Founder @dotproduct3d Prev. Founder @ Enkin AR (acq‘d by Google) SLAM / Sensor Fusion / 3D Capture GPU poor
Joined June 2013
- Tweets4.5K
- Following1.4K
- Followers2K
- Likes84.8K
Pinned Tweet
Dear friends, dear internet, dear enjoyers of splats,
today I'm releasing not just one, but two projects to the world that I'm extremely proud of:
The first is Frankensplat:
an all-new, browser-based, free-to-use, superfast 3DGS training studio built entirely in Javascript and WebGPU!
Frankensplat is a complete BEAST of an engine.
In fact, it's the fastest 3DGS training engine that we know!
Despite using "only" WebGPU and JS in the browser it outperforms even the fastest native apps while being SOTA on quality (study link in reply).
Since it runs natively in your browser, it works
* offline (data never leaves your device)
* on virtually all modern computers
* without any download, setup, build or install
And since it's so fast you can enjoy training great Gaussian Splats even on older or cheap consumer GPUs. No 4090 required!
If you like 3DGS, go give it a try!
>>> Links in reply <<<
-----
The second project is the Frankensplat launch video.
This is my first time getting really serious with AI for video production.
It's only 3:44 minutes long but *a lot of work* went into it.
Every single clip in the video was prompted as a still frame (many, many times!), then style-adjusted using img2img, then animated using different video models until it was right.
Final editing, mixing, color grading and rendering was done in Davinci Resolve.
In the end, it really did come out exactly the way I imagined it, which is a great feeling and which I didn't assume was possible when I started it.
If you have four minutes, I think it'll be worth your time.
PS: the 3DGS jokes start in the second half of the video, so be sure to watch till the end!
PPS: due respect to @janusch_patas who helped bring 3DGS to a broader audience with LFS and has helped the community in countless ways. Janusch, thanks for starring in the video -- don't take it too seriously!
Links:
Frankensplat: offlinetools.io/frankensplat
Product page: offlinetools.io/fsp-product
Benchmark Study: offlinetools.io/fsp-study
1
103
The only person on the planet that's entitled to a judgement about Jensen right now is is wife Lori.
Everybody else please just stfu.
136
This workflow is becoming soooo good :)
Why draw at your desk when you can draw onto reality? (and trust it!) 🎯
Simple (yet precise!) mapping of a hotel's architectural footprint with:
1. Point and shoot capture with #Dot3D
2. Automatic geo-control with @bad_elf
3. Custom deliverables with @OpenAI
Now let’s break it down for you:
1. Place 1 AprilTag for loop closure
2. Establish RTK fix and start scan
3. Take scan-time annotations at all “draw points” (corners)
4. Finish scan back at the AprilTag
5. Review results (manually add/edit annotations only if necessary)
6. Export JSON annotation file straight from Dot3D (no need to “complete annotations”)
7. Optional: Export geoTIFF
8. Ask ChatGPT to produce: DXF drawing, PDF report, GeoPackage, etc. (whatever you need!)
9. Remember to include any custom requests (i.e. flattening z coordinates to zero)
10. Bring it all straight into CAD, GIS, or just move on with the completed PDF!
Scan time: 7 minutes
Optimization time: 40 seconds
AI processing time/costs: Negligible
5
337
Rafael Spring retweeted
【全世界がこの瞬間に感動した？】
#dot3d と #drogger 最新状況
dot3dがdroggerのGパッケージに対応
希望者またはPRO版の方は30日限定のテストライセンス番号を交付して貰えるので、まずは公式サイトから申し込みしてください
※PRIME版は1年間$999てすが、月額$139の契約もあるので、まずは必要な時に月額サブスク契約するのもやり方もありですね
@DotProduct3D
#dot3d 更新キター❗️
まだBETA版ですが #drogger を機種選択できません😭
PRIME版に契約しないいけないの？
PRO版で動かないと動作テストも出来ないよ😵
最初から難題が起きて躓いてしまった…🤦
どうすれば試用できるのでしょうか？
@DotProduct3D
I have a weird feeling "vibe coding" will go the same way as "prompt engineering".
Will anyone even be interacting with code in 2028? I'm not so sure actually.
3
1
400
My fellow Gaussian Splat enjoyers,
I created a monster.
I'm about to set it free this week.
Be prepared.
Rafael Spring retweeted
Y Combinator just published the list of startups it wants to fund next winter. One of them is AI for City Hall. In some American cities, YC writes, getting a building permit takes longer than building the house. By its own math, if every permit in the country were approved in a tenth of the time, cities would collect $16 billion more a year in property tax. There are 19,000 cities in America, and almost all of them still check drawings against the code by hand.
I like the idea. My first company, while I was still at Columbia in the 1980s, redeveloped buildings in Tribeca and SoHo, so I have stood on the other side of that counter.
But look at what the waiting actually costs. In 2021 the National Association of Home Builders surveyed more than 2,000 builders and found that regulation accounts for $93,870 of the price of an average new home, 23.8% of it. Only about $2,400 of that is the pure cost of delay. The rest is fees, building codes that keep changing, design standards beyond the ordinary, and land the government takes or forces you to leave empty.
So AI can make City Hall faster. It cannot make it cheaper. A model that approves a bad rulebook in one day still leaves the same $94,000 inside the average new house, just sooner.
The startup I would fund does both: it reads every permit, and it also measures what each rule costs the family that will buy the house. Then it publishes the numbers and lets voters decide which rules are worth paying for.
Faster bureaucracy is progress. Less bureaucracy is reform.
Rafael Spring retweeted
Can it really be that easy?
Can it really be that fast?
Actually yes.
The real numbers:
Prep: 0 min
Scanning: 13 min
Optimization: 3.5 min
Software: Dot3D Prime
Hardware: iPhone Pro Max + General Laser GLRM RTK
Constraints: 1 AprilTag for loop closure, MaCORS RTK corrections
my github is pure white*
*I don't even have an account
“You look at their GitHub and they’re not dark green.”
@saradu co-founded Alloy Automation and now runs Ando solo, with $20M from Accel, Index and Emergence. She doesn’t think most of your competitors are real.
Every market, in her view, has only a few serious players. The rest talk about the mission and then drift toward quick money.
“A lot of solo founders should take solace in that.”
1
353
Rafael Spring retweeted
@Rafael_L_Spring COLMAP with no install. Drop images or a video into this unofficial browser port and watch keypoints, matches, camera poses and a sparse point cloud appear. Nothing leaves your computer. We made a 30s intro video for it.
offlinetools.io/colmap-landi…
OK, launch day!!
Over the past few weeks I built a fully browser-resident implementation of COLMAP, using WASM and WebGPU (-- link in reply --).
No uploads, no installs, no setup. Just drop images or video and run the pipeline. Then export and drop results into a 3DGS tool of your choice.
The first naive implementation was about 10x slower than native but with many optimizations it's now faster than the original!
Go give it a try!
You know what? I'm here for it.
Make them cost 50k, 100k, the higher the better.
It will make the market value efficient code and punish sloppy code again.
Y'all know we're overdue for this.
Don’t shoot the messenger: Why would Nvidia sell any RTX 5090 at <$5K when they can sell the same die as RTX 6000 Pro for $16K? It has massive demand. Today every RTX 5090 they sell is a charity project. Capitalism basics = sell for highest bidder.
3
419
My razor confirmed again.
If a person or organization aggressively wants you to believe they are X, they're almost certainly the exact opposite of X.
Look at this horror show of a UI called "ComfyUI".
Plus their docs are several gens behind and nothing applies anymore.
Avoid.
1
6
842
How much edge compute is needed you ask?
If you hard code your physics, control systems are surprisingly cheap.
Inputs are a bunch of IMUs + altitude.
At a rate of 400 / sec, that's ~ 64 KB / sec of input processing.
Output is next to nothing.
Could run on a Raspberry Pi easy.
Hate to dox myself a bit but have worked on just about every fighter jet the US has. I have seen some impressive technology over my career but this is a whole other level. I have so many questions and need to know how this works. How much edge compute is needed to make this happen? My first thought is dual processor + 1TB ram. That way it can handle the fuckquintillion process per second. There was zero correction it nailed it 100%.
3
28
5,847