Hacker News Reader

Comment by ndr | Hacker News Reader

The image generation results are extremely poor, but it's exciting that it does _anything_ in the browser.

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-...

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.

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.

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.

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.

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.

Oh wow okay thank you for the context

Yep! Less than 1B in total [0]:

> 860M UNet and 123M text encoder

[0] https://github.com/CompVis/stable-diffusion/blob/main/README...

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.

Reply on news.ycombinator.com