Marigold V2: A Single-GPU DiT Depth Estimator Hits SIGGRAPH Asia 2026

Marigold V2 has been released and accepted to SIGGRAPH Asia 2026. Author Anton Obukhov continues the V1 "research anyone can do" recipe of post-training an image generator into a depth estimator; V2 moves to a Diffusion Transformer (DiT) backbone with razor-sharp edges, 2K resolution without OOM, and training in days on a single consumer GPU. Website, code, weights, and an online demo are all public.

Confirmed

• Released by Huawei Bayer Lab; key improvements include single-step flow-matching inference balancing speed and quality, and a Sinkhorn-based semantic alignment mechanism (per @huawei-bayerlab).
• Open low-cost recipe: start from a pretrained DiT (Qwen-Image-Edit-2509), quantize to 4-bit, add rank-128 QLoRA, fine-tune in days on a single 32GB consumer GPU.
• All modalities (depth, surface normals, albedo) improve over the predecessor; on depth, it beats all models trained on comparable data across NYUv2, KITTI, ETH3D, ScanNet, and DIODE, with RGB→V1.1→V2 comparison sliders provided.
• Training follows the "unconventional but effective" tradition: V1 stuffed depth maps into an image VAE; V2 encodes ground-truth depth via DINOv3 and aligns DiT features to it (iREPA-depth).

Why it matters

• Shows high-quality depth estimation is achievable in days on one consumer GPU, dramatically lowering the research barrier for small labs.
• Demonstrates the potential of generative image models (especially DiTs) as backbones for dense prediction; the recipe (4-bit quantization + QLoRA, DINOv3 alignment, single-step flow matching) is a reference for other dense prediction tasks.

2026-09-09 ~ 2026-09-10 · 11 related posts

• [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
• Marigold V2 tops NYUv2, KITTI, ETH3D, ScanNet and DIODE depth benchmarks — AntonObukhov1 · 2026-09-09
• [source] Marigold V2 retools diffusion transformers for sharper monocular depth estimation — huawei-bayerlab · 2026-09-09
• Marigold V2: single-step DiT depth estimation trainable on one 32GB consumer GPU — _akhaliq · 2026-09-10
• Marigold V2 arrives as Qwen-Video-Edit repurposes an image editing DiT for video — AntonObukhov1 · 2026-09-10
• Marigold V2 depth estimation demoed on an ECCV 2026 art gallery piece — AntonObukhov1 · 2026-09-10
• Marigold V2 ships with an online demo for interactive depth estimation — AntonObukhov1 · 2026-09-10

2 near-duplicate retellings: AntonObukhov1 · AntonObukhov1