I guess there are other uses?? But this is just more abstracted reality. It will be innacurate just as summaried text is, and future peoples will again have no idea as to reality.
In fact you can already turn any photo into spatial content. I’m not sure if it’s using this algorithm or something else.
It’s nice to view holiday photos with spatial view … it feels like you’re there again. Same with looking at photos of deceased friends and family.
SHARP, an approach to photorealistic view synthesis from a single image - https://news.ycombinator.com/item?id=46284658 - Dec 2025 (108 comments)
Each assumes you already have their developer environment configured to have the tool work, but simply don’t have it compiled yet.
The “scenes” from that feature are especially good for use as lock screen backgrounds
Though in this particular case, you don't even need conda. You just need python 3.13 and a virtual environment. If you have uv installed, then it's even easier:
git clone https://github.com/apple/ml-sharp.git
cd ml-sharp
uv sync
uv run sharp"Exclusively for research purposes" so not actually open source.
I'm writing open desktop software that uses WorldLabs splats for consistent location filmmaking, and it's an awesome tool:
https://youtube.com/watch?v=iD999naQq9A
This next year is going to be about controlling a priori what your images and videos will look like before you generate them.
3D splats are going to be incredibly useful for film and graphics design. You can rotate the camera around and get predictable, consistent details.
We need more Gaussian models. I hope the Chinese AI companies start building them.
Everyone will be able to flex their muscles as a creative. Everyone will be able to become an artist (expressing themselves though their unique lens) without putting points into a mechanical skill that is dimensionally orthogonal to idea expression and communication.
This is the "bicycle of the mind" that Steve Jobs talked about 40 some years ago. We've all had keyboards with which to express ourselves and communicate, but soon everyone will be able to visually articulate themselves and their thoughts. It's going to be so uplifting for society.
In fifty years we'll even be able to render our direct thoughts and mold them like clay. Share them directly with one another. Co-think.
Any of the code that wraps the model or makes it useful is subject to copyright. But the weights themselves are as unrestricted as it gets.
Ok, but there's clearly more nuance there. Otherwise I could claim that any mp3 file I wanted to distribute is just a table of 8-bit integers and therefore not subject to copyright.
I don’t know where the catch is, but that sentence can not be true in general.
Weights are numbers that come from an optimization process. To the extent that weights encode any creativity, they encode the creativity of the training data. But any company using AI models (including Apple) does not want that interpretation because they are using AI models that were trained on other people's copyrighted works. If weights could be copyrighted, we all of us would own them.
You either a deliberately misrepresenting the facts or been livoning under a rock. I mean read any discussion about M laptops and you see apple fanboys noncritucally declaring them a revolution in computing.
I see a lot of people extolling the battery life, displays, and trackpads. And probably an equal amount of complaining about the increasingly locked-down and un-customisable nature of macOS. We all like the hardware, and fight the software more by the day.
The blind zealots of the "I'm a Mac, and I'm a PC" era just aren't very common anymore.
Yes, but the most important reason to pay attention to ANY license for most people is because it is a signal for under what conditions the licensor is likely to sue you (especially in the US, which does not have a general “loser pays” rule for lawsuits), not because of the actual legality, because a lawsuit is a cost most people don’t want to bear while it is ongoing or cover the unrecoverable costs of once it is done, irrespective of winning and losing, and, on the other hand, few people care about being technically legal with their use of copyright protected material if there is no perceived risk of enforcement.
But even if that wasn’t true, and being sued was of no financial or other costs until the case is finally resolved, and only then if you lose, I wouldn't bet much, in the US, in the court system ultimately applying precedent in the most obvious way instead of twisting things in a way which serves the interest of the particular powerful corporate interests involved here.
I know this is a long, nuanced, ongoing discussion. I'm very interested in it, but haven't read up on it for years. Could you elaborate a bit on the latest?
I was always in the camp that opined that "weights" are too broad a term for any sensible discussions about conclusions like "are (not) copyrightable". Clearly a weight that's the average of its training data is not copyrightable. But also, surely, weights that are capable of verbatim reproduction of non-trivial copyrightable training data are, because they're just a strange storage medium for the copyright data.
What am I missing?
Open source does not require full working implementations. There's no requirement that a code snippet that I release be fully working and identical to a complete solution.
About the access of binaries or providing working implementations, where did those come from? I don't think this thread was discussing those at all.
Indeed I would be willing to call something an "open source model" if it came without weights, but did come with the training data and with a documented process (preferably executable); and a release with just the training data could be called "open dataset" while the software to run the training would be just plain old open source software.
And, of course, a model with only the model data distributed with an open license is relatively commonly called "open weights", this being pretty self-explanatory term.
You already have access to all the training data everyone else is using.... You can download an offline version of Wikipedia. Here's every Reddit comment for a decade: https://academictorrents.com/details/ba051999301b109eab37d16...
Though, I do think it's still acceptable if you just point how to get the data (i.e. if it was the offline version of Wikipedia and then URL to that) if actually providing the source data is overwhelming. Offering to provide a copy at cost would be quite acceptable (i.e. I deliver the media to you to make a copy).
But if there's no way another person can acquire that data, even in theory, then I think it's pretty clear the source was not open. Just use the more appropriate term and everyone is on the level what the release is about.
The only reference seems to be in the acknowledgement, saying that this builds ontop of open source software
https://github.com/apple/ml-sharp/blob/main/LICENSE
Between this and the model's license, it seems like one is stuck with using this for personal use?
[1] https://github.com/apple/ml-sharp/blob/main/LICENSE
[2] https://fedoraproject.org/wiki/Licensing/Apple_MIT_License
It effectively prevents the community from using Apple's solution, but gives the Chinese everything they need to duplicate the results and push their own version.
I expect a Hunyuan-branded version of this model in six months. Probably with lots of improvements.
I'm all for Chinese model takeover if this is how US tech giants treat AI. You can't horde the flames forever, US hyperscalers.
The DoD ought to be advocating for a strong domestic open source stance to ensure our ecosystem doesn't get washed away. AI czar David Sacks has this view, but I suppose it's been falling on deaf ears when the hyperscalers crowd out the conversation.
People slowly waking up to how daft and hypecycle misusing a term was all along has been amazing.
Though I'm sure they will shut their shop asap now that Nvidia basically bought them.
I know, I'm joking: Trump likes Nvidia, but maybe he'll bump the Chinese tax to 30% to approve this deal? In a way I hope he pulls something like that, to punish Huang for his boot shining manipulations.
#iwantRAM
Open Source =/= free or software, just readable
so it wasn't a new campaign, it is at best re-appropriating the term open source in the software community in a way communities outside of software have always been using it, in a way that predates software at all, exists in parallel to the software community, and continues to exist now
I have also never once heard anyone use the term FOSS outside of the written form.
So the opposite of what you said, I guess.
You also seem to be saying that the term "open source" existed before software did, so I feel compelled to ask: what do you think "source" stands for in "open source"?
people need to re-evaluate their relationship with open source instead of as a synonym for FOSS, because it clearly doesn't mean that regardless of the colloquialism
and FOSS has an adjective and noun for a reason, its older than the colloquialism
this is just a reversion to the mean
New movements like "fair source", which is a form of source available + free use + expiring copyright is the ideal license. It's effectively unlimited use for customers and users, but has a baked in "non-compete" preventing low effort third parties from coming in and eating the market of managed services established by the authors.
We need to kill open source purity. Especially if we want to erode the hyperscalers that aren't even giving away the magic parts or valuable parts of their kingdoms.
Open source purity is a socialist dream while we're still living under the Empire. And it prevents the formation of a salient that can punch through. It's a very low suboptima.
I don't see any reason why you would want fair source authors to go "OSI" open other than taking their revenue stream as your own. The license bakes in contingencies in case the authors go out of business to open the license up for community ownership. That's good enough.
If these businesses were OSI open, the businessss become unsustainable and impossible to scale into something formidable that could chip away at entirely closed hyperscalers.
They just ask you not to compete with them for a few years.
How is that any way comparable to AWS?
Perfect truly is the enemy of good.
In this case, perfect murders good and locks you in the dungeon of eternal bad so you can think endlessly about perfect. It also stabs any good that comes along while crying about perfect.
No Open Source license actually permits this - by definition of Open Source.
Also the notion that copyright "expires" is ludicrous - we only just saw work from the 1920s enter the public domain (and source is no different to that). Laundering via AI clearly does not count, either.
this is a reversion to the mean
Linux kernel and GCC are probably the only thing left they tolerate, and even then, it is less relevant in the cloud, with containers powered by type 1 hypervisors.
> “Research Purposes” means non-commercial scientific research and academic development activities, such as experimentation, analysis, testing conducted by You with the sole intent to advance scientific knowledge and research. “Research Purposes” does not include any commercial exploitation, product development or use in any commercial product or service.
I don't think any in this particular space (image-to-3d gaussian representation) are, but then this is the first model I’ve seen in that space at all.
If (which the courts seem to be pretty consistently finding) training models on copyright-protected works generally is fair use, though using models to produce works which would violate copyright if made by other means with reference to the source material is still a copyright violation, then training has no bearing on the legality of copying the models. (Even if it wasn't, then copying and using the models at all would violate the copyright of the original owners of the training material again and be illegal irrespective of the “license” offered by the model trainer.)
Morally? Well, pretty much the same dichotomy applies; if training the model isn't a violation of the source material's creators' rights, then the fact it was trained without permission has no bearing on the morality of using the model without the trainers permission, if it is a violation of the source material's creators' rights, then so is using the model irrespective of the trainer's “license”, as the trainer has no right to permit further use of the material they had no right to create.
The idea that the model is an intrusion on the rights of the creators of the materials used in training and that this makes use of the model more rather than less permissibly, legally or morally, takes some bizarre mental gymnastics.
https://github.com/TencentARC/StereoCrafter https://github.com/TencentARC/GeometryCrafter
I’d be keen too.
doesn't seem very accurate, no idea of the result with a photo of large scene, that could be useful for level designers
I managed to one-shot it by mixing in the mesh exporter from https://github.com/Tencent-Hunyuan/HunyuanWorld-Mirror but at that point you might as well use HWM, which is slower but much better suited to the level design use case.
Note that the results might not be as good as you expect, because this does not do any angled inpainting -- any deviation from the camera origin and your mesh will be either full of holes or warped (depending on how you handle triangle rejection) unless you layer on other techniques far outside the scope of this model.
And note that although HWM itself does support things like multi-image merging (which ml-sharp does not), in my testing it makes so many mistakes as to be close to useless today.
If you want something different that is designed for levels, check out Marble by World Labs.
As rare as Apple is to admit it, there is this mercurial thing called "competition" that haunts the free market. OpenCL would have had an excellent chance if Apple took it as seriously as Nvidia took CUDA. But they didn't, it was thrown over the fence and expected that everyone else would do the work. While Nvidia was shipping Linux and BSD-native CUDA drivers, Apple was just distributing loose specs and begging the OpenCL working group to stop rewarding their competitor. Not for a lack of funding or motivation, Apple lost because they were butthurt.
OpenCL was DOA the moment Apple stopped treating Nvidia as a proper threat. Everyone else in the industry supported CUDA and was fine with it.
Once you start using even a small fraction of their tech it instantly metastasises throughout the entire organisation because of lock in and "open standards" that weirdly only work with their own tech. If the MS tech creates a problem the solution is to pour more MS tech onto the festering wound.
You apparently have been so insulated from how actual companies have to deal with tech that you think your little forays using computers are what everything should be measured by. All you have is a developer and hobbyist point of view.
You know Apple releases/funds a lot of open source, right?
Projects like WebKit, LLVM/clang, or CUPS (the print drivers for all of Linux)...
"Less than a second" is not "instantly".
> (...) Now, if I tell someone: "You should come to dinner more punctually; you know it begins at one o'clock exactly"—is there really no question of exactness here? because it is possible to say: "Think of the determination of time in the laboratory or the observatory; there you see what 'exactness' means"? "Inexact" is really a reproach, and "exact" is praise. (...)
2. People who were born outside the United States but moved here to do research a while back don’t suddenly stop doing research here.
That said, the US only has some 5% of the worlds population (albeit probably a larger proportion of the literate population), so you'd only expect some fraction of the world's researchers to be US born. Not to mention that US born is an even smaller fraction of births (2.5-3%, by Google), so you'd expect an even smaller fraction of US born researchers. So even if we assume that we're on par with peer countries, you'd only expect US born researchers to be a fraction of the overall research population. We'd have to be vastly better at educating people to do otherwise, which is a longshot.
Obviously this makes turning away international students incredibly stupid, but what are we to do against stupidity?
I checked the first, middle, and last author: Lars Mescheder got his PhD in Germany, Bruno Lecouat got his PhD in France, Vladlen Koltun got his PhD in Israel.
(Edit: or maybe they did not actually immigrate but work remote and/or in Europe)
Approximately 96% of the world's population is not American, so you should expect that really.
It's because American education culture is trash. American parents are fine with their kids getting Bs and Cs. Mediocrity is rewarded and excellence is discouraged in our schools, both socially and institutionally.
Meanwhile you have hundreds of millions of foreign born children pulling out all the stops to do the best they possibly can at school precisely so they can get into the US and work at one of our top companies.
It was never even a competition. Immigrants and children of theirs will continue to outperform because it is literally baked into their culture - and it is baked out of ours.
Was somewhat surprised to learn that the pipeline wasn't built by industry demand, it was supply pressure from abroad that happened to arrive just as US universities needed the money (2009/10). In 1999, China's government massively expanded higher education, combined with a system where the state steers talent into stem via central quotas in the "gaokao", it created an overflow of CS capable graduates with nowhere to go domestically, India's 1991 liberalization created the IT services boom (TCS, Infosys, Y2K gold rush) and made engineering THE middle class ticket, so same overflow problem. US phd programs became the outlet for both countries.
In that light, the university side response probably wasn't state side industry demand for loads of PhDs, who was hiring those then? Google Brain didn't exist until 2011, FAIR until 2013. It wasn't really till 2012+ that industry in tech started to hire big research groups to actually advance the field vs specialized PhDs here and there for products... so not a huge amount of pull from there. Then, at the same time, universities were responding to a funding crisis... there was a 2008 state budget collapse, so it was backfilled with international Master's students paying $50-80k cash (we do this in Canada heavily also), that revenue cross-subsidized PhD programs (which are mostly cost centers remember). I also read some say PhD students were also better labor: visa constraints meant they couldn't easily bounce to industry, they'd accept $30k stipends, tho I saw other research contradicting this idea. The whole system was in place before "AI Researcher" was even a real hiring category. Then deep learning hit (2012), industry woke up, and they found a pre built pipeline to harvest: The authors on that Apple paper finished their PhDs around 2012-2020, meaning they entered programs 2009-2015 when CS PhDs were already 55-60% foreign born. Those students stayed, 75-85% of Chinese and Indian STEM PhDs are still here a decade later. They're now the senior researchers publishing papers you read here on HN.
This got me wondering, could the US have grown this domestically? In 2024 they produced ~3,000 CS PhDs, only ~1,100 domestic. To get 3,000 domestic you'd need 2.7x the pipeline...which traces back to needing 10.8 million 9th graders in 2018 instead of 4 million (lol), or convincing 3x more CS undergrads to take $35k stipends instead of $150k industry jobs. Neither happened. So other countries pay for K-12 and undergrad, capture the talent at PhD entry, keep 75%+ permanently.
Seems like a reasonable system emerged from a bunch of difficult constraints?
(and just to reiterate, even tho it was an interesting research project for me, you can't infer where someone is directly from based on their name)
https://sccei.fsi.stanford.edu/china-briefs/highest-exam-how...
https://en.wikipedia.org/wiki/Economic_liberalisation_in_Ind...
https://ncses.nsf.gov/pubs/nsf24300/data-tables
https://www.aau.edu/newsroom/leading-research-universities-r...
https://ncses.nsf.gov/pubs/nsf25325
https://www.science.org/content/article/flood-chinese-gradua...
https://www.insidehighered.com/quicktakes/2017/10/11/foreign...
The US is the largest research hub in the world, and it offers (or offered) outstanding conditions for research. I believe this to be as old as WW2, and it certainly didn't start with AI. Higher salaries, more diverse career opportunities (academia is more porous to industry in the US than many other countries), and the ability to hire more and better candidates for the workhorses of a lab: PhD students, postdocs, technicians, research scientists.
Re: supply side, undergraduate education (including Master's in some countries) has become basic infrastructure in a developed (or developing) country, and countries like China, ex-USSR or the western European nations have solid traditions in this regard, with many offering comparable (or surpassing) education to the best US universities in specific STEM topics. However, save for China, I believe a majority of these countries have not invested in research to match their growing pool of Master's (or even PhD) graduates.
I'm not kidding. That's going to be >80% of the images/videos synthesized with this.
The output is not automatically metrically scaled (though you can use postprocessing to fix this, it's not part of this model). And you can't really move around much without getting glitches, because it only inferences in one axis. It's also hard capped at 768 pixels + 2 layers.
Besides depth/splatting models have been around for quite a while before this. The main thing this model innovates on is inference speed, but VR porn isn't a use case that really benefits from faster image/video processing, especially since it's still not realtime.
This year has seen a lot of innovation in this space, but it's coming from other image editing and video models.
So far the best looking results are still achieved with good old mesh warping and no inpainting at all. This may change that.
> So far the best looking results are still achieved with good old mesh warping and no inpainting at all.
I agree
> This may change that.
Seems not to be the case in my testing. The splats are too fine and sparse to yield an improvement. There are actually better (slower) image -> splat models than ml-sharp (with much higher dynamic range for the covariance) but I still don't use them over meshes for this.
The only improvements ml-sharp seems to add to the SOTA is 1) speed and 2) an interesting 2-focal layer architecture, but these are somewhat tangential steps.
Additionally, we might need better categories. With software, flow is clear (source, build and binary) but with AI/ML, the actual source is an unshippable mix of data, infra and time, and weights can be both product and artifacts.
There has to be an easier combination of words for conveying the same thing.
neom•1mo ago
Paper: https://arxiv.org/abs/2512.10685
the8472•1mo ago