Skip to content

[Feature Request] Metric checkpoints at Base/Small scale (DA3Metric-Base / DA3Metric-Small) #261

Description

@M-Colley

Is your feature request related to a problem?

The Monocular Metric series currently ships only at Large scale (DA3METRIC-LARGE, 0.35B).
The any-view Main series already provides DA3-BASE (0.12B) and DA3-SMALL (0.08B), and
there's a DA3MONO-LARGE relative model — but all of the smaller checkpoints are relative-only.
There is no lightweight checkpoint that outputs real-world metric scale.

This is a blocker for real-time / on-device deployment. On edge runtimes — Unity Inference
Engine (Sentis), mobile, standalone XR headsets — the Small/Base models are the only ones that
run at interactive frame rates and are small enough to convert and quantize for the target,
but they're relative-only. The Large metric model is both too heavy for these targets and, in
practice, harder to get running in these runtimes. So any application needing absolute distance
on-device (AR/XR, robotics, assistive/accessibility tools) currently has to choose between
metric scale and a deployable footprint.

For context: Depth Anything V2 did release metric checkpoints at Small/Base/Large, so the move
to DA3 effectively removes the small-footprint metric option for edge users.

Describe the solution you'd like

Release metric checkpoints at the smaller scales — DA3Metric-Base and DA3Metric-Small — by
fine-tuning the metric head onto the existing Base/Small backbones, mirroring DA3METRIC-LARGE.
Even a single additional Small metric checkpoint would be valuable, and the Apache-2.0 license on
the current metric model makes a small variant especially useful for production/edge use.

Additional context

  • The backbones already exist (DA3-BASE, DA3-SMALL), so this is primarily a metric fine-tune
    at smaller scale rather than new architecture.
  • Happy to benchmark candidate checkpoints on Unity Inference Engine / mobile XR and report back
    on accuracy and runtime if that's useful.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions