Marigold V2, a new research effort from the HUAWEI Bayer Lab alongside EPFL and the University of Bologna, revisits the original Marigold approach for monocular depth estimation by rebuilding it around the diffusion transformer (DiT) architecture. The work is set to appear at SIGGRAPH Asia 2026.
Monocular depth estimation remains a highly ill-posed problem, and the authors note that even recent models struggle to generalize to out-of-distribution inputs or produce sharp, detailed depth maps. Marigold V2 targets single-step inference from pretrained multi-step flow-matching models, with quantization where needed, so the model keeps its capacity while staying cheap to run.
The team analyzes artifacts that arise from naive training and proposes two remedies: aligning the model's internal representations with semantic features extracted from ground-truth data, and a two-stage fine-tuning protocol built around a novel Sinkhorn-based loss. The result is crisper, cleaner depth maps that generalize well out-of-distribution, with a 16-26% improvement in AbsRel over the previous best on KITTI and ETH3D. Qualitatively, the model resolves fur, foliage and hair-thin edges that eluded prior models.
Beyond depth, Marigold V2 reports state-of-the-art results on other dense regression tasks such as surface normals estimation and intrinsic image decomposition. The recipe is designed to be accessible: a large pretrained DiT (Qwen-Image-Edit-2509) is quantized to 4-bit, augmented with a rank-128 QLoRA, and fine-tuned on a single 32 GB consumer GPU, with inference completing in one step without running out of memory at 2K resolution.
The authors also describe a Sinkhorn-based matching loss (SinkLoss) that tolerates imperfect ground truth by tiling an image into 5x5 blocks and softly matching 25 predicted depths to 25 ground-truth depths via transport cost, rather than pixel-to-pixel comparison. The same model can be fine-tuned on the last ray hit of LayeredDepth-Syn to see through glass for computational photography, and can perform metric depth completion by fitting a small LoRA at test time on sparse lidar points without retraining.