It's titled "What's New At Blockade Labs", a title which continues to display even when you scroll down and find the terms of service below the news... which is presumably why the button is labeled "Confirm"
I was hoping someone had found a better way than radiance fields to generate actual 3D models. This is creating 360 panoramic images, which are not "3D" unless I'm missing something. Still impressive tech, though.
I do it with heightmaps. It can generate human faces, rocks, planets, terrains, all at high detail. There is no need for attempting to generate 3d models for everything, when you can create “blueprints”.
If you check out their Discord they have tools in (closed I think) beta that generative full 3d scenes. Lots of really cool demos get posted their daily. I'd actually say that as someone using it purely for panorama generation that I'm in the minority of use cases.
while nerf.studio is impressive, the parent comment is asking about radiance field free methods. NeRF stands for Neural Radiance Fields and all of the algorithms in that repo use radiance fields
1. Point-NeRF: Point-based Neural Radiance Fields - 2022: This paper presents Point-NeRF, a localized neural representation that combines volumetric radiance fields with point clouds to model complex local scene geometry and appearance. It achieves better rendering quality than NeRF and is more efficient in reconstructing complex and large-scale scenes.
2. TensoRF: Tensorial Radiance Fields - 2022: TensoRF proposes a tensorial radiance field representation that can achieve fast reconstruction and high rendering quality. It is significantly faster than NeRF and many previous methods.
3. GazeNeRF: 3D-Aware Gaze Redirection with Neural Radiance Fields - 2022: GazeNeRF incorporates 3D awareness with gaze redirection by applying a rotation matrix to the disentangled eyeball feature maps. It represents a static scene with multi-layer perceptrons and maps the 3D positions and viewing directions to a density and radiance for generating novel views with volume rendering.
This is an example of using a simple AI technique (stable diffusion) and really getting the most out of it. It's really just a clever execution and presentation of what is now just about the most commonly deployed type of generative AI.
As opposed to what I am doing, trying to write a whole program from scratch, probably the most complex and difficult to pull off application of AI. It's quite challenging to explain what you want and get a functioning attractive output. Also to run arbitrary backend code I need to deploy VMs which are relatively expensive. People don't seem interested in paying for my website.
So right now I am trying to think of something similar to my original idea but maybe easier to accomplish successfully and with less work for the user.
Maybe if I can finally do a good job with filling web page templates that is what I should start with. Because even if it can synthesize code for some unique requirement, the fact that the web pages it outputs don't look attractive is probably making it hard for people to understand that it's useful.
But basically I think this blockade thing is a good lesson for me if I can find a way to adapt a similar approach to what I have been trying to do.
I gave it this prompt, which I didn't spend any time refining, and it came up with a world that does a pretty good job of matching it.
> A world made of very strong rock that has lots of rain. The rain contains acid, which wears the rock down. There is lots of soil, but it s in very thin layers because the world is relatively new.
I prompted it "small heavily forested islands in rough seas" and it gave me one large forested land area in snow, with a bunch of rocky treeless islands in calm water. <shrug>
Serious question: How is this tool free? Are they burning money making it publicly accessible? Or something else? Every other AI gen tool I've seen is pay-per-use.
But theoretically the devs could easily add a more "direct" pipeline with better results. There is a HF diffusers function for encoding input images like they are text prompts.
I'm familiar with interrogation, and image -> interrogation -> t2i will usually generate a completely different image that the original (and it should!), unlike i2i.
Well there is interrogate, there is i2i, and then there is something in between. You encode the image directly into latent space vectors, and it applies the general "style" of the input, much closer than a text prompt would.
Skybox Labs makes really great backgrounds, although if you ask it to add people it will try and generate AI monstrosities!
Being able to take the same scene and just change the time of day is really cool, my generative storytelling demonstration site[1] uses that feature to show the passage of time.
My largest wish is that generation wasn't so slow. I get their main use case is actually generating 3d environments that can be navigated around (and their Discord has some cool examples of early adopters using it for just that!) but I just need fast generation of panoramas!
Really what it comes down to is that OpenAI and their obscene levels of funding have made everyone used to AIs that give results almost instantly, but smaller scrappier startups cannot afford to buy a large % of the world's GPU supply.
Skyboxes are used in 3D games. They are usually made of images plastered to the inside of a cube that is larger than the game area, with the player at the center. It makes the game area appear larger than it really is, because clouds, mountains and buildings appear far away.
As the founder of Ayvri which made 3D virtual worlds/digital twins of outdoor sports events and other scenic flythroughs, I love seeing this.
Part of me loves it because I recall telling VCs about 5 years ago that we would be increasing the resolution of our 3D worlds, and making more engaging by using AI to enhance the 3D world we had already built.
I got laughed at, and told that I didn't know what I was talking about. This was pre-facebook getting into the metaverse, and I was labelling it "spatial media".
Even with customers and revenue, nobody wanted to have anything to do with us.
We sold the technology and our patents last year, as the team had moved onto bigger and better things, but I love seeing how well this future is progressing.
This is very cool, which is why I'm so annoyed by the onboarding modals. Don't take away focus from what the user is currently messing with to show them a wall of text. Use tooltips.
It's titled "What's New At Blockade Labs", a title which continues to display even when you scroll down and find the terms of service below the news... which is presumably why the button is labeled "Confirm"