ZICQ
中 Log in / Sign up
Newsroom LLMs & Foundation Models #Depth Estimation #Diffusion Transformer #Huawei #Computer Vision #AI Models

Marigold V2 Released: Diffusion Transformer-based Breakthrough in Monocular Depth Estimation

Avatar of Mr.Xu

By Mr.Xu Compiled & Reviewed by Editorial

Published: · 9 views

中文阅读 (Chinese) English Version
Co-reading Select any sentence in the article to highlight, annotate, or copy a quote card. Join discussion ↓

Summary:Huawei Bayer Lab has released Marigold V2, a set of techniques repurposing diffusion transformer (DiT) architecture for state-of-the-art monocular depth estimation. Marigold V2 employs single-step inference and quantization to preserve model capacity while reducing computational costs. Its key innovations include aligning internal representations with semantic features and adopting a two-stage fine-tuning protocol based on a novel Sinkhorn loss. Experiments show a 16-26% improvement in AbsRel on


Technical Breakthroughs and Core Innovations

Marigold V2, developed by Huawei Bayer Lab, is a set of techniques leveraging the diffusion transformer (DiT) architecture for monocular depth estimation. The key innovations include:

  • Single-step Inference and Quantization: By employing single-step inference and necessary quantization, Marigold V2 preserves model capacity while reducing computational costs, making it more efficient for practical applications.
  • Model Representation Alignment: Aligning the model's internal representations with semantic features extracted from ground truth addresses the artifacts of naive training.
  • Two-stage Fine-tuning Protocol: The adoption of a two-stage fine-tuning protocol based on a novel Sinkhorn loss further enhances the model's refinement capabilities.

Experimental Results and Performance

Experiments on KITTI and ETH3D datasets show a 16-26% improvement in AbsRel over the previous best methods. Additionally, Marigold V2 produces sharper and more detailed depth maps, particularly excelling in handling complex details like fur, foliage, and hair-thin edges.

Application Scenarios and Industry Impact

Marigold V2 is not only effective in monocular depth estimation but also applicable to other dense regression tasks such as surface normal estimation and intrinsic image decomposition. Its application scenarios include:

  • Scene Reconstruction: Providing more accurate depth information for virtual and augmented reality.
  • Computational Photography: Enhancing the precision of image processing and editing.
  • Robot Navigation: Offering more reliable depth perception for robots.

Developer Recommendations

For researchers and developers in computer vision and deep learning, Marigold V2 offers an efficient and powerful solution for monocular depth estimation. We recommend visiting the project website (https://hf.co/spaces/huawei-bayerlab/marigold-v2-web) for more technical details and code examples.

Future Directions

With the release of Marigold V2, we anticipate its applications in various fields will drive further advancements in monocular depth estimation technology. Future research directions may include optimizing the model's generalization capabilities across different environments and exploring its applications in more dense regression tasks.


Source: Hugging Face Daily Papers (2026-09-08)

— END —

Tags: #Depth Estimation #Diffusion Transformer #Huawei #Computer Vision #AI Models

Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.

What's your take on this story?
🚀 Explore & Submit Works

Community Comments

Loading live comments and annotations…