Monocular Depth Estimation

The "3D from one photo" tier. Modern monocular depth networks won't replace photogrammetry or LiDAR — the absolute scale is wrong, and back-of-object geometry is invented — but they're useful for relighting, parallax effects, depth-aware compositing, faux-3D thumbnails, and as a prior inside larger reconstruction pipelines. Pairs with Photogrammetry Software and Photogrammetry Mobile Apps (the proper full-3D capture path), Gaussian Splatting and NeRF & Radiance Fields (where monocular priors sometimes feed in), and AI / LLMs for the broader ML context.

State of the art (free / OSS)

• ★ Depth Anything V2 (Bytedance / TikTok) — free, OSS (Apache-2); the 2024–2025 SOTA on monocular depth; large + small + ViT variants. Drop-in via Hugging Face / ComfyUI / Diffusers. The default in 2026 for almost any use.
• ★ MiDaS (Intel ISL) — free, OSS (MIT); the long-running monocular depth baseline; v3.1 still useful; many tutorials reference it. Often pre-installed in ComfyUI / A1111.
• ZoeDepth — free, OSS (MIT); MiDaS + metric-scale; Intel.
• Marigold — free, OSS; diffusion-based depth (slow but very crisp); ETH Zurich.
• Depth Pro (Apple) — free, OSS (Apple Sample Code License); 2024; metric-scale, single-image, sub-second.
• GeoWizard — free, OSS; multi-task depth + normals.
• UniDepth — free, OSS; metric-scale across domains.

Hosted / device APIs

• ★ iOS LiDAR / TrueDepth — Apple's native depth APIs; free for developers; real depth, not estimated. iPhone 12 Pro+, iPhone X+ TrueDepth (front).
• ARCore Depth API (Google) — free for developers; depth on supported Android phones; partly estimated, partly sensor.
• Replicate / Hugging Face Spaces — host any of the OSS models; pay-per-call.
• MiDaS / ZoeDepth on Replicate — paid hosted endpoints.

Where it actually fits

• Relighting / re-rendering 2D photos — drive parallax / depth-of-field with a depth map.
• Faux-3D thumbnails / GIFs — common in social marketing; "Live photo"-style.
• Compositing / matte extraction — soft-edged subject masks via depth-thresholding.
• NeRF / splat priors — monocular depth as a regularizer in single-view radiance-field training.
• Robot vision / perception — when the only sensor is a camera.
• Heightmaps for terrain prototyping — quick-and-dirty, then replace with real DEM.

What it doesn't do

• Metric 3D reconstruction — the absolute scale is unknown without calibration. Use Photogrammetry Software instead.
• Reproduce occluded geometry — only the visible front is real.
• Replace LiDAR for safety / surveying — accuracy is in the percent range, not the millimetre range.

Tooling / integration

• ★ Hugging Face transformers / diffusers pipelines — free; one-liner inference for Depth Anything / MiDaS / ZoeDepth.
• ComfyUI — free, OSS; workflow graph; many depth nodes pre-built.
• Automatic1111 / Forge / SwarmUI — free, OSS; depth via ControlNet preprocessors.
• Blender add-ons — community add-ons that turn a depth map into a heightmap mesh.
• OpenCV / PyTorch — if you're rolling your own pipeline.

Single-image to 3D (full reconstruction)

A different lineage from depth estimation; mentioned for completeness:

• TripoSR / Tripo — free / paid; image-to-3D mesh in seconds.
• Stable 3D / Stable Zero123 — research; image-to-3D.
• InstantMesh, OpenLRM, Wonder3D — research; single-image 3D.
• Luma AI Genie — paid; text-/image-to-3D.
• Meshy AI, Rodin AI — paid SaaS; text-/image-to-3D.

These don't replace photogrammetry either — accuracy is invented — but they're useful for concept / placeholder geometry.

Pick this if…

• Default OSS monocular depth in 2026: Depth Anything V2.
• Need metric scale in addition to relative depth: Apple Depth Pro or ZoeDepth.
• Inside ComfyUI / A1111 already: MiDaS or Depth Anything V2 ControlNet.
• iPhone Pro app, real depth: ARKit / TrueDepth / LiDAR APIs.
• Want actual full-3D scan, not estimated depth: Photogrammetry Software or Handheld 3D Scanners.
• Single-image-to-3D for concept work: TripoSR / Meshy / Luma Genie (results are hit-or-miss).