Depth Anything V3 (Multi-View)

DepthAnythingV3_MultiView

Multi-view Depth Anything V3 - processes multiple images TOGETHER with cross-view attention. Key difference from standard nodes: - Standard: Processes images one-by-one (sequential, independent) - Multi-view: Processes all images together (cross-attention, geometrically consistent) Use this for: - Video frames (temporal consistency) - Multiple angles of same scene (SfM/reconstruction) - Stereo pairs (left/right cameras) **Normalization Modes:** - Standard: Original V3 min-max normalization (0-1 range) - V2-Style: Disparity-based with content-aware contrast (default, best for ControlNet) - Sky appears BLACK, content-only normalization - Contribution by Ltamann (TBG) - Raw: No normalization, outputs metric depth (for 3D reconstruction) **Optional Inputs:** - resize_method: How to handle patch size alignment (resize/crop/pad) - invert_depth: Toggle output convention. OFF (default): close=bright. ON: far=bright. - keep_model_size: Keep model's native output size instead of resizing back (intrinsics stay accurate) Input: Batch of images [N, H, W, 3] Outputs (all normalized across views together for consistency): - depth: Batch of consistent depth maps [N, H, W, 3] - confidence: Confidence maps [N, H, W, 3] - ray_origin: Ray origin maps (for 3D, normalized for visualization) - ray_direction: Ray direction maps (for 3D, normalized for visualization) - extrinsics: Predicted camera poses for each view (JSON) - intrinsics: Camera intrinsics for each view (JSON) - auto-scaled if resized - sky_mask: Sky segmentation [N, H, W] (Mono/Metric/Nested only) - resized_rgb_image: RGB images matching depth output dimensions - gaussian_ply_path: Path to raw 3D Gaussians PLY (Giant model only, empty string if not supported) Note: All images must have the same resolution. Higher N = more VRAM usage but better consistency.

Pack: ComfyUI-DepthAnythingV3

custom_nodes.ComfyUI-DepthAnythingV3