You can use the multi-view node to use the cross attention feature of the main class of models. This is done to have a more consistent depth across frames of a video.

You can reconstruct 3D point clouds!

Even from multiple views, with the option to either match them (with icp) or leave them to use the predicted camera positions.

Description

Depth Anything V3 is the latest depth estimation model that predicts spatially consistent geometry from visual inputs.

Published: November 14, 2025 Paper: Depth Anything 3: Recovering the Visual Space from Any Views

Nodes

Model Loading

Download And Load Depth Anything V3 Model - Downloads and loads model from HuggingFace

Inputs: model (variant), precision (auto/fp16/fp32/bf16)
Output: da3_model
Models auto-download to ComfyUI/models/depthanything3/

DA3 Enable Tiled Processing - Configure model for high-resolution images

Inputs: da3_model, tile_size, overlap
Output: da3_model (with tiled config)
Use for 4K+ images to avoid OOM errors

Inference Nodes

Depth Anything V3 - Unified depth estimation with multiple normalization modes

Inputs: da3_model, images, normalization_mode, camera_params (optional), resize_method, invert_depth, keep_model_size
Outputs: depth, confidence, ray_origin, ray_direction, extrinsics, intrinsics, sky_mask
Normalization Modes:
- Standard: Original V3 min-max normalization (0-1 range, includes sky in normalization)
- V2-Style (default): Disparity-based with content-aware contrast - best for ControlNet workflows
  - Sky appears BLACK (like Depth Anything V2)
  - Content-only normalization excludes sky from range calculation
  - Percentile-based contrast enhancement (1st-99th percentile)
  - Enhanced depth gradations via contrast boost
  - Subtle edge anti-aliasing for natural transitions
  - Contribution by Ltamann (TBG)
- Raw: No normalization, outputs metric depth values for 3D reconstruction/point clouds

Note: The unified node always outputs all available data. Connect only the outputs you need - unused outputs are simply ignored.

Depth Anything V3 (Multi-View) - Process multiple images with cross-view attention

Inputs: da3_model, images (batch), normalization_mode, resize_method, invert_depth
Outputs: depth, confidence, ray_origin, ray_direction, extrinsics, intrinsics, sky_mask
Normalization modes: Standard, V2-Style (default), Raw
Use for video frames or multiple angles of same scene
Normalizes depth across all views together for temporal/spatial consistency

3D Processing

DA3 to Point Cloud - Convert depth to point cloud using geometric unprojection

Inputs: depth_raw, confidence, sky_mask (optional but recommended), intrinsics (optional), source_image (optional)
Output: pointcloud
Sky filtering: When sky_mask is connected, sky pixels are automatically excluded from the point cloud
Input validation: Automatically detects and rejects normalized depth inputs (requires Raw mode)

DA3 to Point Cloud (Multi view) - Convert depth to point cloud using geometric unprojection

DA3 Save Point Cloud - Export to PLY format

DA3 to 3D Gaussians - Extract 3D Gaussian splats (placeholder)

DA3 Save 3D Gaussians - Export Gaussians to PLY

Camera Utilities

DA3 Create Camera Parameters - Create camera conditioning input

DA3 Parse Camera Pose - Parse camera parameters from JSON

Model Variants

Model	Size	Features
DA3-Small	80M	Fast, good quality
DA3-Base	220M	Balanced quality and speed
DA3-Large	350M	High quality, balanced
DA3-Giant	1.15B	Best quality, slower
DA3Mono-Large	350M	Optimized for monocular depth
DA3Metric-Large	350M	Metric depth estimation
DA3Nested-Giant-Large	1.4B	Combined model with metric scaling

Model Capabilities

Different models support different features:

Feature	Small/Base/Large/Giant	Mono-Large	Metric-Large	Nested
Sky Segmentation	❌	✅	✅	✅
Camera Conditioning	✅	❌	❌	✅
Multi-View Attention	✅	⚠️	⚠️	✅
3D Gaussians	✅*	❌	❌	✅*
Ray Maps	✅	❌	❌	✅

✅ = Fully supported
❌ = Not available (returns zeros/ignored)
⚠️ = Works but no cross-view attention benefit (images processed independently)
✅* = Requires fine-tuned model weights (placeholder in current release)

Choose your model based on needs:

Need sky masks? → Use Mono/Metric/Nested (required for V2-Style normalization)
Need camera conditioning? → Use Main series or Nested
Processing video/multi-view? → Use Main series or Nested for consistency
Single images only? → Any model works

Workflow Tips

For ControlNet Depth Workflows

Use Mono or Metric models (they provide sky segmentation)
Set normalization_mode to V2-Style (default)
Connect the depth output to your ControlNet node
Enjoy clean depth maps with proper sky handling!

For 3D Reconstruction (Point Clouds)

Use any model (Mono/Metric recommended for sky filtering)
Set normalization_mode to Raw
Connect depth → depth_raw, confidence → confidence, sky_mask → sky_mask to DA3 to Point Cloud
Sky pixels will be automatically excluded if sky_mask is connected
Important: Point cloud nodes validate input and will raise an error if normalized depth is detected (prevents incorrect 3D output)

Credits

Original Paper: Haotong Lin, Sili Chen, Jun Hao Liew, et al. (ByteDance Seed Team)
Original Implementation: PozzettiAndrea
V2-Style Normalization: Ltamann (TBG) - See TBG Takeaways: Depth Anything V3 for workflow examples
Based on: Official Depth Anything 3 repository
Inspiration: kijai's ComfyUI-Depth-Anything-V2

License

Model architecture files based on Depth Anything 3 (Apache 2.0 / CC BY-NC 4.0 depending on model).

Note: Some models (Giant, Nested) use CC BY-NC 4.0 license (non-commercial use only).