igzi/depth-anything-v2-large-lora-booster

Depth Anything V2 Large — LoRA adapter (Booster dataset)

LoRA fine-tuning of Depth Anything V2 Large on the Booster stereo depth dataset.

Training details

Base model	depth-anything/Depth-Anything-V2-Large-hf
Dataset	Booster stereo (prepared split, metric depth in metres)
Loss	Affine-invariant L1 + gradient loss (scale+shift in disparity space)
LoRA rank / alpha	16 / 32
LoRA targets	query, key, value
Best val AbsRel	0.0333 (scale+shift aligned)
Epochs trained	2

Usage

from transformers import AutoModelForDepthEstimation, AutoImageProcessor
from peft import PeftModel
from PIL import Image
import torch

processor = AutoImageProcessor.from_pretrained(
    "depth-anything/Depth-Anything-V2-Large-hf",
    size={"height": 518, "width": 518},
)
base = AutoModelForDepthEstimation.from_pretrained(
    "depth-anything/Depth-Anything-V2-Large-hf"
)
model = PeftModel.from_pretrained(base, "igzi/depth-anything-v2-large-lora-booster")
model.eval()

image = Image.open("image.jpg").convert("RGB")
inputs = processor(images=image, return_tensors="pt")
with torch.no_grad():
    outputs = model(**inputs)
depth = outputs.predicted_depth   # relative disparity; use scale+shift to get metric depth