Are they using me to refine the model in some way?
Is that a wide spread myth? I thought it’s widely accepted that GAN is really good generating these artificial pictures (it’s what started DeepFake after all) when you know your model’s “button”. Similar to how this uses GAN since they have a model “boundary condition”. While humans are diverse, we have a set of repeatable features (two legs, two arms, etc). Diffusion models are great because you can control the latent space with something way more generic, like text hence why it’s been so much more mainstream.
Edit: actually I might be misremembering, I think Deepfake used VAEs?
You come across as severely, _severely_ biased and reactionary.
Get off of Twitter, maybe? I don't know but I have seen a ton of work done on large scale GANs. The incentives are pretty fucking obvious - they are faster to use. Of _course_ people want them to work for the same usecases as diffusion.
Maybe they will someday! But they aren't currently. You're referencing a zeitgeist that I am not familiar with and therefore come across overly defensive.
GigaGAN is the best quality out of those and requires 7 loss functions and is incredibly complicated.
Sure GANs can scale, but Diffusion models are drastically easier to scale.
Diffusion models are only 'drastically easier to scale' because researchers have spent the past 3 years researching pretty much nothing but diffusion models, discovering all sorts of subtle issues with them and how to make them scale, which is why it took them so long to become SOTA, and why massive architectural sweeps like https://arxiv.org/abs/2206.00364#nvidia were necessary to discover what makes them 'easier to scale'. If this level of brute force and moon math is 'easy', lord save us from any architecture which is 'hard'!
Reading your comments make me feel like that you believe that just every researchers (even extremely smart dude like Karras) just switch to diffusion models because they are idiots, they should have instead focus on GANs and today we will have GANs that are as powerful or more than the diffusion models we have today and also work one step; this is just a weird delusion. Diffusion models are just simply much easier to train (just a L1/L2 loss in most cases), write (for example your buggy BigGAN implementation), they usually work out-of-box on different resolutions and aspect ratios, you can just finetune them if you want to create an inpainting model; and for what is right now you just need much less compute to reach a good image coherency or maybe just reaching a coherence that as not been achieved by GAN models; like I would be curious even on a small scale experiment what a GAN (with ~55M parameters) would be able to perform after a 1-day/2-day GPU time of training on Icon645 dataset, because my diffusion model I can assure is much better than I could have imagine while being trivial to implement (I just implemented a Unet as I remember one, nothing rigorous and of course no architecture sweep).
> for example your buggy BigGAN implementation
The bug was that the batchnorm gamma was initialized to zero instead of one, so all the model params were being multiplied by zero instead of one. Literally any model that uses batchnorm is susceptible to this bug, and it certainly has nothing to do with GANs.
I think it’s too easy to spout dogma, and calling gwern delusional is something I’ve learned from experience will prove foolish as t approaches infinity.
Also no I don't take "he is gwern" as an argument ahah
This is what tends to happen when you find a superior method.
GAN’s are fine, they have plenty of promise for tasks requiring rapid inference. But diffusion models beat GANs on robustness and image quality every time.
I assume its cheap because they are burning money to build a business, its not fast at all, and the quality... sucks.
> And if you're wondering how it can look so good
I’m not.
I'm wondering why they’re trying to get people to use something worse than using a decent photorealistic SD1.5-based checkpoint with some basic prompt templating.
Not saying GANs can't be awesome, just that this site isn't what I’d use to make that case.
https://images.generated.photos/0wV1dBnZ15hGneEfqfZT7SdEIill...
My prompt was simply "Standing in front of a rocket.".
That looks more like a machine, an android perhaps, than an actual human.
It's not nightmare fuel it's a nightmare engine
One could argue that such an outlet could even reduce incidents of direct sexual assault of children by pedophiles, but there is also a counter-argument that it would instead serve to "whet the appetite" and encourage such behaviour. And of course there are other counter-arguments; it could make actual CP more difficult to detect, for one. Finally there is the argument from the perspective of fundamental morality, that depicting children in a sexual manner is wrong in and of itself, and therefore the various potential effects are irrelevant. (Much like it's wrong to murder an innocent, even if you could harvest their organs and save five others as a result.)
I don't think those two scenarios are usefully similar. Murder is wrong because of its effects. This would be more like if you could murder someone without _anyone_ dying.
I have to say though that this argument starts making a lot more sense with current VR tech. Sometimes it's really disgustingly realistic.
While I'm still not in favor of banning violent games, I do think a valid argument is emerging. I've never truly had a feel of fright playing on a 2d screen but since VR the spiders headcrabs of Half Life Alyx really make me panic sometimes.
On the csam side I'm just not sure if this would diminish or increase their interest. I would leave that up to someone who studied for this kind of insight.
2. Efforts to stamp out real child abuse may be undermined by a flood of AI-generated false positive imagery
3. When people see something over and over again they start to think that it's normal. AI generation of this kind of material (something which can be done at a huge scale) risks normalizing the sexual abuse of children.
I'm sure there are many other arguments beyond these.
Don't get me wrong. I have a built in revulsion toward CP. But I wonder how much is that built in revulsion helping vs hindering rational discussion - just like it took me like 15 years of reading and pondering to remove the indoctrinated deference I had toward religion, and I hope I can have a more stable, less knee jerk discussion about religion. I think CP is wrong, but I also Think murder and torture are wrong. The latter two are amply represented in entertainment media. What is the Venn diagram, what is the delta that makes cp different?
Currently? The fact that all the models need training data and the law will see that as victimizing the people who were used in the data set be they adults of children.
Overall? The fact that it's disgusting and pedophiles deserve things which I can say IRL and everyone agrees with, but on HN will get me banned.
Many countries ban underage anime porn too. Children and their likeness are off limits.
But its not hard to find many other sexual acts that at least some people, if not most people, find disgusting. So I'm not sure disgust is a useful criteria here.
Most of what makes pedophilia so disgusting is the fact that a child is involved. Remove the child from the equation and it's more in line with other sex preferences.
People can't really control their preferences. We've grown to understand that there are healthy, and unhealthy ways to satiate those preferences. Most of those revolve around the consent of those involved (taking as fact that children cannot consent.)
One could argue that AI allows for a healthy outlet for a pedophile.
You mention children being off limits. In other societies homosexuality is still criminal. Even where it is legal many consider it "off limits" or "disgusting". Laws, and emotions, in this space are not always logical.
There will be second-order arguments (think - the gays are trying to turn my child gay) but most of them fundamentally misunderstand how human sexuality works.
In fact in a lot of those countries where fully consensual homosexuality is highly illegal, marrying (with all its implications) what we would consider a child is legal and "normal" and so is forced/arranged marriage (again with all that implies). I'm really lucky to have grown up here.
As a complete aside, forced marriages are not the same as arranged marriages. In many cultures arranged marriages (which seem completely bonkers to someone grown up in a western mindset) are both common, work well, and often good.
Clearly neither arranged marriages, nor "choice marriages", are guaranteed. The divorce rate in the west bears this out. Both work well, both work badly.
Again arranged is not the same as forced - clearly forced marriages are not something I consider a good thing.
Would you say the same about homosexuality? Nobody chooses their own sexual preference. Pedophiles are unfortunate victims having a sexual preference that they can never satisfy without causing others harm. Even if people could pick their sexual preferences - which they can not - among the countless possibilities, children would be one of the worst choices.
https://en.wikipedia.org/wiki/Wrongful_execution
Encouraging state violence based on the fact that something is "disgusting" is not a good idea. Law needs to be grounded in concrete harm. The lines currently seem ill-defined, and existing laws in this area have been used against children (see State of Washington v. E.G., where an autistic child was charged and convicted with distribution of child pornography for taking a selfie and sending it to an adult).
As something you generate in a photorealistic image generator you are building a business around?
The fact that it is a serious crime in many jurisdictions and, even where it isn’t, photorealistic child porn images that get noticed anywhere or going to result in uncomfortable conversations for everyone involved in the process of establishing that they aren’t evidence of a crime.
Blatant disregard for children aside, the larger issue here is the presupposition that we should someone permit all urges. Not every thought that passes through your brain should be “explored”; some thing are just bad and should be fought.
Keep in mind that science can tell any story you want it to tell as long as you interpret the data in the way you need. That’s why some things, such as our children, cannot be left to slippery slopes such as this.
We’ve already got more than enough data proving that pornography is absolutely destructive and a net negative on all of humanity. I don’t think we need additional data to prove that bringing kids into the mix would somehow make it better.
Paedophilia is not synonym for child sex abuse, and to the extent that conjecture hasn't been ruled out by study already, that's at most an argument that its an appropriate function for tightly-controlled research-supporting tools, not a general public business.
So I was relatively unimpressed with the first few images, they looked incredibly unrealistic. But then after reading this message i decided to add "topless" to the notes just to see how crazy the result was.
I was actually impressed, it returned an image that actually looks lifelike once I gave it an adult request. Which reinforces the idea that this was optimized for porn.
For those curious, this was the result I was provided. NSFW of course, topless female gender (ai generated): https://generated.photos/human-generator/64e6c26e190809000fb...
[NSFW] https://generated.photos/human-generator/64e7e1aee4125d000cd...
And then I requested a full bush...
[NSFW] https://generated.photos/human-generator/64e7e47ff56326000b9...
>See for example BigGAN
I remember when you try training a BigGAN model on anime images, the quality was bad. Now look at this example, one single GPU, 1.5M images with a diffusion model: https://medium.com/@enryu9000/anifusion-diffusion-models-for... ,the difference in quality is absurd, you can say this or that is not true but the quality speak for itself, obtaining good quality on complex distribution is much easier with a diffusion model than a GAN.
For example in the case of the site linked they have conditioned the model on poses because you're not going to get anything close to be coherent without them with a simple StyleGAN as they say they're using.
Because there was a bug in the code, in a part unrelated to the GAN itself.
> the difference in quality is absurd
Yes, it does help to train on anime with code that isn't buggy. (BTW, Skylion was getting good results with GANs on anime similarly restricted to centered figures like those samples, he just refuses to ever publish anything.)
Also the bug is probably related to the added complexity of training a GAN model in comparison to a diffusion model.
I don't think anyone is wondering this, especially if they are used to playing with diffusion models.
"And if you're wondering how it can look so good when 'everyone knows GANs don't work because they're too unstable', a widespread myth, repeated by many DL researchers who ought to know better, GANs can scale to high-quality realistic images on billion-image scale datasets, and become more, not less, stable with scale, like many things in deep reinforcement learning."
My interpretation from how this is written is that you are saying: "Researchers say that GANs are unstable at inference time." Is that the correct reading? If so, where have you seen this sentiment? More commonly I've heard that people criticize GANs at inference time for mode collapse and monotony (which comes from training), not instability.
Or are you saying that researchers say GANs are unstable during training, which is a common criticism. Don't you feel this is the case? A lot of different tricks are used for getting generators that are roughly balanced with more powerful discriminators so that the adversarial game is balanced, like TTUR and encodec's weight balancer etc. In this case, are you saying that GAN training is as straightforward as diffusion training?
In my experience, GAN training involves an unholy number of heuristics and best-practices are still quite murky.
I am eager to hear your response.
[edit: my experience with GANs and diffusion models isn't about taking image problems with existing working models and scaling or refining them, but about applying GANs and diffusion models to different domains (audio) and on novel problems with different kinds of conditioning. I would love to see more controlled experiments comparing weak and sophisticated generative backbones as the experimental control and varying the training regime across different GAN + diffusion flavors.]
eg (NSFW!): https://generated.photos/human-generator/64e741e6c4c0f80009e...
Wrong number of limbs, wrong number of fingers on hands, etc.
We trained for 36 days on a v4-8 on 558k images.
https://nyx-ai.github.io/stylegan2-flax-tpu/
We were hopeful GANs would beat diffusion models when trained on specific domains. But we’ve now switched to Stable Diffusion and Dreambooth training which has proven more much efficient for this purpose.
I still have hopes for GANs! I miss their insane inference speed.
(And if you're wondering how it can look so good when 'everyone knows GANs don't work because they're too unstable', a widespread myth, repeated by many DL researchers who ought to know better, GANs can scale to high-quality realistic images on billion-image scale datasets, and become more, not less, stable with scale, like many things in deep reinforcement learning. See for example BigGAN on JFT-300m https://arxiv.org/pdf/1809.11096.pdf#page=8&org=deepmind , GigaGAN https://arxiv.org/abs/2303.05511 , Projected GAN https://arxiv.org/abs/2111.01007 , StyleGAN-XL https://arxiv.org/abs/2202.00273 , or Tensorfork's chaos runs https://gwern.net/gan#tensorfork-chaos-runs . 'Computing is a pop culture'...)