p.s. This was lobbed onto the frontpage by the second-chance pool (https://news.ycombinator.com/item?id=26998308) and I need to make sure we don't end up with duplicate threads that way.
Released last week. Looks like all the weights are now out and published. Don’t sleep on the SAM 3D series — it’s seriously impressive. They have a human pose model which actually rigs and keeps multiple humans in a scene with objects, all from one 2D photo (!), and their straight object 3D model is by far the best I’ve played with - it got a really very good lamp with translucency and woven gems in usable shape in under 15 seconds.
Are those the actual wireframes they're showing in the demos on that page? As in, do the produced models have "normal" topology? Or are they still just kinda blobby with a ton of polygons
I’ve only used the playground. But I think they are actual meshes - they don’t have any of the weird splat noise at the edge of the objects, and they do not seem to show similar lighting artifacts to a typical splat rendering.
I haven’t tried it myself, but if you’re asking specifically about the human models, the article says they’re not generating raw meshes from scratch. They extract the skeleton, shape, and pose from the input and feed that into their HMR system [0], which is a parametric human model with clean topology.
So the human results should have a clean mesh. But that’s separate from whatever pipeline they use for non-human objects.
For the objects I believe they're displaying Gaussian splats in the demo, but the model itself can also produce a proper mesh. The human poses are meshes (it's posing and adjusting a pre-defined parametric model).
Side question: what are the current top goto open models for image captioning and building image embeddings dbs, with somewhat reasonable hardware requirements?
For pure image embedding, I find DINOv3 to be quite good. For multimodal embedding, maybe RzenEmbed. For captioning I would use a regular multimodal LLM, Qwen 3 or Gemma 3 or something, if your compute budget allows.
I would suggest YOLO. Depending on your domain, you might also finetune these models. Its relativly easy as they are not big LLMs but either image classification or bounding boxes.
For a long time I've wanted to use something like this to remove advertisements from hockey games.The moving ads on the boards are really annoying. Maybe I'll get around to actually doing that one of these days.
I wonder if this can be used to track an object's speed. E.g. a vehicle on a road. It would need to recognize shapes, e.g. car model or average size of a bike, to guess a speed.
Alternative downloads exist. You can find torrents, and match checksums against the HF downloads, but there are also mirrors and clones right there in HF which you can download without even having to log in.
This would be convenient for post-production and editing of video, e.g. to aid colour grading in Davinci Resolve. Currently a lot of manual labour goes into tracking and hand-masking in grading.
I do a test on multimodal LLMs where I show them a dog with 5 legs, and ask them to count how many legs the dog has. So far none of them can do it. They all say "4 legs".
Segment anything however was able to segment all 5 dog legs when prompted to. Which means that meta is doing something else under the hood here, and may lend itself to a very powerful future LLM.
Right now some of the biggest complaints people have with LLMs stems from their incompetence processing visual data. Maybe meta is onto something here.
This was front page for a while last week
https://news.ycombinator.com/item?id=45982073
Thanks! Macroexpanded:
Meta Segment Anything Model 3 - https://news.ycombinator.com/item?id=45982073 - Nov 2025 (133 comments)
p.s. This was lobbed onto the frontpage by the second-chance pool (https://news.ycombinator.com/item?id=26998308) and I need to make sure we don't end up with duplicate threads that way.
what is old is new again
Released last week. Looks like all the weights are now out and published. Don’t sleep on the SAM 3D series — it’s seriously impressive. They have a human pose model which actually rigs and keeps multiple humans in a scene with objects, all from one 2D photo (!), and their straight object 3D model is by far the best I’ve played with - it got a really very good lamp with translucency and woven gems in usable shape in under 15 seconds.
Between this and DINOv3, Meta is doing a lot for the SOTA even if Llama 4 came up short compared to the Chinese models.
https://ai.meta.com/blog/sam-3d/ for those interested.
Are those the actual wireframes they're showing in the demos on that page? As in, do the produced models have "normal" topology? Or are they still just kinda blobby with a ton of polygons
I’ve only used the playground. But I think they are actual meshes - they don’t have any of the weird splat noise at the edge of the objects, and they do not seem to show similar lighting artifacts to a typical splat rendering.
I haven’t tried it myself, but if you’re asking specifically about the human models, the article says they’re not generating raw meshes from scratch. They extract the skeleton, shape, and pose from the input and feed that into their HMR system [0], which is a parametric human model with clean topology.
So the human results should have a clean mesh. But that’s separate from whatever pipeline they use for non-human objects.
[0]: https://github.com/facebookresearch/MHR
For the objects I believe they're displaying Gaussian splats in the demo, but the model itself can also produce a proper mesh. The human poses are meshes (it's posing and adjusting a pre-defined parametric model).
you can download them at https://github.com/facebookresearch/sam3. for 3d https://github.com/facebookresearch/sam-3d-objects
I looked quickly but it does not generate a 3d model file right?
Surprisingly, SAM3 works bad on engineering drawings while SAM2 kinda works, and VLMs like Qwen3-VL works as well
Had good luck with Gemini 2.5, SAM3 failed miserably with PIDs.
yeah I tried too. Im trying a fine tuning on PIDs.
Side question: what are the current top goto open models for image captioning and building image embeddings dbs, with somewhat reasonable hardware requirements?
For pure image embedding, I find DINOv3 to be quite good. For multimodal embedding, maybe RzenEmbed. For captioning I would use a regular multimodal LLM, Qwen 3 or Gemma 3 or something, if your compute budget allows.
Try any of the qwen3-vl models. They have 8, 4 and 2B models in this family.
I would suggest YOLO. Depending on your domain, you might also finetune these models. Its relativly easy as they are not big LLMs but either image classification or bounding boxes.
I would recommend bounding boxes.
What do you mean "bounding boxes"? They were talking about captions and embeddings, so a vision language model is required.
I suggested YOLO and non llm-vl as a lot faster alternative.
Of course CLIP would be otherwise the other option than a big llm-vl one.
Which YOLO?
Any current one. they are easy to use and you can just benchmark them yourself.
I'm using small and medum.
Also the code for using it is very short and easy to use. You can also use ChatGPT to generate small exepriments to see what fits your case better
There aren’t any YOLO models for captioning and the other models aren’t robust enough to make for good embedding models.
You can get labels out of the classifier and bounding box models.
They are super fast.
Its just an alternative i'm mentioning. I would assume a person knowing a little bit of that domain.
Otherwise the first option would be CLIP i assume. llm-vl is just super slow and compute intensive.
Which (if any) of these models could run on a RaspberryPi for object recognition at several FPS?
For a long time I've wanted to use something like this to remove advertisements from hockey games.The moving ads on the boards are really annoying. Maybe I'll get around to actually doing that one of these days.
I wonder how effective this is medical scenarios? Segmenting organs and tumors in cat scans or MRIs?
I wonder if this can be used to track an object's speed. E.g. a vehicle on a road. It would need to recognize shapes, e.g. car model or average size of a bike, to guess a speed.
Been waiting days to get approval to download this from huggingface. What's up with that?
Alternative downloads exist. You can find torrents, and match checksums against the HF downloads, but there are also mirrors and clones right there in HF which you can download without even having to log in.
I was approved within about 10 minutes for "Segment Anything 3"
same here, didn't get approval
Miss the old segment anything page, used it a lot. This UI I found very complex to use
Same.
Checkout https://github.com/MiscellaneousStuff/meta-sam-demo
It's a rip of the previous sam playground. I use it for a bunch of things.
Sam 3 is incredible. I'm surprised it's not getting more attention.
> I'm surprised it's not getting more attention.
Remember, it's not the idea, it's the marketing!
This would be convenient for post-production and editing of video, e.g. to aid colour grading in Davinci Resolve. Currently a lot of manual labour goes into tracking and hand-masking in grading.
[dead]
I do a test on multimodal LLMs where I show them a dog with 5 legs, and ask them to count how many legs the dog has. So far none of them can do it. They all say "4 legs".
Segment anything however was able to segment all 5 dog legs when prompted to. Which means that meta is doing something else under the hood here, and may lend itself to a very powerful future LLM.
Right now some of the biggest complaints people have with LLMs stems from their incompetence processing visual data. Maybe meta is onto something here.
Segmentation doesn't need to count legs. I'd guess something like YOLO could segment 5 legged dogs too.
YOLO is not a segmentation model.
https://docs.ultralytics.com/tasks/segment/
Thanks! TIL there's a class of segmentation models with the YOLO naming scheme.
I thought it was a joke about YAML
Lol you obviously haven't seen what cheats for FPS games look like in the last 3 years.
https://github.com/Babyhamsta/Aimmy
You don’t need segmentation to count legs. Object detection can do that. DeepLabCut from 2020 perhaps.
I doubt that gemini 3 cannot do it.