hwers

If you have even just 4gb stable diffusion will run fine if u go for 448x448 instead (basically the same quality).

Here’s another one in gif form (right image) https://twitter.com/johnowhitaker/status/1565710033463156739 These seem really useful to get intuition about im2im

We could always do im2tex via just clip embedding the image. The idea that you could hide/sell prompts is silly. (Having human interpretable im2tex is cool tho.)

This message brought to you by replicate.com

I have 4 gb and it takes about 9 seconds for me :) (Tho at 448x448 but there’s no real difference in quality.)

Whats wasm32 (vs just normal wasm)?

whats the difference between float16 and bfloat16?

Same. Things are getting so noisy and hard to find that even if good howtos exist somewhere on the internet, finding it can waste hours vs just word searching your own documents.

That’s because $20 is still humiliatingly low for the workload

Yeah stylegan is rarely this good at super diverse data like this

Lol imagine if someone said this as a prediction when photoshop was invented.

> It puzzles me when people say that small companies can’t retrain every day because they don’t have FAANG-style budgets. It costs a few dollars, at best, to retrain many xgboost or scikit-learn models. Most models are not large language models

So true. This is a pretty easy flag that someone hasn’t actually been down and dirty with trying to train any actual models and doesn’t know what they’re talking about

What are you using it for if can share? I’ve thought about training some of these and releasing the weights but I’ve never found a reason they’d really be useful personally so it never really happened

Anyone know what SIMD means at the bottom layer is here? I know what SIMD is I mean in the context of the preprocessor (and being the worst offender apparently).

He mentions open source and education which I’m really excited about. There’s been a big lack of accessible from-first-principles resources on this (plenty of abstracted libraries though). Now is a great time to build up excitement among amateurs to learn about this stuff imo. (Right near the super cool text-to-image breakthroughs.) It shouldn’t just be openai and google building these things.

Surely 1bn could have afforded a lot more and better marketing..

So embarrassing if he ends up being forced to own it and everyone knows he tried to back out.

This is such a terrible look for Elon. I’ve really liked him up until now but this just feels like playing with our emotions. Honestly starting to question how much hope there is for mars to happen at all.

I don’t even think it blocks the channel long term from your recommendations. I’m pretty sure they’ve tried to sneak in an old channel I’ve explicitly said not to recommend again in my recommendations via whatever multi armed bandit with noise algo they’re running.

I guess an interesting way to translate this technique to text-to-image would be to get an image from a database that matches the text query (via CLIP) and feed that + noise into a diffusion model that only does a few denoising iterations (and no clip guidance maybe). Would be a lot faster than from-scratch diffusion.

(Another way could be to redo the architecture to include a “inspired by this image” input, which is queried from an image server at inference time.) Anyone have other ideas?

New product lines are always on the horizon

Love that analogy at the end, very well said

It’s silly that you’re being downvoted because this is absolutely true. The idea that it would go to zero is mad, it will without a doubt go back up to similar numbers again. This is just a little gift

This whole “what matters is the founder” idea really discourages me. It’s such a VC perspective. The VC has a flock of people they pick from to invest in and if your traits doesn’t fit the necessary requirements you shouldn’t start a startup. And I’m sure it’s true too. It’s just that it feels like if you don’t have the grit etc necessary you’re just out of luck and out of this club and it’s not really something you can work on. (After all if it was something you could improve then it could be taught but PG et al seems to think - and I think they’re right - that it’s an innate thing.)

Well my point was that you could extract it into a wasm blob without onnx (library independent). See tensor4 as linked in the other thread.

It's just a form of stress coping. It's entirely delusional and I think it's very important we not allow this type of thinking to be a normal form of thought (as uncomfortable as it is).

Yup. In torch you can use torchscript and in jax you can 'jit' a function and both turn your NN function into a fairly easily compliable form which could then be translated to wasm.

Here's a cool project that does something like this (though I believe it's basically superseded by torchscript): https://github.com/podgorskiy/tensor4 Here it is running stylegan in the browser (via wasm): http://podgorskiy.com/static/stylegan/stylegan.html

It's hard for me to see "X gets more funding" as something to celebrate. Funding is essentially zero sum so that just means something else is now less funded as a consequence.

I know even publishing the code for stuff like this is generous but it'd be so cool if the standard practice included publishing a little wasm binary blob as a library independent network with weights to play with without having to go through colab. (Many frameworks have a way of creating that automatically already, e.g. torchscript or jax jit.)

I agree. The limit encourages creativity and brevity