Marigold V2 turns image-editing DiT into single-step depth model, beats SOTA by 16-26%

bdsqlsz · x · 2026-09-14

Marigold V2 (Apache 2.0) adapts a pretrained Qwen-Image-Edit DiT into a single-step dense predictor via lightweight fine-tuning.

Best zero-shot depth results across comparable-data models; AbsRel improves 16%–26% over prior best on KITTI and ETH3D
Two-stage iREPA and SinkLoss fix smoothing and flying-pixel artifacts, keeping fur, foliage, and fine wires crisp
Same framework hits SOTA in depth completion (lowest RMSE on all four benchmarks), see-through depth, surface normals, and intrinsic decomposition
Training takes under a week on a single 32 GB GPU

More from Multimodal