Huawei Marigold V2 Trains on a Single 32 GB GPU
Huawei and its partners have released Marigold V2, a single-step depth estimation model that achieves state-of-the-art results while training on a single 32 GB consumer GPU in under a week.
Researchers from HUAWEI Bayer Lab, EPFL, and the University of Bologna have introduced Marigold V2, an open-source depth estimation model. The system repurposes Qwen-Image-Edit-2509, an open-source image-editing diffusion transformer, converting it into a single-step depth estimator. By replacing the Stable Diffusion U-Net used in the original Marigold with a diffusion transformer, the team reduced inference to a single denoising pass.
To make training highly accessible, the researchers quantized the backbone to 4-bit precision and applied rank-128 QLoRA adapters. This setup allows Marigold V2 to be trained in less than a week on a single 32 GB consumer GPU. During deployment, the model can process high-resolution 2048x2048 images within 29 GB of GPU memory without running out of memory. This represents a massive hardware reduction compared to competing detail-oriented depth systems, which often require 80 GB accelerators or distributed training setups.
The model introduces two key innovations: iREPA-depth, which utilizes DINOv3 feature alignment, and SinkLoss, which applies Sinkhorn optimal transport per tile to address unreliable boundaries in synthetic training datasets. These techniques helped Marigold V2 achieve state-of-the-art zero-shot depth estimation on major benchmarks, including NYUv2, KITTI, ETH3D, ScanNet, and DIODE. Beyond standard depth mapping, the system supports see-through depth to predict behind glass, metric depth completion from sparse lidar, and can estimate surface normals and albedo by loading different LoRA adapters over the same frozen backbone.
For AI practitioners, Marigold V2 democratizes high-fidelity depth estimation by eliminating the need for expensive enterprise hardware. The release includes Apache 2.0-licensed code, model weights, and a browser demo, with an upcoming paper slated for ACM Transactions on Graphics and SIGGRAPH Asia. Developers can now deploy a state-of-the-art, multi-task spatial AI model on consumer-grade hardware.
This is our own summary of reporting by AlphaSignal
