Normal & Depth Map Generation¶
3D Generation
Open Source¶
-
DepthART: Scaling Foundation Monocular Depth to Tiny Models Discovered by AI
2026
A compact monocular depth estimation model designed for on-device deployment. It bridges the gap between heavy foundation models and tiny models, supporting both affine-invariant relative depth and camera-aware metric depth with high zero-shot generalization.
depthmetric-depthon-devicetiny-model -
GeoNeXt: Video Generative Models as Geometry Learner Discovered by AI
2026
A unified geometry estimation framework that repurposes pretrained video generative models for monocular depth and surface normal estimation. By formulating geometry estimation as a next-frame prediction task, it leverages temporal priors to co-generate highly consistent depth and normal maps.
depth-estimationnormal-estimationvideo-diffusion -
InfiniDepth Discovered by AI
2026
A monocular depth estimation framework that represents depth as neural implicit fields, enabling arbitrary-resolution and fine-grained depth estimation by querying depth at continuous 2D coordinates.
arbitrary-resolutiondepth-estimationimplicit-fields -
Marigold V2: Revisiting Diffusion Transformers for Monocular Depth Estimation Discovered by AI
2026
An upgraded version of the Marigold depth estimation framework that leverages Diffusion Transformers (DiT) to produce highly detailed, sharp, and out-of-distribution robust monocular depth and surface normal maps.
depthdiffusionditnormal-map -
MetricAnything: Scaling Metric Depth Pretraining with Noisy Heterogeneous Sources Discovered by AI
2026
MetricAnything is a scalable pretraining framework designed to learn accurate metric depth from noisy, diverse, and heterogeneous 3D sources. It establishes scaling laws for metric depth estimation, achieving strong zero-shot generalization across various real-world scenarios without requiring manual calibration.
depth-estimationfoundation-modelmetric-depth -
Sapiens2: Foundation for Human Vision Models Discovered by AI
2026
Sapiens2 is a family of high-resolution vision transformers (0.4B to 5B parameters) for human-centric tasks, including depth estimation and surface normal prediction. Pretrained on 1 billion human images, it supports native 1K and 4K resolution inference, delivering highly detailed and temporally consistent dense geometry predictions for human bodies.
depthgeometryhuman-centricnormal-mapGitHub · arXiv · Hugging Face
-
TransNormal-2: Geometry-Grounded Rectified Flow with Edge-Aware Decoding for Precise Normal Estimation Discovered by AI
2026
A FLUX.2-based rectified-flow framework designed for precise monocular surface normal estimation, correcting VAE reconstruction errors at object boundaries using geometry-aware training losses.
fluxnormal-maprectified-flow -
ViGeo Discovered by AI
2026
A feed-forward video geometry foundation model that recovers spatially dense and temporally consistent depth, point maps, and surface normals from video sequences.
depthnormal-mapvideo-geometry -
Weather-Conditioned Depth Anything Discovered by AI
2026
A weather-robust monocular depth estimation framework built on Depth Anything V2. It explicitly disentangles style from content using a Style Filter to achieve robust zero-shot relative depth estimation under adverse conditions like fog, rain, snow, and low light.
depthstyle-filterweather-robust -
ZipDepth: Bringing Lightweight Zero-Shot Monocular Depth Anywhere, on Any Device Discovered by AI
2026
A compact, lightweight monocular depth estimation network designed for zero-shot generalization and efficient on-device deployment across diverse platforms.
depthlightweighton-devicezero-shot -
Dens3R: A Foundation Model for 3D Geometry Prediction Discovered by AI
2025
A unified geometric foundation model that jointly predicts depth, surface normals, and point maps from single images. By modeling the inherent correlations between these geometric quantities, it ensures cross-modality consistency and improves reconstruction accuracy.
3d-reconstructiondepthfoundation-modelnormal-map -
Depth Any Camera: Zero-Shot Metric Depth Estimation from Any Camera Discovered by AI
2025
Depth Any Camera (DAC) is a zero-shot metric depth estimation framework that extends perspective-trained models to handle any camera type, including wide field-of-view (FoV) fisheye and 360-degree cameras. It enables accurate metric depth prediction across diverse camera geometries without requiring camera-specific fine-tuning.
depth-estimationmetric-depthzero-shot -
Depth Anything 3 Discovered by AI
2025
A geometry foundation model that predicts spatially consistent geometry from arbitrary visual inputs, with or without known camera poses, achieving state-of-the-art results in camera pose estimation, any-view geometry, and monocular depth estimation.
depthfoundation-modelgeometry -
Depth Anything at Any Condition Discovered by AI
2025
A robust monocular depth estimation model fine-tuned from Depth Anything V2, designed to handle diverse and challenging environmental conditions such as low light, adverse weather, and sensor distortions.
depthmonocularrobustnesszero-shot -
Lotus-2 Discovered by AI
2025
A two-stage deterministic framework for monocular geometric dense prediction that leverages pre-trained generative models as a deterministic world prior for depth and normal estimation.
depthdiffusiongenerativenormal-map -
Marigold: Affordable Adaptation of Diffusion-Based Image Generators for Image Analysis Discovered by AI
2025
An updated unified framework and fine-tuning protocol that adapts pretrained latent diffusion models like Stable Diffusion for dense image analysis tasks, including monocular depth estimation and surface normals prediction.
depthdiffusionnormal-mapstable-diffusion -
MoGe-2: Accurate Monocular Geometry with Metric Scale and Sharp Details Discovered by AI
2025
An advanced monocular geometry estimation model that recovers a metric-scale 3D point map, depth map, and surface normals of a scene from a single image with enhanced detail recovery.
3d-reconstructiondepthmetric-depthnormal-map -
NormalCrafter: Learning Temporally Consistent Normals from Video Diffusion Priors Discovered by AI
2025
A video-based surface normal estimation framework that leverages the temporal priors of video diffusion models. It introduces semantic feature alignment to ensure high-fidelity and temporally coherent normal maps across video sequences.
diffusionnormal-maptemporal-consistencyvideo-normal -
Online Video Depth Anything Discovered by AI
2025
An online video depth estimation framework based on Video Depth Anything that achieves temporally consistent depth prediction with low memory consumption by caching latent features.
depthonline-inferencevideo-depth -
UniDepthV2 Discovered by AI
2025
An upgraded version of the universal monocular metric depth estimation model that simplifies the architecture while improving zero-shot generalization and metric 3D scene reconstruction from a single image across diverse domains.
depth-estimationmetric-depthzero-shot -
Video Depth Anything Discovered by AI
2025
A video depth estimation model that extends Depth Anything to produce temporally consistent, high-quality depth sequences for super-long videos without requiring camera poses or optical flow.
depthtemporal-consistencyvideo-depth -
ChronoDepth: Learning Temporally Consistent Video Depth from Video Diffusion Priors Discovered by AI
2024
A video depth estimation framework that reformulates depth prediction into a conditional generation problem using video diffusion priors. It achieves high temporal consistency and per-frame spatial accuracy across long video sequences.
depthdiffusiontemporal-consistencyvideo-depth -
Depth Any Video with Scalable Synthetic Data Discovered by AI
2024
A video depth estimation model trained on a scalable synthetic dataset of 40,000 video clips from diverse games, leveraging generative video diffusion priors to handle varying video lengths and frame rates.
depthdiffusionvideo-depth -
Depth Anything V2 Discovered by AI
2024
A highly capable foundation model for monocular depth estimation that improves upon its predecessor by utilizing synthetic data, larger scale models, and a more robust training pipeline. It provides fine-grained details and robust generalization across diverse scenes.
depthfoundation-modelmonocularGitHub · arXiv · Hugging Face
-
Depth Anything: Unleashing the Power of Large-Scale Unlabeled Data Discovered by AI
2024
A highly practical foundation model for robust monocular depth estimation trained on a massive dataset of 1.5M labeled and 62M unlabeled images. It achieves strong zero-shot generalization across diverse environments, making it highly useful for generating depth maps from arbitrary game or anime scenes.
depth-estimationfoundation-modelmonocular-depth -
Depth Pro Discovered by AI
2024
A foundation model for zero-shot metric monocular depth estimation that synthesizes high-resolution depth maps with sharp details in less than a second without relying on camera metadata.
depthmetric-depthmonocularzero-shot -
DepthART: Monocular Depth Estimation as Autoregressive Refinement Task Discovered by AI
2024
A monocular depth estimation framework formulated as an autoregressive refinement task, utilizing a depth autoregressive transformer to iteratively refine depth predictions.
autoregressivedepthtransformer -
DepthCrafter Discovered by AI
2024
A state-of-the-art video depth estimation model that generates temporally consistent, high-resolution, and detailed depth sequences for open-world videos without requiring camera poses or optical flow.
depth-estimationtemporal-consistencyvideo-depth -
DepthFM Discovered by AI
2024
A fast and versatile monocular depth estimation model based on flow matching, framing depth estimation as a direct transport between image and depth distributions.
depthflow-matchinggenerativemonocular -
DSINE Discovered by AI
2024
A surface normal estimation model that utilizes per-pixel ray direction and encodes the relationship between neighboring surface normals by learning their relative rotation.
3d-reconstructionnormal-mapsurface-normal -
ECoDepth: Effective Conditioning of Diffusion Models for Monocular Depth Estimation Discovered by AI
2024
A monocular depth estimation model that leverages a diffusion backbone conditioned on embeddings from a pre-trained Vision Transformer (ViT) model to improve zero-shot transfer and capture fine-grained details.
depthdiffusionvitzero-shot -
GeoWizard: Unleashing the Diffusion Priors for 3D Geometry Estimation from a Single Image Discovered by AI
2024
A generative foundation model designed for estimating depth and surface normals from a single image by leveraging pre-trained diffusion priors. It achieves high-quality zero-shot generalization and captures fine geometric details.
depthdiffusionnormal-map -
Lotus Discovered by AI
2024
Lotus is a diffusion-based visual foundation model designed for high-quality dense geometry prediction, including zero-shot depth and surface normal estimation. It optimizes the diffusion formulation specifically for dense prediction tasks, achieving state-of-the-art performance with minimal training data.
depth-estimationdiffusion-modelnormal-map -
Metric3D v2 Discovered by AI
2024
A versatile geometric foundation model for zero-shot metric depth and surface normal estimation from a single image, enabling accurate 3D reconstruction in the wild.
depthfoundation-modelmetric-depthnormal-map -
MoGe: Unlocking Accurate Monocular Geometry Estimation for Open-Domain Images with Optimal Training Supervision Discovered by AI
2024
A powerful model from Microsoft that directly predicts a 3D point map, depth map, and surface normal map from a single open-domain image, utilizing novel global and local geometry supervisions.
depthnormal-mappoint-cloud -
Prompting Depth Anything for 4K Resolution Accurate Metric Depth Estimation Discovered by AI
2024
A novel paradigm that prompts depth foundation models with low-cost sparse LiDAR data to guide Depth Anything models, achieving highly accurate metric depth estimation at up to 4K resolution.
4kdepthlidarmetric-depth -
Scalable Autoregressive Monocular Depth Estimation Discovered by AI
2024
An autoregressive model for monocular depth estimation that treats depth maps of different resolutions as tokens, performing low-to-high resolution autoregressive prediction.
autoregressivedepthmulti-resolution -
SharpDepth: Sharpening Metric Depth Predictions Using Diffusion Distillation Discovered by AI
2024
SharpDepth is a monocular metric depth estimation framework that combines the metric accuracy of discriminative models with the sharp, fine-grained boundary details of generative diffusion-based methods. It utilizes a diffusion distillation process to achieve high-fidelity depth maps with crisp edges without the heavy computational overhead of multi-step denoising.
depth-estimationdiffusion-distillationmetric-depth -
StableNormal Discovered by AI
2024
A surface normal estimation framework designed to reduce the stochastic variance of diffusion-based estimators, producing stable, sharp, and highly accurate normal maps from monocular colored inputs.
diffusion-priornormal-mapzero-shot -
UniDepth: Universal Monocular Metric Depth Estimation Discovered by AI
2024
A universal monocular metric depth estimation model capable of reconstructing metric 3D scenes from a single image across diverse domains without relying on camera intrinsics.
3d-reconstructiondepthmetric-depth -
Marigold Discovered by AI
2023
A state-of-the-art affine-invariant monocular depth and surface normal estimation framework that repurposes the rich generative priors of Stable Diffusion, achieving exceptional zero-shot generalization on unseen real-world images.
depthdiffusion-priornormal-map -
PatchFusion Discovered by AI
2023
PatchFusion is an end-to-end tile-based framework designed for high-resolution monocular metric depth estimation. It processes images patch-wise and fuses them globally to preserve high-frequency details and prevent error propagation, making it ideal for high-resolution consumer camera inputs.
depth-estimationhigh-resolutionmetric-depth -
ZoeDepth Discovered by AI
2023
A monocular depth estimation framework that combines relative and metric depth estimation to achieve both strong zero-shot generalization and accurate metric scale prediction. It uses a framework of metric bins module that dynamically adjusts to the scene's depth range.
depthmetric-depthzero-shot
