Marc Mapeke

3d vision research (big + small models) @realitylabs. @brownvisualcomp alum.

tweet = my personal opinions.

creativity, learning, perception, and action ≥≥≥

[Marc Mapeke @marcmapeke]

@jonstephens85 sounds fun!

[Stefan Baumann @StefanABaumann]

The internet is full of video. So why can't novel view synthesis just scale on it?

Real-world video is simultaneously unposed, messy, and dynamic, breaking self-supervised NVS.

We fixed that. RayDer learns static-scene NVS from dynamic internet video, scaling like an LLM. A🧵

[Marc Mapeke @marcmapeke]

@pablovelagomez1 @Nik__V__ I’ll keep this in mind for future releases. I can see efficient models + multitask being helpful for open source releases, even if not SOTA on all tasks.

[Marc Mapeke @marcmapeke]

Depth and surface normal models are slow at high resolutions.

HyDen (ICLR 2026) is a dual-path CNN+ViT recipe that is 4–10× faster and more accurate.

Thread + code below.

[Marc Mapeke @marcmapeke]

@pablovelagomez1 @Nik__V__ These are 3 separate models. We mainly want to show Hyden works independently on 3 tasks across public SOTA checkpoints.

[Marc Mapeke @marcmapeke]

This is a simple recipe for extending base models to high-resolution dense prediction.

We integrate HyDen encoders into 3 tasks/models:
• relative depth
• metric point maps
• surface normals

HyDen is also faster than recent fast depth models like DepthPro and FlashDepth.

We think the recipe can generalize to other pixel-heavy tasks as well.

[Marc Mapeke @marcmapeke]

Code, Model, and Instructions: https://t.co/xx3qKQKs0E

[Saksham Suri @_sakshams_]

Introducing Efficient Universal Perception Encoder (EUPE)🚀

A family of compact vision encoders that match or exceed domain experts across diverse tasks, in a single model.

📄 Paper: https://t.co/u4DtnmQISr
💻 Code: https://t.co/NaSt0aTOGN
🤗 Models: https://t.co/BUumD6LhCe
🧵

[Marc Mapeke @marcmapeke]

@natolambert why do you think open frontier and open small will be separate?

[Marc Mapeke @marcmapeke]

@yacineMTB did you pick those sensors because of the weight limit?

[Marc Mapeke @marcmapeke]

@gabriberton @chris_j_paxton how about if it only used for distillation or supervision signals? we don’t know the details still

[Marc Mapeke @marcmapeke]

@gabriberton @chris_j_paxton we don’t know though!! E2E only means that the big model is E2E.

it is very possible that they have a standalone frozen detection model AND that model is only used to bootstrap the intermediate prediction and visualization of E2E driving.

[Marc Mapeke @marcmapeke]

@gabriberton @chris_j_paxton this recent figure is from Tesla, October 2025 - https://t.co/tz33SL3ySU

[Ashok Elluswamy @aelluswamy]

https://t.co/Jr4i7tHUCr

[Marc Mapeke @marcmapeke]

@gabriberton @chris_j_paxton E2E models does not mean 0 intermediate outputs though

[Marc Mapeke @marcmapeke]

@yacineMTB in a recent podcast with the Matic ceo, he mentioned the hidden major cost of adding sensor types to hardware is the amount of team bandwidth and headcount required to calibrate, simulate, and iterate with the sensor

[Marc Mapeke @marcmapeke]

@DrCatharineY the higher-order effects of this over the coming years will be interesting to see

[Marc Mapeke @marcmapeke]

@sainingxie always run the experiments!

[Marc Mapeke @marcmapeke]

@Yuchenj_UW higher leverage for OpenAI**

this is definitely not true in general tbh

[Marc Mapeke @marcmapeke]

@Yuchenj_UW I feel like this ordering is because of leverage NOT abilities. Researchers are higher leverage (and harder to hire) than engineers.

The actual act of engineering/experimentation/implementation is easier to verify (and train big model for) then research ideas imho.