HN Explorer

25 comments

[ 2.8 ms ] story [ 70.7 ms ] thread
hathat
wow this is indeed fast

actually I don't understand why other SDXL services are much slower than SD 1.5 or SD 2.1 models, they seem to be about the same size, but normally takes 2x to 3x time comparing to the previous models

letonlyter
Glad to see SDXL models becoming more and more accessible, the images are much better
sylindome
have you checked out finetuned models?
uIvspro
perhaps it's just my use case being not complicated (or specialized?) enough, the original model is already pretty good at generating realistic images, unlike previous models, which still need a lot of prompt engineering or fine-tuning or bunch of hacks
ytitbb
agree soon everyone will use sdxl to generate high resolution images instead of the old models
9thtown
how does it compare to midjourney
sylindome
prop model is probably gonna stay better for quite a while
no-maaagic
not necessarily, the power of open source community has been proven over and over in the AI industry
no-maaagic
Foe example there were once so many closed source “deep learning frameworks”, but these days only pytorch matters
Daddyshoma
is openai working on gpt5 that will cover all modals?
logpta
nice work. although I'm wondering what's the roofline performance, and how much more optimizations can users expect
juswth
> Our joint endeavor has resulted in 46% (12.6 steps/s vs. 8.6 steps/s) speedup compared to stock PyTorch.

How about jax

yytab3
not sure why isn't this thread up, but imo this is great advance in image generation. can not emphasize more on iteration speed improvement
tiantiandss
what are the optimizations done in this model? I would like to know more details
uIvspro
rather than A100 or H100, I would love to run these models on more affordable gpu cards like A10 or even T4
wohenniubi
A100 is more suitable for running stable diffusion, especially if you can afford doing batching