Somewhere between "base compute" and "4x compute".
So maybe you "just" need to know how to create a certain type of diffusion transformer model and then train on a ton of videos, but with an adequate amount of compute. Which is probably a LOT for training and inference to get more realistic results.
I’ve launched apps with free trials in the past and when I allowed email, there were 100 fake emails from temporary email services for every 1 real one. Had to do more work to block those as they were eating up all the resources from paying users
If you have an extensive demo page covering my use case, I would prefer a low-cost entry offer to use my Google account. Upgrading to a more expensive option later is usually not such a hard task for me when I really love the product.
Lately, I had a case when someone informed me that he would take my e-mail from his newsletter mailing list because I seemed to be a dead address (and every address generated cost for him). I replied that I did not like to be tracked (implying that I had read the newsletter) and we had a nice conversation. Not every fake e-mail is as fake as it seems ;-)
Apparently, Thunderbird can be quite good at beating E-Mail Tracking.
Doesn't this just further the point that your potential customers don't want to identify themselves to try your product? Punting on anti-sybil or anti-spam by requiring a strongly identity-linked account (Google accounts can't be created from VPNs or without phone numbers, for instance) is user-hostile.
If your free users are using up paid user resources, that's a different problem.
One of the nice parts about starting a business is that you get to choose your customers. There is no ethical or moral rule that you must provide access to everyone who wants to use your app. You are even allowed to make decisions that reduce your revenue in favor of other priorities.
The problem is not that I want peoples real information, I really couldn't care less. It's the fact that some of these are clearly the same people creating hundreds of emails to just keep resubscribing to a free trial.
I could tell because there were emails with the sequence of:
I've become even more drastic: if you want me to sign in at all before I can try out your product "for free", you already have one less potential customer.
Not much choice with a GPU backed service. Way too expensive to not have at least a verified email filter in front of it, and Google is the simplest most "universal" plug and play OIDC provider.
Ya it's really the Google-as-only-option problem. I immediately close the window when I see that and move on.
Are there any "privacy-first" SSO options? Maybe Proton or Mullvad or some similarly positioned company should think about this. Or maybe it's just way too hard to break into the Google/Apple duopoly.
If you’re going to offer SSO, at least take the lowly amount of time to add some other providers. I genuinely don’t believe anyone is launching so quickly that they don’t have time to add a second or third option.
Am I missing something? The AI video examples on this just look like non-temporally stable image gen frames with the wobbly/morphy look. I don't think the comparison is there to what Luma shipped.
The problem for me is that they took our design, styles, fonts, colors, gradients, and name for a very similar product and shipped it with quite an impressive team that is well-funded by A16z.
Their product is much slicker than what we are doing with Pollinations, of course. We are a few people working on it in our spare time.
It would have just been nice to get a kind of mention or have someone reach out. I'm all for getting inspired and copying.
The current state of the Pollinations site is very different. The Luma Labs page looks like what Pollinations had on its site about nine months ago when the Dream Machine was a more significant focus of ours.
I could get into the weeds and show side-by-side screenshots but it seems like the wrong direction to invest time. The person who did our design didn't think it was very funny.
In my opinion Gen AI startups should be extremely careful in crediting because the whole copyright thing is already such a hot potato with artists and designers.
If it's any consolation, the Luma website is so laggy and poorly-designed that I don't think I could recommend their service simply as a matter of principle. The Pollinations site is much more static and doesn't have any distracting or finicky on-scroll effects, which at least makes it enjoyable to read.
We're giving free access to a model that other people trained on other people's data. The code for all improvements or tweaks is accessible in public repositories. From a copyright perspective, I think one can make a fair use argument.
Nevertheless, it would be great to credit artists and designers whose work was used for training. Practically, it's not easy because I think one would have to credit humanity.
Every image in the training data influenced the model's output to a certain degree, so one would have to credit every person who uploaded an image to a public place.
One could argue that a model should not be trained on public data in the first place because we can't fairly credit all individuals who contributed to it.
Like the alternatives, it appears to be limited to adding forced movement (shift each frame by X, Y) and only the most predictable motion, like moving fog, water or sand.
These constraints (and most of the morphing) come from these type of image generators always going for the most plausible interpretation of noise. They don't have any actual temporal coherence yet.
It's very cool, but fairly limited in practical application with the current capabilities.
I really wish AI generators displayed the prompt used to generate content so that others can learn from it. E.g, https://storage.cdn-luma.com/lit_lite_inference_text2vid_v1.... was generated from "A gray and silver Maine Coon playing with an android" and it's unclear how to make it as good as the official demos.
Awesome video quality hahah, the frogman wearing a leather jacket was on point! But apparently they said themsellves that accurate text rendering has a long way to go
Also found these guys providing API for Dream Machine: https://piapi.ai/dream-machine-api, would be interesting to see what developers can do with this in their own chat bot!
Reminder that "This is the worst version of the AI video creation we'll have from here on out."
As we mock the new Will Smith Eating Spaghetti video, every update is an improvement on where we're at and a new, higher baseline of the worst functionality.
Doesn't mean it'll get any better though. Most technologies and algorithms have an end point. People thought flying cars were inevitable and imminent because after all, cars got so much better and feature-rich within a very short timespan. And then innovation practically stopped because there's a limit to possible energy storage and utilisation unless some magical new innovation comes along. But it hasn't.
These generated videos are exactly the same nonsense as the spaghet videos, they've just got more data crammed in there. But nothing about them is consistent.
I suspect that this is a compute tradeoff they decided to make. Uses much less compute than SORA, so feasible to scale up for the public, but that much less coherent. If you look at OpenAI's technical report for SORA they show examples with "base compute" and "4x compute". This output looks like an estimated "2x" to me.
I wonder if they are really talking about the model size.
That's the version they released as an "open" model. Their SD3 model behind an API works just fine, the released medium model has been basically been lobotomized to hell and back in the name of "safety", because apparently human anatomy is icky and nipples are the devil. Unfortunately, that sort of heavy handed model lobotomizing leads to nasty side effects that turn even innocent renders of people into Eldritch nightmares as the model has been forced to forget that a human doesn't have limbs growing out of random places.
I asked it for "A dog catching a treat in slow motion, it's chops flapping around comically"
https://imgur.com/a/dog-catching-treat-slow-motion-its-chops...