Marigold V2 Released: Single-GPU Fine-tuned DiT for Depth Estimation

Marigold V2 has been officially released and accepted by SIGGRAPH Asia 2026. Author Anton Obukhov continues V1's "research anyone can do" approach, post-training a pretrained image generation model into a depth estimator. V2 upgrades to a diffusion Transformer (DiT) architecture with extremely sharp edges, and the entire training runs on a single GPU.

Confirmed

A low-cost reproduction recipe is public: starting from a pretrained DiT (specifically Qwen-Image-Edit-2509), quantized to 4-bit with rank-128 QLoRA, fine-tuning completes in a few days on a single 32GB consumer GPU.
Results: all modalities—depth, surface normals, albedo—outperform the previous generation; on depth, it beats all models trained on comparable data volumes across five benchmarks: NYUv2, KITTI, ETH3D, ScanNet, and DIODE.
Training follows the "unconventional but effective" tradition: V1 stuffed depth maps directly into an image VAE (theoretically shouldn't work, but it did); V2 encodes ground-truth depth via DINOv3 and aligns DiT features with it (iREPA-depth).

Why it matters

It demonstrates that high-quality depth estimation can be trained in a few days on a single consumer GPU, dramatically lowering the research barrier for this direction.
It showcases the potential of generative image models (especially the DiT architecture) as backbones for dense prediction tasks; the recipe (4-bit quantization + QLoRA fine-tuning, feature alignment) can serve as a reference for other dense prediction tasks.

2026-09-09 ~ 2026-09-09 · 5 related posts

Primary sources

Marigold V2 launches at SIGGRAPH Asia 2026: sharp diffusion-transformer depth estimation — AntonObukhov1 · 2026-09-09 00:12 (UTC+08:00)
Marigold V2 tops NYUv2, KITTI, ETH3D, ScanNet and DIODE depth benchmarks — AntonObukhov1 · 2026-09-09 00:12 (UTC+08:00)
Fine-tune Qwen-Image-Edit into a depth estimator on one 32GB GPU with 4-bit QLoRA — AntonObukhov1 · 2026-09-09 00:12 (UTC+08:00)

[source] Marigold V2 launches at SIGGRAPH Asia 2026: sharp diffusion-transformer depth estimation — AntonObukhov1 · 2026-09-09
[source] Fine-tune Qwen-Image-Edit into a depth estimator on one 32GB GPU with 4-bit QLoRA — AntonObukhov1 · 2026-09-09
Marigold V2 recipe: 4-bit quantized DiT + rank-128 QLoRA, fine-tuned on one 32GB GPU — AntonObukhov1 · 2026-09-09
[source] Marigold V2 tops NYUv2, KITTI, ETH3D, ScanNet and DIODE depth benchmarks — AntonObukhov1 · 2026-09-09

1 near-duplicate retellings: AntonObukhov1