[HN Gopher] Janus Pro 1B running 100% locally in-browser on WebGPU
       ___________________________________________________________________
        
       Janus Pro 1B running 100% locally in-browser on WebGPU
        
       Author : fofoz
       Score  : 198 points
       Date   : 2025-01-28 14:04 UTC (8 hours ago)
        
 (HTM) web link (old.reddit.com)
 (TXT) w3m dump (old.reddit.com)
        
       | ndr wrote:
       | The image generation results are extremely poor, but it's
       | exciting that it does _anything_ in the browser.
        
         | jjice wrote:
         | I don't know a lot about image generation models, but 1B sounds
         | super low for this kind of model, so I'm pretty impressed,
         | personally.
        
           | diggan wrote:
           | If I remember correctly, SD had less than 1B parameters at
           | launch (~2 years ago?), and you could generate pretty
           | impressive images with the right settings and prompts.
        
             | jjice wrote:
             | Oh wow okay thank you for the context
        
             | salviati wrote:
             | Yep! Less than 1B in total [0]:
             | 
             | > 860M UNet and 123M text encoder
             | 
             | [0] https://github.com/CompVis/stable-
             | diffusion/blob/main/README...
        
             | refulgentis wrote:
             | Janus Pro 1B is a multimodal LLM, not a diffusion model, so
             | it's got a bit more things to pack in the parameters. It is
             | super low parameter count, in an LLM context.
        
         | vunderba wrote:
         | Even the full 7b model's results are relatively low-res
         | (384x384) so its hard for me to imagine the generative aspect
         | of the 1b model would be useable.
         | 
         | Comparisons with other SoTA (Flux, Imagen, etc):
         | 
         | https://imgur.com/a/janus-flux-imagen3-dall-e-3-comparisons-...
        
           | littlestymaar wrote:
           | It's still very impressive that it gets the cube order right!
           | 
           | Also it looks like octopuses are suffering the "six finger
           | hand" syndrome with their arms from all models.
        
           | nicce wrote:
           | I am not sure if the results are that comparable to be
           | honest. For example DALL-E expands the prompt by default to
           | be much more descriptive. We would need to somehow point out
           | that it is close to impossible to produce the same results
           | than DALL-E, for example.
           | 
           | I bet there has been a lot of testing that what looks "by
           | default" much more attractive for the general people. It is
           | also a selling point, when low effort produces something
           | visually amazing.
        
         | qingcharles wrote:
         | I actually had some pretty impressive results (and a few duds).
         | I think we've lost sight of how amazing something like this
         | actually is. I can run this on my low-end GPU in a web browser
         | and it doesn't even tax it, yet it's creating incredible images
         | out of thin air based on a text description I wrote.
         | 
         | Just three years ago this would have been world-changing.
        
       | jedbrooke wrote:
       | well it was a long shot anyway but it doesn't seem to work on
       | mobile. (tried on iOS safari on iPhone 11 pro)
       | 
       | a 1B model should be able to run in the RAM constraints of a
       | phone(?) if this is supported soon this would actually be wild.
       | Local LLMs in the palm of your hands
        
         | nromiun wrote:
         | I don't know about this model but people have been running
         | local models in Android phones for years now. You just need a
         | large amount of ram (8-12 GB), ggml and Termux. I tried it once
         | with a tiny model and it worked really well.
        
           | kittikitti wrote:
           | This is from Reddit, what were you expecting?
        
         | bla3 wrote:
         | This needed a 4 GB renderer process and about that much
         | additional memory use in the GPU process for me, in Chrome.
        
         | rahimnathwani wrote:
         | Local LLMs in the palm of your hands
         | 
         | https://apps.apple.com/us/app/mlc-chat/id6448482937
        
       | amelius wrote:
       | The reason why this doesn't work on Firefox:
       | 
       | https://news.ycombinator.com/item?id=41157383
        
       | pentagrama wrote:
       | Happy to have these models running locally on a browser. However,
       | the results are still quite poor for me. For example:
       | https://imgur.com/a/Dn3lxsU
        
         | sdesol wrote:
         | It's not too bad given that it runs in your browser. I took
         | your prompt and asked GPT-4o mini to elaborate on it and got
         | this https://imgur.com/a/qmQ7ZHl
         | 
         | The burger looks good.
        
       | n-gauge wrote:
       | I like the local running of this and learning about how it works.
       | 
       | Q:These models running in WebGPU all seem to need nodejs
       | installed. I that for just the local 'server side', can you not
       | just use a python http server or tomcat for this and wget files?
        
         | andrewmackrodt wrote:
         | Had a peek at the repo and it looks to be a react frontend, so
         | a JavaScript runtime is needed to "bundle" the application in a
         | way browsers can consume. If you had the dist folder then I
         | imagine you can use whatever web server you want to serve the
         | static files.
        
       ___________________________________________________________________
       (page generated 2025-01-28 23:01 UTC)