Marigold V2: Monocular Depth from a Single Frame in One Step

🤖 Marigold V2: Depth from a Single Image in One Step

HUAWEI Bayer Lab, together with EPFL and the University of Bologna, has released Marigold V2 — a one-step monocular depth estimator based on the pretrained diffusion transformer Qwen-Image-Edit-2509. Training — LoRA (rank 128) on top of a 4-bit DiT and distillation from a multi-step flow-matching model. The authors claim an improvement in AbsRel of 16–26% on KITTI and ETH3D and SOTA for normals and albedo.

🌍 Training is possible on a single consumer GPU with 32 GB (~5 days, batch=1), and inference takes one step instead of dozens: ~17 GB VRAM at 1024×1024. This reduces the cost of dense prediction for robotics and 3D reconstruction.

👤 The model is available in a web demo on Hugging Face and can be run locally (Apache-2.0). There is a Log-layered checkpoint that can "see" through transparent objects — for example, glass.

Source 1: https://github.com/huawei-bayerlab/marigold-v2 Source 2: https://arxiv.org/abs/2609.08084