Marigold V2 Cuts KITTI, ETH3D AbsRel 16-26% With QLoRA-Tuned DiT

Marigold V2 reports a 16-26% AbsRel improvement on KITTI and ETH3D over the previous best in monocular depth estimation, using a diffusion transformer fine-tuned to output depth in one flow-matching step. The paper, authored by researchers at HUAWEI Bayer Lab, EPFL and the University of Bologna, starts from the pretrained Qwen-Image-Edit-2509 DiT and adapts it with a rank-128 QLoRA that fits on a single 32 GB consumer GPU.

The authors describe the recipe compactly: Marigold V2 "repurposes diffusion transformers for monocular depth estimation via single-step flow-matching inference, semantic alignment, and a Sinkhorn-based two-stage fine-tuning protocol, yielding sharper out-of-distribution depth maps and strong results on related dense regression tasks." The alignment piece, called iREPA-depth, pulls DINOv3 features from ground-truth depth to keep the model's semantics honest. The Sinkhorn loss does soft optimal-transport matching over 5×5 image tiles, a way to fine-tune when the ground truth is noisy.

Beyond depth, the same recipe carries over to surface normals, intrinsic image decomposition, sparse-to-dense metric depth completion, and transparent-object depth via LayeredDepth-Syn. Inference is single-step and handles 2K resolution without going OOM. Evaluations cover NYUv2, KITTI, ETH3D, ScanNet and DIODE.

The abstract does not publish per-dataset AbsRel tables for the non-KITTI/ETH3D benchmarks. It arrives days after Scal3R, another KITTI-focused result that moved through our computer-vision feed last week, one of 61 computer-vision stories we have tracked in the last 90 days.