Fine-Tuning Image-Conditional Diffusion Models is Easier than You Think

Benchmark Model Rank Results
monocular-depth-estimation-on-kitti-relative-depth-alignedMarigold + E2E FT (zero-shot)#3AbsRel: 0.096
monocular-depth-estimation-on-kitti-relative-depth-alignedStable Diffusion + E2E FT (zero-shot)#6delta1: 0.921
monocular-depth-estimation-on-diode-relative-depth-alignedMarigold + E2E FT (zero-shot)#4AbsRel: 0.302delta1: 0.779
surface-normals-estimation-on-nyu-depth-v2Marigold + E2E FT(zero-shot)#5% < 11.25: 61.4Mean Angle Error: 16.2
monocular-depth-estimation-on-scannet-relative-depth-alignedMarigold + E2E FT (zero-shot)#7AbsRel: 0.058
monocular-depth-estimation-on-scannet-relative-depth-alignedStable Diffusion + E2E FT (zero-shot)#13delta1: 0.965
surface-normals-estimation-on-sintelGeoWizard + E2E FT (zero-shot)#7mAE: 33.4
monocular-depth-estimation-on-eth3d-relative-depth-alignedMarigold + E2E FT (zero-shot)#8AbsRel: 0.062delta1: 0.959
monocular-depth-estimation-on-nyu-depth-v2Marigold + E2E FT(zero-shot)#9absolute relative error: 0.052Delta < 1.25: 0.966