Thanks very much, yes, I always enjoy the installation/testing of new models like this. I was happy to use my desktop for something, as I had not used it for anything intensive lately!
Good to see you're covering these two. Love the content an your approach is gonna be awesome for people who are still getting started with AI. Thanks for showcasing the models and helping people use them! We definitely need as much AI literacy as possible
Thanks for the kind words. I try to make things as understandable as possible even if I'm not doing a full tutorial. Agreed, AI literacy is important regardless of how one feels about the technology!
We can only hope they unleash the R2 (multi-modal supreme) version with all these modalities built-in to one. 2025 is shaping up to be even more insane than predicted.
No mistakes in logs, but i got endless ui symbol calculation loop, should i install cuda drivers? just have fresh install of ubuntu Ubuntu 24.04.1 LTS. My CPU intel 10700kf, and 3080ti rtx videocard, 64GB ram
@@Bijanbowen how do non tech people get this? We need a video for dummies who want this but don’t know how to code, step by step like you are explaining to a child
Hi, i have an error when genereting picture. In logs it sounds like: RuntimeError: Input type (struct c10::BFloat16) and bias type (struct c10::Half) should be the same I have 4060 ti 16gb vram and 128gm ram. Vram didn't utilize at all, only Ram. Do i need to change somthing in app_januspro py file?
When using the official website, all images are 1024x1024. Can the local version create full HD images or arbitrary sizes and ratios (e.g. 16:9)? Does the local version run with a CPU when no extra GPU is available (GPU built in CPU)?
It seems like when generating it takes as much vram as available. I did the same test and the 7b model used about 40gb on my 48 gb card that was doing nothing else.
Help, I ran the 7B, but when I click 'Chat' or 'Generate Images,' it just loads infinitely. I have a 4060 Ti and 16 GB of RAM. By the way, thanks for the video!
legit questions are those 1B and 7B versions real deepseek? or an algamation hibrid of llama finetuned to output deepseek like responses? do the 1B and 7B "deepseek models" share if infrastructure or are they fake?
I have not yet tried that, aside from the little distilled 1.5B I tried a few days ago (which definitely wouldn't do that). I would imagine it would be able to without much issue, but I will try to do some more Deepseek content and do a game test in one shot as you describe here.
Hahaha I'm becoming more open to it.. I actually initially tried to do this on the super for the video, but unfortunately this is far, far too resource intensive to run on the super. It is perhaps possible to try some very hardcore optimizations just to get it to run, but I still don't even know if it would then.
Thanks! Unfortunately this is very unlikely to be able to run on the pi 16gb. It is a great device but I wouldn't want to run more than some general purpose Ollama llms on it.
Thank you for showing us! Is there a version for computer illiterate people like myself? I honestly have no idea how you did 99% of the coding lol i couldn't start if someone held a gun to my head. Is there a version like deepseek, where you go in, type what you need and it generates the images or do you have to do all the steps you just showed?
Not a problem! In regards to a simpler to use version - sort of. You can follow this link to get to the huggingface spaces for this, which is basically a cloud hosted instance of what we see running in the video here: huggingface.co/spaces/deepseek-ai/Janus-Pro-7B
Anyone who has a link to a video who teach beginners exactly the step by step, from the link, what to do etc. this video starts typing code to a cmd? i have no frcking idea. sorry not a tech savy person
what's the point in doing a test when you don't even read the prompt that you're testing. sigh. Testing a prompt means testing prompt coherence. PS use your own.
For 99.9% of users, their data isn’t individually significant, and framing everything as 'us vs. them' is just narrow-minded fearmongering. while security concerns can be valid, blindly dismissing tech based on its origin ignores the bigger picture
legit useless to people who wanna use the thing yet know 0 bout coding, u r the 3rd person that did a vid on how to install it and all of u were useless
Why promote a model trained on a foreign language? You get foreign language constraints, mentality, and potentially controlled narrative from source selection? I just don't get it. Even free does not make sense tome. Upsetting the market, I love that part!!
Regardless of the origin, it is a popular model and is worth looking into and seeing for a lot of folks interested in AI. I understand your concerns and I do, albeit in a roundabout way, touch upon them in the closing thoughts of the video.
Open source vs money hungry, of cource we choose freedom of choosing "open". And i hope another deepseek with "open minded" will flourish, for freedom, for humanity. ❤
imagine the entire thing reversed... mightn't you think, as a natural chinese speaker, that the american models are constraining their language, mentality, and narrative differently from how you would prefer?
Interesting, I can tell you are having a blast exploring this new model
Thanks very much, yes, I always enjoy the installation/testing of new models like this. I was happy to use my desktop for something, as I had not used it for anything intensive lately!
Riding the deepseek hype 😊 it worked, you pulled me in.
LOL I am glad!
Good to see you're covering these two. Love the content an your approach is gonna be awesome for people who are still getting started with AI. Thanks for showcasing the models and helping people use them! We definitely need as much AI literacy as possible
Thanks for the kind words. I try to make things as understandable as possible even if I'm not doing a full tutorial. Agreed, AI literacy is important regardless of how one feels about the technology!
Great video! Thanks so much for putting this together! This was right to point - what I was looking for!! 🙂
Great video, easy to understand and follow 👍
I really love your videos. Clear and to the point. Couldnt get it to work on my mac because of sentencepiece errors.
Thanks for the kind words! That is frustrating about the install error. Unfortunately I can't offer any suggestions RE that issue.
perfect timing! subscribed
Thanks very much, welcome!
We can only hope they unleash the R2 (multi-modal supreme) version with all these modalities built-in to one. 2025 is shaping up to be even more insane than predicted.
What a freakin fantastic video! Thanks for putting this together!
Thanks very much for the kind words!
Nicely done mate - thanks
Love your video's, easy to follow for beginners.
Love your content!
Thanks for the kind words!
Thank you for another great program and informations
Absolutely, thanks for the kind words!
I'd like to see if you could run the 4bit Quantized version on the Jetson Orin Nano.
So I still wonder, would the 1B model fit on 8G VRAM?
No mistakes in logs, but i got endless ui symbol calculation loop, should i install cuda drivers? just have fresh install of ubuntu Ubuntu 24.04.1 LTS. My CPU intel 10700kf, and 3080ti rtx videocard, 64GB ram
just in case, installing cuda drivers as a driver from nvidia helped
great video, however, when you changed eye to car, you did not change all instances of eye to car so the output was not as expected
can you generate uncensored images with deepseek janus pro?
I honestly did not try, I would assume no, but you never do know...
@@Bijanbowen how do non tech people get this? We need a video for dummies who want this but don’t know how to code, step by step like you are explaining to a child
thanks 4 sharing the channel exp! good luck! 🫶
I appreciate it a lot!
Hi, i have an error when genereting picture. In logs it sounds like: RuntimeError: Input type (struct c10::BFloat16) and bias type (struct c10::Half) should be the same
I have 4060 ti 16gb vram and 128gm ram. Vram didn't utilize at all, only Ram. Do i need to change somthing in app_januspro py file?
When using the official website, all images are 1024x1024. Can the local version create full HD images or arbitrary sizes and ratios (e.g. 16:9)?
Does the local version run with a CPU when no extra GPU is available (GPU built in CPU)?
For someone new to AI, everyone is talking about that deepseek does not need NVIDIA cards anymore, is there an way this runs only on a CPU?
so janus pro is a full-software to be install? not like bing image creator that u go to its website and input prompt?
It seems like when generating it takes as much vram as available. I did the same test and the 7b model used about 40gb on my 48 gb card that was doing nothing else.
What is the program you are running everything on?
Help, I ran the 7B, but when I click 'Chat' or 'Generate Images,' it just loads infinitely. I have a 4060 Ti and 16 GB of RAM.
By the way, thanks for the video!
How to utilize GPU for this model?
Your the man!
Thx for video
For sure!
legit questions are those 1B and 7B versions real deepseek? or an algamation hibrid of llama finetuned to output deepseek like responses? do the 1B and 7B "deepseek models" share if infrastructure or are they fake?
I got it running, though Janus-Pro-1B did not fit in 4GB VRAM on my side, don't know why. But it runs on CPU, due to your explanations, thanks.
What was the issue with not working on the Jetson Orin Nano Super?
Far too performance intensive to run on the Jetson I'm afraid.
nvidia-smi supongo. Buen test bro
Have you tried the text based only? To see whether it is able to generate a basic game such snake or whatever, on first try
It can
I have not yet tried that, aside from the little distilled 1.5B I tried a few days ago (which definitely wouldn't do that). I would imagine it would be able to without much issue, but I will try to do some more Deepseek content and do a game test in one shot as you describe here.
Let's change the Eye for a Sports car.... And you get an Eye.
I know you don't want to be the Orin guy, but would this run on the Orin Nano Super?
Hahaha I'm becoming more open to it.. I actually initially tried to do this on the super for the video, but unfortunately this is far, far too resource intensive to run on the super. It is perhaps possible to try some very hardcore optimizations just to get it to run, but I still don't even know if it would then.
Ignore me I just got to the end of your video :D
@ Hahahah
It is possible on Win11?
Thanks from Ukr ))
thanks friend. but ... would you make a test with raspberry 16gb? minimal hardware performances are what the most of people look for. thanks
What graphics card do you have connected to it? I'm not so sure there would be enough compute power in that even for the 1B version.
Thanks! Unfortunately this is very unlikely to be able to run on the pi 16gb. It is a great device but I wouldn't want to run more than some general purpose Ollama llms on it.
@@Bijanbowen I found another video showing one token per second. Not so bad in the end for an sbc
Thank you for showing us! Is there a version for computer illiterate people like myself? I honestly have no idea how you did 99% of the coding lol i couldn't start if someone held a gun to my head. Is there a version like deepseek, where you go in, type what you need and it generates the images or do you have to do all the steps you just showed?
Not a problem! In regards to a simpler to use version - sort of. You can follow this link to get to the huggingface spaces for this, which is basically a cloud hosted instance of what we see running in the video here: huggingface.co/spaces/deepseek-ai/Janus-Pro-7B
Jetson Orin Nano’s RAM keeps getting full, the board crashes
increible, me suscribo
Thanks very much, welcome!
Listen to King Buffalo!
Всё, художники теперь тоже безработные 😢
Anyone who has a link to a video who teach beginners exactly the step by step, from the link, what to do etc. this video starts typing code to a cmd? i have no frcking idea. sorry not a tech savy person
great
Thanks very much!
what's the point in doing a test when you don't even read the prompt that you're testing. sigh. Testing a prompt means testing prompt coherence. PS use your own.
i try to run its on my pc, but its dont use my 3060 videocard, i think
Based on what I saw with vram utilization, I am not sure it would run on the 3060 either.
@@Bijanbowen
I am running the 7b on my 3060ti without issues.
I have only 6 GB vRAM
For 99.9% of users, their data isn’t individually significant, and framing everything as 'us vs. them' is just narrow-minded fearmongering. while security concerns can be valid, blindly dismissing tech based on its origin ignores the bigger picture
Cannot talk and type - very funny!
I just end up starting to type the word I was saying instead LOL
maybe the R1 is very good, but this image model seems pretty cheap to me. Very bad results comparing to other models like flux.
this is poor. but there R1 is really nice
Not great, but paper is progress, I guess.
Agreed, my prompting also didn't do it any favors hahah.
too bad it uses so much vram
legit useless to people who wanna use the thing yet know 0 bout coding, u r the 3rd person that did a vid on how to install it and all of u were useless
Stop following tech channels then. 🤣 Use DeepSeek to explain to your monkey brain how to use it step by step.
Try writing proper English. Use question marks. Punctuation😊 is key.
Really....
Assuming you mean for my bad prompting, yeah that would probably help haha...
Rorri Maesu says useaMirroR
Why promote a model trained on a foreign language? You get foreign language constraints, mentality, and potentially controlled narrative from source selection? I just don't get it. Even free does not make sense tome. Upsetting the market, I love that part!!
Regardless of the origin, it is a popular model and is worth looking into and seeing for a lot of folks interested in AI. I understand your concerns and I do, albeit in a roundabout way, touch upon them in the closing thoughts of the video.
Are you getting mad that the free Chinese ai model is so good that it made US stock market loose 1 trillion dollars?
Open source vs money hungry, of cource we choose freedom of choosing "open". And i hope another deepseek with "open minded" will flourish, for freedom, for humanity. ❤
It speaks English bro 😂
imagine the entire thing reversed... mightn't you think, as a natural chinese speaker, that the american models are constraining their language, mentality, and narrative differently from how you would prefer?