/2025/01/27/deepseek_suspends_new_registrations_amid/
[20] https://whitepapers.theregister.com/
Released on [1]Hugging Face on Monday amid an ongoing [2]cyberattack , Janus Pro 1B and 7B are a family of multimodal large language models (LLMs) designed to handle both image generation and vision processing tasks. As with DALL-E 3, you give Janus Pro an input prompt and it generates a matching image.
The models are said to improve upon the Chinese lab's first 1.3B Janus model [3]released last year. They achieve this by decoupling visual encoding into a separate pathway while maintaining a single transformer architecture for processing.
[4]
In a research [5]paper [PDF] detailing the model and its architecture, the boffins behind the neural network noted that the original Janus model showed promise, but suffered from "suboptimal performance on short prompts, image generation, and unstable text-to-image generation quality." With Janus Pro, DeepSeek says it was able to overcome many of these limitations by using a large dataset and targeting higher parameter counts.
[6]
[7]
Pitted against a variety of multimodal and task-optimized models, the startup claims Janus Pro 7B narrowly outperforms both Stable Diffusion 3 Medium and OpenAI's DALL-E 3 in the GenEval and DPG-Bench benchmarks. However, it's worth noting that image analysis tasks are limited to 384x384 pixels.
[8]
DeepSeek claims its Janus Pro image models offer higher performance than either OpenAI's DALL-E 3 or Stability AI's SD3-Medium ... Click to enlarge
Much like DeepSeek V3, the model maker claims it was able to achieve these results using only a few hundred GPUs running the HAI-LLM framework on PyTorch. The process, detailed in the paper above, claims the "whole training process took about 7/14 days on a cluster 16/32 nodes for 1.5B/7B model, each equipped with eight Nvidia A100 (40GB) GPUs."
Training times may have been aided by the reuse of earlier models rather than training an entirely new one from scratch. We've reached out to DeepSeek for clarification.
But while competitive with other multimodal LLMs and diffusion models, DeepSeek admits there's still more work to be done. "In terms of multimodal understanding, the input resolution is limited to 384x384, which affects its performance in fine-grained tasks, such as OCR," the researchers explained. Meanwhile, for image generation, they note the limited resolution also results in images that lack fine details.
[9]
The Janus codebase is available under an MIT license, with the use of the Pro models subject to DeekSeek's Model License, which you can find [10]here .
DeepSeek latest at 2315 UTC
[11]China's DeepSeek releases free challenger to OpenAI's o1
[12]How did DeepSeek train its AI model on a lot less – and crippled – hardware?
[13]DeepSeek's R1 tells Register reader: 'My guidelines are set by OpenAI'
[14]Tech stocks tank as US AI dominance no longer a sure bet
[15]DeepSeek limits new accounts amid cyberattack
Meanwhile, Nvidia, whose GPUs were used to train China-based DeepSeek's models, told El Reg the LLMs were "an excellent AI advancement," and used technologies that are "fully export control compliant."
If you're interested in giving either of the Janus Pro models a go, DeekSeek has a pair of quick-start scripts available on their [16]GitHub page for local testing or you can check out their demo running in Hugging Face Spaces [17]here . Note: it took several minutes for the HuggingFace demo to load during our testing.
DeepSeek's model releases caused significant market reactions, sending Silicon Valley stocks [18]sliding precipitously on Monday as US superiority in AI and the need for billions of dollars of infrastructure was called into question. However, it hasn't been without a few hiccups including challenges with censorship.
If that weren't enough, DeepSeek was [19]forced to limit new signups for its AI chatbot on Monday amid an ongoing cyberattack. ®
Get our [20]Tech Resources
[1] https://huggingface.co/deepseek-ai/Janus-Pro-7B
[2] https://status.deepseek.com/incidents/666k4t024szr
[3] https://huggingface.co/deepseek-ai/Janus-1.3B
[4] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=2&c=2Z5hkfFpb01qdnHHrD3NBLQAAAc0&t=ct%3Dns%26unitnum%3D2%26raptor%3Dcondor%26pos%3Dtop%26test%3D0
[5] https://github.com/deepseek-ai/Janus/blob/main/janus_pro_tech_report.pdf
[6] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Z5hkfFpb01qdnHHrD3NBLQAAAc0&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[7] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=3&c=33Z5hkfFpb01qdnHHrD3NBLQAAAc0&t=ct%3Dns%26unitnum%3D3%26raptor%3Deagle%26pos%3Dmid%26test%3D0
[8] https://regmedia.co.uk/2025/01/27/ds_janus_pro_bench.jpg
[9] https://pubads.g.doubleclick.net/gampad/jump?co=1&iu=/6978/reg_software/aiml&sz=300x50%7C300x100%7C300x250%7C300x251%7C300x252%7C300x600%7C300x601&tile=4&c=44Z5hkfFpb01qdnHHrD3NBLQAAAc0&t=ct%3Dns%26unitnum%3D4%26raptor%3Dfalcon%26pos%3Dmid%26test%3D0
[10] https://github.com/deepseek-ai/DeepSeek-LLM/blob/HEAD/LICENSE-MODEL
[11] https://www.theregister.com/2025/01/26/deepseek_r1_ai_cot/
[12] https://www.nextplatform.com/2025/01/27/how-did-deepseek-train-its-ai-model-on-a-lot-less-and-crippled-hardware/
[13] https://www.theregister.com/2025/01/27/deepseek_r1_identity/
[14] https://www.theregister.com/2025/01/27/tech_stocks_tank_as_us/
[15] https://www.theregister.com/2025/01/27/deepseek_suspends_new_registrations_amid/
[16] https://github.com/deepseek-ai/Janus?tab=readme-ov-file
[17] https://huggingface.co/spaces/deepseek-ai/Janus-Pro-7B
[18] https://www.theregister.com/2025/01/27/tech_stocks_tank_as_us/
[19] https://www.theregister.com/2025/01/27/deepseek_suspends_new_registrations_amid/
[20] https://whitepapers.theregister.com/
beast666
"However, it hasn't been without a few hiccups including challenges with censorship."
The challenge is that it doesn't censor enough for the-powers-that-be in the West.
Scepticism
Sorry that handle is already taken.
Amy Castor and David Gerard have [1]a few interesting things to say about DeepSeek's claims .
Most interesting to me were the suggestions that DeepSeek is using just as much computing power as anyone else, they just can't admit to it because that would imply they've breached export controls on Nvidia GPUs, and the suggestion that OpenAI is pushing the DeepSeek hype as a scare tactic to attract more funding (because they're rapidly running out).
Perhaps a little bit conspiratorial, sure, but at least they list their sources.
[1] https://pivot-to-ai.com/2025/01/26/deepseek-r1-is-wildly-overhyped-but-you-can-try-it-at-home/
Janus, the two-faced God
Howard Sway
Symbolising the transition from one year to the next, hence January, looking backwards to the time just gone and forward to the future. In this case backwards to a time when the money was flowing into certain companies like a river, and forwards to a time when all that money just evaporated into thin air. This is looking more and more like possibly the largest, most meticulously planned trolling operation of all time.
> Crouching tiger, hidden layer(s)
Argh!
Next you'll be telling us that the CEO/Founder's name is "Hugh"
The last thing the world needs is yet another disinformation machine.