wanx-troopers
Segmentation
Sam3
Kijai: “I’m just about to finish native SAM 3.1 support”
Community is experimenting with SAM3 model from Facebook. The model provides several capabilities
- segmentation
- depth maps
Sam3D praised as new in segmentation models.
Several ComfyUI implementations:
- “we have that in core too”
- GH:Ltamann/ComfyUI-TBG-SAM3
- GH:PozzettiAndrea/ComfyUI-SAM3
- GH:PozzettiAndrea/ComfyUI-SAM3DBody “They don’t release VRAM quite as well as you’d expect. They work, but lack optimization”
Sentiment: “crazy”, “Sec4B still doing better”
- David Show’s workflow for Sam3 masking: ds-sam3d-Mask-Generator-multi
SAM2 can arguably be better with points; 3 (and 3.1) is all about the text prompting and video tracking
LocateAnything
GH:alisson-anjos/ComfyUI-LocateAnything based on research.nvidia.com/labs/lpr/locate-anything
Depth Anything
Anyone know why depthanything V3 looks like unadulterated garbage compared to V2?
Sam3.1 And Sam3D
Preview of future nodes to come utilizing Sam3.1 for segmentation and Sam3D for mesh reconstruction: sam3d-wip; sam3d-object.
sam3d-object … it will be able to do whole scene too and then sam3d-body plugs into the scene
Q: it does do fingers tho?
A: sam3d does, but kimodo doesn’t; sam3d can even do ASL [American Sign Language], it’s really good; … did even add HAMER to refine the hands, but I’m not sure it’s … necessary anymore
Sam3 Native Model Loader node has been spotted in the wild.
SAM3 Video Track Node, SAM3 Tracks to Mask, SAM3 Trackes Preview, SAM3 Detect
Sam3 masking wf from djbfilmz: 
cant we use sam 3 insted of sam 3.1?
why? they perform about the same, 3.1 better with multiple people
people:2 may be required to segment out two characters
ucren:
sam assigns the colors based on what you pick in the drop down, by default left to right. but if you want to mask three separate ids you need to have them in a single image so sam detects them as three distinct ids