Fuser Apps are here 🚀
Fuser Apps are here. Free generations for the next month 💫
Let's GoDepth Anything V2
byTikTok / ByteDance & HKU
Extract high-contrast relative depth maps with razor-sharp silhouette boundaries and clean structural separation
How Depth Anything V2 works
Convert any 2D image into clean, high-resolution relative depth conditioning in three simple steps.
Supply source scene
Provide any clear 2D photograph, render, or illustration with defined spatial depth and foreground contrast.
Extract relative geometry
The vision transformer analyses spatial contours and converts scene distance into a high-contrast grayscale field.
Export depth conditioning
Feed the clean depth map directly into ControlNet conditioning pipelines, 3D parallax plates, or post-production blur shaders.
What Depth Anything V2 is good at
Engineered with a distilled vision transformer backbone that delivers clean silhouette isolation without haloing or diffusion latency.
Sub-pixel silhouette boundaries
Isolates complex contours like delicate foliage, fine hair strands, and slender furniture legs without edge bleeding, haloing, or artifact smearing.
Continuous spatial gradients
Maps continuous distances from pure foreground white to deep background black, preserving subtle slope transitions across complex floorplans and landscapes.
Distilled transformer speed
Runs over ten times faster than diffusion-based estimators by using an efficient teacher-student architecture trained on 62 million synthetic depth labels.
ControlNet-ready luminance scale
Generates standardized grayscale depth maps calibrated for direct spatial guidance in modern diffusion workflows without manual histogram rebalancing.
Made with Depth Anything V2
A gallery of architectural, product, and cinematic compositions highlighting structural fidelity and edge precision in depth estimation.
Documentary archival plate with distinct geological separation
Commercial product study with dynamic fluid detail
Graphic album sleeve with sharp fragmented silhouettes
E-commerce product still highlighting tactile materials
Monolithic interior layout demonstrating spatial gradience
What people build with Depth Anything V2
From generative ControlNet conditioning to post-production optical blur, discover how spatial engineers and visual artists leverage relative depth.
Frequently asked questions
Depth Anything V2 is designed for extracting high-contrast relative depth maps from 2D images. It is ideal for ControlNet spatial conditioning in diffusion pipelines, post-production depth-of-field blur, foreground subject isolation, and 2.5D parallax plate generation.
Depth Anything V2 replaces noisy real-world depth annotations with synthetic data distillation, eliminating the edge haloing, boundary bleeding, and smeared silhouettes common in older models. It produces significantly cleaner structural outlines and runs over ten times faster than diffusion-based alternatives.
No, Depth Anything V2 generates relative depth maps rather than metric physical measurements. The output maps proximity as a continuous grayscale gradient from near (white) to far (black), rather than true real-world units like meters or feet.
The model struggles with mirrors, highly reflective surfaces, and transparent elements like glass or smoke. Because monocular depth estimation relies on visual cues, reflections and refractions can cause depth inversion or false holes in the resulting depth map.
Depth Anything V2 was created by researchers at TikTok/ByteDance and the University of Hong Kong (HKU), published in June 2024 by Lihe Yang and collaborators.
Try Depth Anything V2 on Fuser
Extract high-contrast relative depth maps with razor-sharp silhouette boundaries and clean structural separation