search

prs-eth / Marigold

[CVPR 2024 - Oral, Best Paper Award Candidate] Marigold: Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation
https://marigoldmonodepth.github.io
Apache License 2.0
2.25k stars 124 forks source link

The purpose of using v_prediction as the target? #88

Closed yongmayer closed 3 months ago

yongmayer commented 3 months ago

Thank you so much for sharing this amazing work! Could you shade some light on why v_prediction is chosen as the prediction type and "get_velocity" as the target? What are the benefits of doing so over other options, say using noise as the target? Looking forward to your comments.

markkua commented 3 months ago

Hi, we simply follow the pretrained model we used (SD-v2), which was fine-tuned from SD-v2-base using v_prediction.