Marigold

Marigold is a computer vision method that adapts latent diffusion models for dense image analysis tasks like depth estimation.

Marigold is a family of conditional generative models and a fine-tuning protocol that extracts visual knowledge from pretrained latent diffusion models like Stable Diffusion. It adapts these models for dense image analysis tasks, including monocular depth estimation, surface normal prediction, and intrinsic image decomposition. The method requires minimal architecture modification and trains with small synthetic datasets. It is built for computer vision tasks and achieves zero-shot generalization on unseen data. The project supports inputs like images and outputs depth maps, surface normals, albedo, diffuse shading, and roughness.

Marigold at a glance

Best for: Researcher, Software engineer
Runs on: Web, Linux, API, self-hosted, open source
Pricing: Free
Released: November 2023
Category: AI image tool
Website: github.com/prs-eth/Marigold

Key features

Monocular depth estimation
Surface normal prediction
Intrinsic image decomposition
Zero-shot generalization
Half-precision float execution
Ensemble inference support
Customizable processing resolution

Use it for

Estimating high-resolution depth maps from single images
Predicting surface normals for computer vision tasks
Decomposing images into albedo, shading, and residual

Alternatives to Marigold

Diffusers — A modular toolbox for inference and training of diffusion models.
stable-diffusion.cpp — Runs diffusion models efficiently in pure C and C++ with no external dependencies.
Lance — It supports unified image and video understanding, generation, and editing at a 3B scale.
Dsh Sd Bridge Guard — Connects DeepSeek directly to local Stable Diffusion without manual setup.
Supervision — It offers a unified Detections object and extensive dataset converters.
Ultralytics YOLO — Delivers high-speed and high-accuracy results in real-time.
LyCORIS — It provides diverse algorithms like LoHa, LoKr, and DyLoRA for Stable Diffusion fine-tuning.
Eagle — It provides data-centric strategies for advanced multimodal research.
KohakuMossTrainer — Fits adapter training onto cards with 6 GB of VRAM.
Lyra — Generates persistent, explorable 3D worlds from camera trajectories and images.
Image or AI — It checks Content Credentials, metadata, and uses a pattern model for evidence.
Viso — Builds complete vision apps from simple prompts without model training or data labeling.

All Marigold alternatives