Hi, I'm the creator of multimodal.art, I didn't overlook it, but there's no "specialized" NSFW content maker to be highlighted - this Vice articles just show people using the model in different iterations to generate NSFW content; you don't need a specialized notebook/tool for that, a few ones on the post can do it (others have a NSFW filter that comes in by default).
Additionally it is important to note that model was licensed under the OpenRAIL-M LICENSE which is not as permissive as an MIT license and forbids certain outputs to be shared or purposes to be built as apps
I thought only the "derivatives of the model" are under the user restrictions in the license, and are very permissive. The outputs of the model are very briefly covered in the license text
> You are accountable for the Output you generate and its subsequent uses. No use of the output can contravene any provision as stated in the License.
Goes without saying, but enthusiasts are working on it and making progress. Here's a quote from a discussion on it yesterday:
as far as I am aware, I may have the only semi-tuned nsfw model, though it's not all that great. There are a lot of concepts that SD needs to learn, and they're difficult to teach. It's very likely that by the time I get something usable, it's going to destroy all of the other concepts. @<another_user> also has a fine tuned model, but it's only tuned on generating closeups of female genitals. If that's all you want to generate, then he's produced some really good results. As for distribution, I wouldn't know where to begin, considering the model is 11GB
I would expect the % of image generation compute power used for making NSFW material to come down over time, just like the % of home video minutes or digital photographs used for making NSFW material went down over time.
My 3 year old GPU generates one image every ~5 seconds, and according to the documentation draws 215 watts at a maximum steady state load (TDP). That's very roughly a kilowatt second per image.
The internet tells me that a 2022 honda civic takes about 0.07 liters of gas per km. And it also tells me that that is equivalent to 2394 kilowatt seconds. I.e. 2394 images on a 3 year old GPU per km travelled using a new and fuel efficient model of car...
I'm not worried about this consuming a significant fraction of the power on earth.
I mean, its not hard to imagine where this goes next: video (and eventually 3d/VR scenes). Say 60fps -> 2394 frames/60fps = 40 seconds. That's equivalent to driving your car at (3600 seconds per hour / 40 seconds / km) == 90 km/hr. Yes, your GPU is older, but there will also be pressure to increase the resolution and fidelity of the generated content to match.
Ok, sure, maybe there's demand there. But how would the logistics work out so that it managed to become a problem?
Are people
a) Waiting 5 seconds/frame * 60 fps = 5 minutes / second for a video to generate on their personal computer, and doing this constantly enough that it manages to become a problem?
b) Buying computers that can do it real-time, but therefore output vastly more heat, requiring thermal management system akin to a car driving at highway speed?
c) Renting these computers at considerable cost to make these videos?
As long as enough people watch each video (or one person watches it enough times), the energy usage washes out to become negligible compared to the amount of human time invested. I just can't see a world where enough people are managing to consume a kw minute/second producing videos for themselves to watch only once or twice that it becomes an issue.
Personally I'm optimistic that energy/compute is going to continue going down substantially (in which case even real-time video generation might not be an issue). If it doesn't and we don't become substantially better at efficiently synthesizing video, I can't see personalized single use video generation being a thing.
> a) Waiting 5 seconds/frame * 60 fps = 5 minutes / second for a video to generate on their personal computer, and doing this constantly enough that it manages to become a problem?
There will be much more compute resources thrown at it to make it render in real time. We're not there yet, but I can see a path to that happening in the next few years.
> b) Buying computers that can do it real-time, but therefore output vastly more heat, requiring thermal management system akin to a car driving at highway speed?
Why not? We already have billions of cars driving around outputting heat. Its an incredible expenditure of energy, sure, but perhaps the value of generated content entertainment will match the value of car transportation.
> c) Renting these computers at considerable cost to make these videos?
I imagine longer term, the opex (i.e energy costs) will dominate the capex (GPU HW). The price of going into a generated world could be similar to going for a drive.
> As long as enough people watch each video (or one person watches it enough times), the energy usage washes out to become negligible compared to the amount of human time invested. I just can't see a world where enough people are managing to consume a kw minute/second producing videos for themselves to watch only once or twice that it becomes an issue.
This is where I strongly disagree. The democratization of skills and tools in creating content will break the one to many media model. You saw this in a large way in what the internet did to content distribution, in how the number of independent people creating content skyrocketed. These models will do the same for content creation. I predict most people will consume content personally generated for themselves or in small groups.
Here's an example: a group of friends puts on their VR headsets for their weekly DnD session. The DM begins describing the scene, which autogenerates around them. Each character can then respond with their own actions / path, and the scenes react dynamically. The hour session costs them $10 in compute/energy.
I'm mostly spitballing. I would imagine that we still have a couple of orders of magnitude reduction in energy costs that can be squeezed out of these models with improvements in specialized HW. But it will be matched against the insatiable demand of consumers for richer interactivity in content.
It really feels like for 3D the better quality/compute trade-off would be to have the ML model generate 3D models and animations, and then use a more traditional 3D rendering pipeline (by then with ray tracing and denoising).
I agree with your analysis, but I don't think that's how most people interpret global compute power consumption. The stored energy in the gasoline is not counted in global energy production figures, but the electricity used to power your GPU is.
As someone who stopped looking at porn about two months ago, after years of porn use, I can tell you: it’s going to be highly subjective.
I used to watch porn basically daily. But then after finally deciding to stop watching porn, the idea of porn itself is downright off putting to me. I don’t even quite know how or why. It just is.
And I imagine it will be the same for others with AI generated porn.
Technically you could get porn of that one exact weird turnon you have, that there is literally zero porn existing now (except if you pay for a costum video/photoshot).
Is this good? maybe... maybe not. Since most of the "normal" stuff already exists, it'll either be something "too extreme" for classic porn studios, or stuff using non porn people to turn into ai-porn stars.
https://www.vice.com/en/article/xgygy4/stable-diffusion-stab...
Why'd they "overlook" it? Probably more culturally significant and controversial than any of the others. It's the natural elephant.