SHARP Predict Depth

Not every SHARP workflow needs a point cloud. Sometimes you just want the depth estimate - as a mask source, a sanity check, or as raw material for the panorama depth pipeline (AlignDepthMaps, ProjectDepthToPanorama) further down this pack. This node gives you exactly that: SHARP's monocular depth prediction as a plain image, with none of the Gaussian decoding or PLY writing SharpPredict does.

Depth estimation as a category is well-trodden ground - Depth Anything, MiDaS, Marigold, MoGe all do some version of "one photo in, a distance map out," and the general rule that applies across all of them applies here too: sharp relative depth is usually more useful for downstream work than technically-more-accurate metric depth, unless something downstream genuinely needs real-world distances. Going by the two node names sharing this pack, SharpPredictMetricDepth is the one that explicitly claims real-world scale; this node is the more general relative-depth output.

How it works

Same underlying SHARP model as every other Predict node, just stopping one step earlier - before the Gaussian decoder runs. The interesting extra it has that the Gaussian-producing nodes don't: an optional reference_depth input, described as accepting depth from something like DepthAnythingV3 "for dense alignment." In plain terms, if you already trust a different depth estimator's geometry for this image, you can hand it to SHARP and get its own prediction nudged to agree with that reference instead of running fully independently.

The inputs and outputs that matter

• model / image (required).
• extrinsics (optional, pass-through) - from SamplePanorama if you're in a multi-view context.
• intrinsics (optional) - used for depth scale.
• reference_depth (optional) - external depth maps to align against. The pack's own tooltip is explicit that this must match the image batch size, which is the one hard constraint worth remembering before you wire it up.

Outputs: depth_maps (the prediction itself), extrinsics / intrinsics (passed through), and alignment_maps - a second image output that reads as a visualization of how much correction got applied where when reference_depth was used.

Installing it

Manager: search "Sharp," install highest (try nightly if it acts up). Or Install via Git URL with https://github.com/PozzettiAndrea/ComfyUI-Sharp.git. Manual: clone, pip install -r requirements.txt --upgrade, python install.py, restart. Weights auto-download to ComfyUI/models/sharp/ on first run.

Common issues

The one concretely documented failure mode is a batch-size mismatch between image and reference_depth - the pack's tooltip states the requirement outright, so if you're wiring in an external depth source, count your batches before you run.

Beyond that, this node inherits the pack's general growing pains rather than anything specific to itself: 8GB VRAM cards have reported OOMs or silent hangs on release, and this is a young pack (first shipped December 2025) where edge cases are still surfacing. If you skip reference_depth entirely and just want SHARP's own unaligned prediction, this node is about as low-friction as the pack gets - no viewer dependency needed, since the output is a plain image you can drop straight into Preview Image.