EfficientViT-Depth-5M-int8

INT8 QDQ version of EfficientViT-Depth-5M: a 4.85M-parameter relative-depth model with an EfficientViT-M2 encoder and a Lite-DPT decoder.

This is an independent community model, not an official Depth Anything release. The float parent was distilled from Depth Anything V2 Small; this repository contains the quantized ONNX graph.

Float vs INT8 size/quality comparison.

NYUv2 indoor scene; float vs INT8 at 336×448.

Qualitative examples on ADE20K/NYUv2 scenes; per-row color scale.

Highlights

• 3.48 MB static ONNX, down from 9.84 MB float
• 0.8319 NYU-654 δ1 at 336×448
• 0.6729 teacher-fidelity δ1 on 700 held-out images
• 24.9 ms / 40.1 FPS on CPU with ONNX Runtime, 8 threads, 336×448
• Static 336×448 graph only

Comparison

All quality numbers are at 336×448 on the 654-image NYUv2 Eigen test split.

model	params	disk	NYU-654 δ1	teacher fidelity δ1	CPU latency
Depth Anything V2 Small	24.79M	99.06 MB	0.9213	1.000	—
Depth Anything V2 Tiny (8-layer)	16.50M	66.04 MB	0.8936	0.8976	—
EfficientViT-Depth-5M float	4.85M	9.84 MB	0.8546	0.7242	12.1 ms / 82.5 FPS
EfficientViT-Depth-5M INT8	4.85M	3.48 MB	0.8319	0.6729	24.9 ms / 40.1 FPS
EfficientViT-Depth-5M 4-bit (experimental)	4.85M	2.31 MB	0.6456	0.4260	—

This INT8 model is 35.4% of the float parent’s size and 94.7% smaller than the 8-layer Tiny, but it loses 2.27 NYU δ1 points and 5.13 fidelity δ1 points relative to the float parent. It also runs about 2× slower than the float ONNX graph on this CPU because QDQ overhead dominates on a small convolutional network.

Why not 4-bit?

Direct ONNX 4-bit QDQ was tested and rejected:

variant	size	NYU-654 δ1	fidelity δ1
INT8 QDQ (shipped)	3.48 MB	0.8319	0.6729
4-bit weights + int8 activations	2.31 MB	0.6456	0.4260
4-bit weights + int16 activations	2.31 MB	~0.66 (100-image)	~0.41 (50-image)

The model is roughly 95% convolutional. ONNX Runtime’s 4-bit path is MatMul-oriented; direct 4-bit convolution QDQ collapses edge and detail quality. A true 4-bit compact model would require 4-bit-aware QAT or a runtime/NPU with native 4-bit convolution support. The valid quantized artifact is this INT8 model.

Quantization

setting	value
format	ONNX QDQ
weights	INT8, per-channel
activations	INT8
calibration	MinMax, 256 stratified images
input	static 1×3×336×448
output	static 1×336×448 inverse depth

Metrics

NYUv2 Eigen-654, 336×448

metric	value
δ1	0.831850
AbsRel	0.136737
RMSE	0.084210

Teacher fidelity, 700 held-out images

metric	value
δ1	0.672902
AbsRel	0.368645
Pearson	0.892266
gradient corr	0.365290

The float parent scores 0.854563 / 0.724169 on the same protocol.

Usage

pip install -r requirements.txt
python example.py --image photo.jpg --output depth_int8.png

The graph is static at 336×448. Resize the RGB image to exactly that size, apply ImageNet normalization, and run:

import numpy as np
from PIL import Image
import onnxruntime as ort

mean = np.array([0.485, 0.456, 0.406], dtype=np.float32)
std = np.array([0.229, 0.224, 0.225], dtype=np.float32)

image = Image.open("photo.jpg").convert("RGB").resize((448, 336), Image.Resampling.BICUBIC)
x = np.asarray(image, dtype=np.float32) / 255.0
x = (x - mean) / std
x = np.transpose(x, (2, 0, 1))[None].astype(np.float32)

session = ort.InferenceSession("efficientvit_depth_5m_int8.onnx", providers=["CPUExecutionProvider"])
depth = session.run(None, {session.get_inputs()[0].name: x})[0][0]
# inverse depth: higher = closer; not metric depth

For non-336×448 inputs use the dynamic float graph in the parent repository, or re-export this model at the target resolution.

When to use this model

Use INT8 when 3.48 MB is the hard constraint or the target runtime has a native INT8 path that actually accelerates it.

Use the float parent when:

• CPU latency matters (12.1 ms vs 24.9 ms here),
• detail/fidelity matters (0.8546 vs 0.8319 NYU δ1),
• you need dynamic resolution.

Limitations

• Relative inverse depth only; no metric scale.
• Static 336×448; no dynamic shape.
• Fine structures and thin objects are weaker than the float model.
• CPU QDQ can be slower than float; benchmark the target runtime before selecting it.
• Not an official Depth Anything release and not affiliated with its authors.

Files

file	description
efficientvit_depth_5m_int8.onnx	static 336×448 INT8 QDQ ONNX
config.json	architecture and quantization metadata
preprocessor_config.json	resize/normalization metadata
metrics.json	full metrics and artifact hash
example.py	ONNX Runtime example
comparison.png	quality and size comparison
example.png	float vs INT8 example on an indoor scene
examples.png	five-scene qualitative gallery
checksums.txt	SHA256 checksums

License

Apache-2.0. See LICENSE and NOTICE. The float parent and teacher are also Apache-2.0.

Contact

Petros Fanioudakis — @Petrouil