ilessio-aiflowlab/project_frigg

FRIGG — AnyDepth Depth Estimation

Paper: AnyDepth: Depth Estimation Made Easy Module: FRIGG (ANIMA Wave 6) Architecture: DINOv2-Base (frozen) + SDT Head (trained)

Model Details

• Encoder: DINOv2-ViT-B/14 (86.6M params, frozen)
• Decoder: SDT Head (9.5M params, trained)
• Total: ~96M params (only 9.5M trainable)
• Input: RGB image (518x518 optimal)
• Output: Dense depth map

Exported Formats

Format	File	Use Case
PyTorch (.pth)	pytorch/frigg_v1.pth	Training, fine-tuning
SafeTensors	pytorch/frigg_v1.safetensors	Fast loading, safe
ONNX	onnx/frigg_v1.onnx	Cross-platform inference
TensorRT FP16	tensorrt/frigg_v1_fp16.trt	Edge deployment (Jetson/L4)
TensorRT FP32	tensorrt/frigg_v1_fp32.trt	Full precision inference

Usage

import torch
from anima_frigg.models import build_anydepth, load_dinov2_backbone

backbone = load_dinov2_backbone(model_path="pytorch/frigg_v1.pth")
model = build_anydepth(backbone)
model.load_state_dict(torch.load("pytorch/frigg_v1.pth"))

image = torch.randn(1, 3, 518, 518).cuda()
depth = model(image)  # (1, 1, H, W)

Training

• Datasets: NYUv2 (1449 samples) + KITTI (7481 pairs)
• Loss: SILog + Gradient Matching
• Optimizer: AdamW (lr=1e-4)
• Hardware: 4x NVIDIA L4 23GB (DDP)

Citation

@article{ren2026anydepth,
  title={AnyDepth: Depth Estimation Made Easy},
  author={Ren, Zeyu and Zhang, Zeyu and Li, Wukai and Liu, Qingxiang and Tang, Hao},
  year={2026}
}

Generated: 2026-03-31T16:39:28.292515