TechieMoon/sam3-libero10-procedural-segmentation

SAM3 LIBERO-10 Procedural Segmentation

This is a SAM3 image segmentation checkpoint fine-tuned for LIBERO-10 object naming and visible-mask segmentation prompts.

The training data used for this checkpoint was generated locally as procedural render-style 2D images with segmentation masks. No LIBERO 3D assets, third-party mesh files, or rendered training images are included in this repository.

Training Summary

• Base architecture: SAM3 image model with segmentation enabled
• Fine-tuning date: 2026-05-22
• Training images: 540
• Validation images: 108
• Training annotations: 1700
• Validation annotations: 338
• Categories: 18 LIBERO-10-related object/environment names
• Trainable parameters: about 32.7M
• Frozen parameters: about 807M backbone parameters
• Epochs: 1
• Steps: 540
• Batch size: 1
• GPU used: RTX 3090 24GB
• Final training core loss: 146.4154

Categories

• white mug
• yellow and white mug
• moka pot
• chocolate pudding cup
• alphabet soup can
• cream cheese box
• tomato sauce can
• butter box
• black bowl
• book
• plate
• left plate
• right plate
• basket
• microwave
• stove
• cabinet drawer
• caddy

Loading

Install the SAM3 codebase and its dependencies first. Then:

from load_model import load_model

model = load_model(
    repo_id="TechieMoon/sam3-libero10-procedural-segmentation",
    device="cuda",
)
model.eval()

The checkpoint file is model.safetensors. It contains the full SAM3 image model state dict, not just a delta, so the loader builds the SAM3 image architecture and then loads this state directly.

Files

• model.safetensors: fine-tuned SAM3 image model state dict
• load_model.py: helper for loading the safetensors checkpoint into SAM3
• training_config.yaml: resolved training config
• source_training_config.yaml: compact source Hydra config
• training_manifest.json: training data counts, category names, and asset policy
• LICENSE: SAM license copied from the upstream SAM3 model repository

Limitations

This checkpoint is a practical starting point, not a fully validated production segmentation model. The fine-tuning data is synthetic/procedural, so there can be a domain gap on real LIBERO frames, especially under heavy Franka occlusion or unusual lighting/viewpoints. It should be tested on the target LIBERO-10 frame distribution before being used as a reliable automatic annotator.

This is not an official Meta checkpoint.

License And Use

This repository does not redistribute the procedural training images or any 3D assets. Use of this checkpoint should follow the license and use terms of the underlying SAM3 model and any SAM3 code/checkpoints used with it.