1kaiser/moge3-litert

MoGe-3 LiteRT & TFLite Models

Quantized and edge-optimized LiteRT / TFLite models for MoGe-3: Fine-Detail Monocular Geometry Estimation with Self-Guided Sparse Volumetric Refinement.

Provided Model Variants

Model File	Precision	Size	Recommended Target
moge3_dense_stage_fp16.tflite	FP16	631.7 MB	Mobile GPU / WebGPU / Vulkan / Apple Metal / Android NNAPI
moge3_dense_stage_dynamic_int8.tflite	Dynamic INT8	316.3 MB	CPU / Edge ARM64 & x86_64 AVX-VNNI
moge3_dense_stage_fp32.tflite	FP32	1.26 GB	Precision Reference Baseline

Model Architecture

The dense stage of MoGe-3 includes the DINOv2 Vision Transformer (ViT-L) backbone, multi-scale FPN neck, and intermediate geometry prediction heads (raw coordinate points, normal map, mask, and metric scale).

Inputs & Outputs

• Input: image [1, 3, 518, 518] (normalized RGB float32)
• Outputs:
  • raw_points: [1, 3, 672, 672] (affine factorized coordinates $x/z, y/z, \log z$)
  • normal: [1, 3, 672, 672] (surface normal unit vectors)
  • mask: [1, 1, 672, 672] (foreground valid mask)
  • metric_scale: [1, 1] (scene metric scaling factor)
  • encoder_features: [1, 1024, 42, 42] (visual tokens for 3D sparse refinement)

Citation

@article{moge3,
  title={MoGe-3: Fine-Detail Monocular Geometry Estimation with Self-Guided Sparse Volumetric Refinement},
  author={Wang, Ruicheng and others},
  journal={arXiv preprint},
  year={2026}
}