Kieran Farr

@sparkjsdev Working great for large area 3D scans! Here's a custom plaza scan using Spark 2.0 LoD streaming combined with Google 3d tile maps and custom models. It takes only 5 seconds (including app download) for this 2.5M splat to start being visible to the end-user, great work!

In Spark 0.1, world scale was constrained by real device limits: real-time splat counts, GPU memory ceilings, multi-GB file loading failures in the browser, and precision artifacts at large coordinates. Spark 2.0 reworks the architecture around LoD trees and streaming. Any splat file can be converted into a hierarchical LoD representation, with fixed, tunable splat budgets per frame to guarantee steady performance across devices. Scenes are streamed coarse-to-fine using a new .RAD format and managed by a shared LRU "splat pager" that keeps GPU memory bounded. We also added higher-precision ExtSplats encoding and ReadableStream support for multi-GB files. Branch: github.com/sparkjsdev/spa…

Spark 2.0 Developer Preview is now available: introducing Level-of-Detail (LoD) rendering and streaming for world-scale Gaussian Splats on the web. Learn more ↓

In 2026 it’s still hard for new users to understand the quest vr controllers. Fine for gamers, but the number of times I have to instruct people how to press the trigger button is insane. I can understand why Apple decided on hands only.

@bai0 @ID_AA_Carmack Yeah same experience, I left, came back, and 0 engagement with posts. Compared to TikTok where I just made a new account and content gets hundreds of likes and thousands of views almost instantly. No joke!

@ID_AA_Carmack The exodus was real, but for those of us that stayed the new algorithm just does not favor the little guy. I post things that used to get a few hundred likes, now it's a ghost town My expectation that you or anyone who follows me will even see this reply is nil. I post anyway 🤷‍♂️

It would be nice if some of the Twitter diaspora returned. So many creatives, but also many developers, that generally enriched the experience are no longer active. Those that performatively left and those with a seething hatred of Elon probably won’t be back soon, but a lot of people just disengaged on vague cultural grounds that can be reevaluated. There are probably some technical tweaks to the algorithm that could make them more comfortable. I don’t mind the existence of independent echo chambers that people are happy within. There is only a problem when some echo chambers are allowed and others aren’t. Reach out to lapsed friends!

@1littlecoder @3dstreetapp @bfl_ml @SFMTA_Muni @fal Thanks! Sent over a request to the @fal grants team to see if it's possible to get some credits for training on our larger data set 😉 happy to open source the weights

@kfarr @3dstreetapp @bfl_ml @SFMTA_Muni @fal this is so cool!

Now available: Convert CAD road striping drawings into aerial views with our AI Street Generator tool (from @3dstreetapp) and our new custom @bfl_ml LoRA we trained with public data from @SFMTA_Muni and aerial images via @fal API.

Try it out here: 3dstreet.app/generator/ from striping drawings sourced from here: sfmta.com/Striping

@_MattHuff @bfl_ml @fal @SFMTA_Muni Correct just images not vectors. There’d need to be a conversion process but that’s do able these days if the image is clean enough. Once training is done inference is cheap

@kfarr @bfl_ml @fal @SFMTA_Muni Interesting, did not realize training time was still this expensive. Just so I understand completely, existing model just uses images of CAD files correct? No actual reading of CAD vector files?

Upload street marking CAD drawings and generate realistic aerial renderings! Demo of a small LoRA fine-tuning of @bfl_ml Flux 2 Dev via @fal from @SFMTA_Muni publicly available striping drawings into aerial satellite map style perspectives.

@_MattHuff @bfl_ml @fal @SFMTA_Muni Yes it is possible, and we have the data, but honestly it's not cheap to get the GPU time to train. Perhaps @fal can help with some credits to do more fine tuning on a larger dataset

@kfarr @bfl_ml @fal @SFMTA_Muni I’m interested in the reverse for rough design purposes! Do you think a model could be trained on aerial images and generate CAD basemaps?

- Daily videos stretched concepts in my mind that I have either done quickly or imagined could be possible with the software, and now I not only proved it but documented the use.

- A cycle of 1 video per day is too much for me at this phase in life. It didn't leave enough time for finishing projects -- there's a lot of testing, bugfixing, docs writing and editing, etc. before deploying many of the projects featured in the last 31 days.

Day 31/31 of sharing what I'm up to this month, meta edition. Every day this month I posted something across LinkedIn, Twitter, TikTok, YouTube, Instagram, and Facebook. My learnings 🧵

- Quick personal or off-topic posts are often just as engaged as "work" ones. - Daily videos provided motivation to finish some functional features quickly to hit deadlines, but this came at the expense of actually finishing many of them.

- Family zoning visualizations were the most popular subject on all platforms! Showing the future at the neighborhood scale is triggering (in good & bad ways) - Bollard Buddy iOS app is popular on TikTok. Bollard Buddy content is also filmed in portrait mode, TikTok seems to like

30/31: There are a lot of ways to show off cutting edge technology like state of the art Gaussian splat 3D scanning, compositing, visualization, and AI rendering technology. Here are some cones in a plaza.

Some great renderings of upzoning in SF Look how inoffensive this is…this ought to be table stakes

29/31: The holidays are a slow time for @3dstreetapp. Even so, in the past 72 hours over 100 scenes were updated by users. Here are 2 I liked and re-rendered...

28/31: Using the Terrain Flattening tool in @3dstreetapp to visualize an underground sewer replacement project in San Francisco's Mission District.