Alias-Free GAN (opens in new tab)

I audibly went "GAH!" when that scrolled into view. Impressive work.

There are some interesting 2d things our eyes do for 3d. If something is on the ground, half is above the horizon and half is below. Parallax is a 2d phenomenon.

Teeth too. Partially covered objects in 3D space have been hard for a GAN to figure out. (See also hands)

I wonder what dataset you could even use to tell a GAN about human internals. 3D renders of a skull with various layers removed?

Mismatched reflections across eyes is the dead give-away for me.

First, that's not the same technique and it's not being used for the same purpose.

Second, Hollywood doesn't care about that problem. They will take the best application of the technique, and they don't care if they have to apply a few manual touchups on the result. As long as there is one way of using the system to do the sort of thing they showed in the sample, it won't matter to them that they can't embed a full video game into the neural network itself. They only care about the happy path of the tech.

Someone's probably already starting the company now to use this in special effects, or putting someone on research in an existing company.

> I do ML every day.

> I go to the github. Maybe model download link is there. I see zero code

Paper was released today. Chill. They said they will release the code in September (I'm guessing late September). The paper is also a pre-print. They're probably aiming for CVPR and don't want to get scooped.

> Model first. Code second. Paper third.

That's how you produce ML code and documentation but that is not how you release it. I guarantee you that they are still tuning and making the model better. They're were still updating ADA till pretty recently (last commit on the pytorch version is 4 months ago, to code).

I originally wasn't in CS, and when I first came over I wasn't in ML. We never had code. The fact that ML publishes models AND checkpoints is a godsend. I love it. Makes work so much easier and helps the community advance faster. I love this, but just chill. The paper isn't peer-reviewed. It is a pre-print. They're showing people what they've done in the last 6 months. It's part publicity stunt, part flex, part staking claim, but it is also part sharing with the community. Even without the code we learn a lot because they attached a paper to it. So chill.

You're clearly disillusioned with the general accessibility of ML research, but I don't think your cynicism is warranted here. Take a look at their prior works[1], and I think you'll agree they go above and beyond in making their work accessible and reproducible. There is no reason to doubt the open-source release of this work will be any different. As to why the release is delayed, I'd speculate it's because they put a significant additional amount of work into releases and because releasing code in a large corporation is a bureaucratic hassle.

[1] https://nvlabs.github.io/stylegan2/versions.html

The repo says the code will be available in September: that's a reasonable timeframe for the necessary polish/legal clearance.

I am not into ML, but from time to time I like to look how this is made and remember only once seeing the code and a model, which I thought was exception from the "norm". Good that more people are calling this out!

Thank you for calling this out. It's critically important that people understand the difference between model, code, and paper and what they mean.

It's also important that people understand that even if code is provided, it's commercially useless. From the NVAE license as an example[1]

> The Work and any derivative works thereof only may be used or intended for use non-commercially.

It's a great example of the difference between open source (which it is) and free software which it is not. So we're back to square one where it is probably best to clean-room the implementation from the paper, which is nearly useless to reproduce the model.

[1] https://github.com/NVlabs/NVAE/blob/master/LICENSE

You won't need that much source code to verify their claims.

Their central improvement is that they limit the generation of high frequencies by ReLU through a upsample-ReLU-filter-downsample sequence.

Their theoretical section explains quite well why high frequencies can be proven mathematically to cause issues. And their practical implementation using filters to cut those off is very straightforward.

If someone tells you "The microphone recording had 50Hz noise so I used a filter to remove it", that's pretty much good enough for someone with experience in the field to replicate their results. This is the equivalent in AI. They uncovered a simple basic issue that everyone else overlooked, but once you know it, it seems obvious in retrospect.

Would you not want the data and code used to train the model, rather than the trained model itself?

Edit: the Debian Deep Learning Team's Machine Learning Policy explains why.

https://salsa.debian.org/deeplearning-team/ml-policy

I hardly understand this comment. Statisticians have been publishing and arguing about models for 100 years now. No one required code to verify authenticity of research. I suppose it is the sorry state of machine learning research that the methodology is so poor that a person cannot verify the research from the paper.