Marigold V2 — 單張照片深度估計 (Monocular Depth Estimation)

• URL: https://www.threads.com/@sw_ai_life/post/DdTB57Agrgj
• Date Saved: 2026-09-15
• Source: Threads (@sw_ai_life)
• Tags: creative-ai, ai-tools
• Repo: https://github.com/huawei-bayerlab/marigold-v2 (⭐677)
• HuggingFace: https://huggingface.co/spaces/huawei-bayerlab/marigold-v2
• Original Marigold: https://github.com/prs-eth/Marigold (⭐3.2k, CVPR 2024 Oral/Best Paper Candidate)

Summary

Marigold V2 by Huawei BayerLab — a diffusion transformer approach to monocular depth estimation from a single photo.

Key features:

• Sharper boundary detection on fine/complex edges: hair, fur, grass blades
• Based on Qwen-Image-Edit-2509 architecture
• Trainable on a single 32GB GPU
• Applications: convert regular photos to 3D, shift video camera angles, spatial understanding for AR, gaming, robotics

The original Marigold (prs-eth) won CVPR 2024 Oral + Best Paper Award Candidate. V2 extends the approach using diffusion transformers instead of the original UNet-based diffusion.