Track bundled vendor runtime sources

vendor/ComfyUI/comfy/ldm/depth_anything_3/camera.py

@@ -0,0 +1,177 @@
+"""Camera-token encoder and decoder for Depth Anything 3."""
+
+from __future__ import annotations
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from comfy.ldm.modules.attention import optimized_attention_for_device
+from .transform import affine_inverse, extri_intri_to_pose_encoding
+
+
+# -----------------------------------------------------------------------
+# Building blocks (mirror depth_anything_3.model.utils.{attention,block})
+# -----------------------------------------------------------------------
+
+
+class _Mlp(nn.Module):
+    """Standard 2-layer MLP with GELU. Matches upstream ``utils.attention.Mlp``."""
+
+    def __init__(self, in_features, hidden_features=None, out_features=None, *, device=None, dtype=None, operations=None):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = operations.Linear(in_features, hidden_features, bias=True, device=device, dtype=dtype)
+        self.fc2 = operations.Linear(hidden_features, out_features, bias=True, device=device, dtype=dtype)
+
+    def forward(self, x):
+        return self.fc2(F.gelu(self.fc1(x)))
+
+
+class _LayerScale(nn.Module):
+    """Per-channel learnable scaling. Matches upstream LayerScale."""
+
+    def __init__(self, dim, *, device=None, dtype=None):
+        super().__init__()
+        self.gamma = nn.Parameter(torch.empty(dim, device=device, dtype=dtype))
+
+    def forward(self, x):
+        return x * self.gamma.to(dtype=x.dtype, device=x.device)
+
+
+class _Attention(nn.Module):
+    """ Self-attention with fused QKV projection. Mirrors upstream utils.attention.Attention;
+    Layout matches the HF safetensors (attn.qkv.{weight,bias} and attn.proj.{weight,bias})."""
+
+    def __init__(self, dim, num_heads, *, device=None, dtype=None, operations=None):
+        super().__init__()
+        assert dim % num_heads == 0
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.qkv = operations.Linear(dim, dim * 3, bias=True, device=device, dtype=dtype)
+        self.proj = operations.Linear(dim, dim, bias=True, device=device, dtype=dtype)
+
+    def forward(self, x):
+        B, N, C = x.shape
+        qkv = self.qkv(x).reshape(B, N, 3, C)
+        q, k, v = qkv.unbind(2)                      # each (B, N, C)
+        attn_fn = optimized_attention_for_device(x.device, small_input=True)
+        out = attn_fn(q, k, v, heads=self.num_heads)
+        return self.proj(out)
+
+
+class _Block(nn.Module):
+    """Pre-norm transformer block with LayerScale. Used by :class:CameraEnc. Layout follows upstream utils.block.Block."""
+
+    def __init__(self, dim, num_heads, mlp_ratio=4, init_values=0.01, *, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.norm1 = operations.LayerNorm(dim, device=device, dtype=dtype)
+        self.attn = _Attention(dim, num_heads, device=device, dtype=dtype, operations=operations)
+        self.ls1 = _LayerScale(dim, device=device, dtype=dtype) if init_values else nn.Identity()
+        self.norm2 = operations.LayerNorm(dim, device=device, dtype=dtype)
+        self.mlp = _Mlp(in_features=dim, hidden_features=int(dim * mlp_ratio), device=device, dtype=dtype, operations=operations)
+        self.ls2 = _LayerScale(dim, device=device, dtype=dtype) if init_values else nn.Identity()
+
+    def forward(self, x):
+        x = x + self.ls1(self.attn(self.norm1(x)))
+        x = x + self.ls2(self.mlp(self.norm2(x)))
+        return x
+
+
+class CameraEnc(nn.Module):
+    """Encode per-view (extrinsics, intrinsics) into a camera token.
+
+    Maps a 9-D pose-encoding vector through a small MLP up to the backbone's
+    ``embed_dim``, then runs ``trunk_depth`` transformer blocks. The output
+    has shape ``(B, S, embed_dim)`` and is injected at block ``alt_start``
+    of the DINOv2 backbone in place of the cls token.
+
+    Parameters mirror the upstream ``cam_enc.py`` so HF weights load directly.
+    """
+
+    def __init__(
+        self,
+        dim_out: int = 1024,
+        dim_in: int = 9,
+        trunk_depth: int = 4,
+        target_dim: int = 9,
+        num_heads: int = 16,
+        mlp_ratio: int = 4,
+        init_values: float = 0.01,
+        *,
+        device=None, dtype=None, operations=None,
+        **_kwargs,
+    ):
+        super().__init__()
+        self.target_dim = target_dim
+        self.trunk_depth = trunk_depth
+        self.trunk = nn.Sequential(*[
+            _Block(dim_out, num_heads=num_heads, mlp_ratio=mlp_ratio,
+                   init_values=init_values,
+                   device=device, dtype=dtype, operations=operations)
+            for _ in range(trunk_depth)
+        ])
+        self.token_norm = operations.LayerNorm(dim_out, device=device, dtype=dtype)
+        self.trunk_norm = operations.LayerNorm(dim_out, device=device, dtype=dtype)
+        self.pose_branch = _Mlp(
+            in_features=dim_in,
+            hidden_features=dim_out // 2,
+            out_features=dim_out,
+            device=device, dtype=dtype, operations=operations,
+        )
+
+    def forward(self, extrinsics: torch.Tensor, intrinsics: torch.Tensor,
+                image_size_hw) -> torch.Tensor:
+        """Encode camera parameters into ``(B, S, dim_out)`` tokens."""
+        c2ws = affine_inverse(extrinsics)
+        pose_encoding = extri_intri_to_pose_encoding(c2ws, intrinsics, image_size_hw)
+        tokens = self.pose_branch(pose_encoding.to(self.pose_branch.fc1.weight.dtype))
+        tokens = self.token_norm(tokens)
+        tokens = self.trunk(tokens)
+        tokens = self.trunk_norm(tokens)
+        return tokens
+
+
+class CameraDec(nn.Module):
+    """Decode the final cam token into a 9-D pose encoding.
+
+    Output layout: ``[T(3), quat_xyzw(4), fov_h, fov_w]``. The translation is
+    always predicted by the network; the quaternion and FoV can either be
+    predicted or supplied via ``camera_encoding`` (used at training time
+    when GT cameras are available -- not exercised at inference here).
+
+    Parameters mirror the upstream ``cam_dec.py`` so HF weights load directly.
+    """
+
+    def __init__(self, dim_in: int = 1536,
+                 *, device=None, dtype=None, operations=None, **_kwargs):
+        super().__init__()
+        d = dim_in
+        self.backbone = nn.Sequential(
+            operations.Linear(d, d, device=device, dtype=dtype),
+            nn.ReLU(),
+            operations.Linear(d, d, device=device, dtype=dtype),
+            nn.ReLU(),
+        )
+        self.fc_t = operations.Linear(d, 3, device=device, dtype=dtype)
+        self.fc_qvec = operations.Linear(d, 4, device=device, dtype=dtype)
+        self.fc_fov = nn.Sequential(
+            operations.Linear(d, 2, device=device, dtype=dtype),
+            nn.ReLU(),
+        )
+
+    def forward(self, feat: torch.Tensor,
+                camera_encoding: "torch.Tensor | None" = None) -> torch.Tensor:
+        """Decode ``(B, N, dim_in)`` cam tokens into ``(B, N, 9)`` pose enc."""
+        B, N = feat.shape[:2]
+        feat = feat.reshape(B * N, -1)
+        feat = self.backbone(feat)
+        out_t = self.fc_t(feat.float()).reshape(B, N, 3)
+        if camera_encoding is None:
+            out_qvec = self.fc_qvec(feat.float()).reshape(B, N, 4)
+            out_fov = self.fc_fov(feat.float()).reshape(B, N, 2)
+        else:
+            out_qvec = camera_encoding[..., 3:7]
+            out_fov = camera_encoding[..., -2:]
+        return torch.cat([out_t, out_qvec, out_fov], dim=-1)

vendor/ComfyUI/comfy/ldm/depth_anything_3/dpt.py

@@ -0,0 +1,489 @@
+"""DPT / DualDPT heads for Depth Anything 3."""
+
+from __future__ import annotations
+
+from typing import List, Optional, Sequence, Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class Permute(nn.Module):
+    def __init__(self, dims: Tuple[int, ...]):
+        super().__init__()
+        self.dims = dims
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return x.permute(*self.dims)
+
+
+def _custom_interpolate(
+    x: torch.Tensor,
+    size: Optional[Tuple[int, int]] = None,
+    scale_factor: Optional[float] = None,
+    mode: str = "bilinear",
+    align_corners: bool = True,
+) -> torch.Tensor:
+    if size is None:
+        assert scale_factor is not None
+        size = (int(x.shape[-2] * scale_factor), int(x.shape[-1] * scale_factor))
+    INT_MAX = 1610612736
+    total = size[0] * size[1] * x.shape[0] * x.shape[1]
+    if total > INT_MAX:
+        chunks = torch.chunk(x, chunks=(total // INT_MAX) + 1, dim=0)
+        outs = [F.interpolate(c, size=size, mode=mode, align_corners=align_corners) for c in chunks]
+        return torch.cat(outs, dim=0).contiguous()
+    return F.interpolate(x, size=size, mode=mode, align_corners=align_corners)
+
+
+def _create_uv_grid(width: int, height: int, aspect_ratio: float, dtype, device) -> torch.Tensor:
+    """Normalised UV grid spanning (-x_span, -y_span)..(x_span, y_span)."""
+    diag_factor = (aspect_ratio ** 2 + 1.0) ** 0.5
+    span_x = aspect_ratio / diag_factor
+    span_y = 1.0 / diag_factor
+    left_x = -span_x * (width - 1) / width
+    right_x = span_x * (width - 1) / width
+    top_y = -span_y * (height - 1) / height
+    bottom_y = span_y * (height - 1) / height
+    x_coords = torch.linspace(left_x, right_x, steps=width, dtype=dtype, device=device)
+    y_coords = torch.linspace(top_y, bottom_y, steps=height, dtype=dtype, device=device)
+    uu, vv = torch.meshgrid(x_coords, y_coords, indexing="xy")
+    return torch.stack((uu, vv), dim=-1)  # (H, W, 2)
+
+
+def _make_sincos_pos_embed(embed_dim: int, pos: torch.Tensor, omega_0: float = 100.0) -> torch.Tensor:
+    omega = torch.arange(embed_dim // 2, dtype=torch.float32, device=pos.device)
+    omega = 1.0 / omega_0 ** (omega / (embed_dim / 2.0))
+    pos = pos.reshape(-1)
+    out = torch.einsum("m,d->md", pos, omega)
+    return torch.cat([out.sin(), out.cos()], dim=1).float()
+
+
+def _position_grid_to_embed(pos_grid: torch.Tensor, embed_dim: int, omega_0: float = 100.0) -> torch.Tensor:
+    H, W, _ = pos_grid.shape
+    pos_flat = pos_grid.reshape(-1, 2)
+    emb_x = _make_sincos_pos_embed(embed_dim // 2, pos_flat[:, 0], omega_0=omega_0)
+    emb_y = _make_sincos_pos_embed(embed_dim // 2, pos_flat[:, 1], omega_0=omega_0)
+    emb = torch.cat([emb_x, emb_y], dim=-1)
+    return emb.view(H, W, embed_dim)
+
+
+def _add_pos_embed(x: torch.Tensor, W: int, H: int, ratio: float = 0.1) -> torch.Tensor:
+    """Stateless UV positional embedding added to a feature map (B, C, h, w)."""
+    pw, ph = x.shape[-1], x.shape[-2]
+    pe = _create_uv_grid(pw, ph, aspect_ratio=W / H, dtype=x.dtype, device=x.device)
+    pe = _position_grid_to_embed(pe, x.shape[1]) * ratio
+    pe = pe.permute(2, 0, 1)[None].expand(x.shape[0], -1, -1, -1).to(dtype=x.dtype)
+    return x + pe
+
+
+def _apply_activation(x: torch.Tensor, activation: str) -> torch.Tensor:
+    act = (activation or "linear").lower()
+    if act == "exp":
+        return torch.exp(x)
+    if act == "expp1":
+        return torch.exp(x) + 1
+    if act == "expm1":
+        return torch.expm1(x)
+    if act == "relu":
+        return torch.relu(x)
+    if act == "sigmoid":
+        return torch.sigmoid(x)
+    if act == "softplus":
+        return F.softplus(x)
+    if act == "tanh":
+        return torch.tanh(x)
+    return x
+
+
+# -----------------------------------------------------------------------------
+# Fusion building blocks
+# -----------------------------------------------------------------------------
+
+
+class ResidualConvUnit(nn.Module):
+    def __init__(self, features: int, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.conv1 = operations.Conv2d(features, features, 3, 1, 1, bias=True, device=device, dtype=dtype)
+        self.conv2 = operations.Conv2d(features, features, 3, 1, 1, bias=True, device=device, dtype=dtype)
+        self.activation = nn.ReLU(inplace=False)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        out = self.activation(x)
+        out = self.conv1(out)
+        out = self.activation(out)
+        out = self.conv2(out)
+        return out + x
+
+
+class FeatureFusionBlock(nn.Module):
+    def __init__(self, features: int, has_residual: bool = True, align_corners: bool = True, device=None, dtype=None, operations=None):
+        super().__init__()
+        self.align_corners = align_corners
+        self.has_residual = has_residual
+        if has_residual:
+            self.resConfUnit1 = ResidualConvUnit(features, device=device, dtype=dtype, operations=operations)
+        else:
+            self.resConfUnit1 = None
+        self.resConfUnit2 = ResidualConvUnit(features, device=device, dtype=dtype, operations=operations)
+        self.out_conv = operations.Conv2d(features, features, 1, 1, 0, bias=True, device=device, dtype=dtype)
+
+    def forward(self, *xs: torch.Tensor, size: Optional[Tuple[int, int]] = None) -> torch.Tensor:
+        y = xs[0]
+        if self.has_residual and len(xs) > 1 and self.resConfUnit1 is not None:
+            y = y + self.resConfUnit1(xs[1])
+        y = self.resConfUnit2(y)
+        if size is None:
+            up_kwargs = {"scale_factor": 2.0}
+        else:
+            up_kwargs = {"size": size}
+        y = _custom_interpolate(y, **up_kwargs, mode="bilinear", align_corners=self.align_corners)
+        y = self.out_conv(y)
+        return y
+
+
+class _Scratch(nn.Module):
+    """Container that mirrors upstream ``scratch`` attribute layout."""
+
+
+def _make_scratch(in_shape: List[int], out_shape: int, device=None, dtype=None, operations=None) -> _Scratch:
+    scratch = _Scratch()
+    scratch.layer1_rn = operations.Conv2d(in_shape[0], out_shape, 3, 1, 1, bias=False, device=device, dtype=dtype)
+    scratch.layer2_rn = operations.Conv2d(in_shape[1], out_shape, 3, 1, 1, bias=False, device=device, dtype=dtype)
+    scratch.layer3_rn = operations.Conv2d(in_shape[2], out_shape, 3, 1, 1, bias=False, device=device, dtype=dtype)
+    scratch.layer4_rn = operations.Conv2d(in_shape[3], out_shape, 3, 1, 1, bias=False, device=device, dtype=dtype)
+    return scratch
+
+
+def _make_fusion_block(features: int, has_residual: bool = True, device=None, dtype=None, operations=None) -> FeatureFusionBlock:
+    return FeatureFusionBlock(features, has_residual=has_residual, align_corners=True, device=device, dtype=dtype, operations=operations)
+
+
+# -----------------------------------------------------------------------------
+# DPT (single head + optional sky head) -- used by DA3Mono/Metric
+# -----------------------------------------------------------------------------
+
+
+class DPT(nn.Module):
+    """Single-head DPT used by DA3Mono-Large and DA3Metric-Large."""
+
+    def __init__(
+        self,
+        dim_in: int,
+        patch_size: int = 14,
+        output_dim: int = 1,
+        activation: str = "exp",
+        conf_activation: str = "expp1",
+        features: int = 256,
+        out_channels: Sequence[int] = (256, 512, 1024, 1024),
+        pos_embed: bool = False,
+        down_ratio: int = 1,
+        head_name: str = "depth",
+        use_sky_head: bool = True,
+        sky_name: str = "sky",
+        sky_activation: str = "relu",
+        norm_type: str = "idt",
+        device=None, dtype=None, operations=None,
+    ):
+        super().__init__()
+        self.patch_size = patch_size
+        self.activation = activation
+        self.conf_activation = conf_activation
+        self.pos_embed = pos_embed
+        self.down_ratio = down_ratio
+        self.head_main = head_name
+        self.sky_name = sky_name
+        self.out_dim = output_dim
+        self.has_conf = output_dim > 1
+        self.use_sky_head = use_sky_head
+        self.sky_activation = sky_activation
+        self.intermediate_layer_idx: Tuple[int, int, int, int] = (0, 1, 2, 3)
+
+        if norm_type == "layer":
+            self.norm = operations.LayerNorm(dim_in, device=device, dtype=dtype)
+        else:
+            self.norm = nn.Identity()
+
+        out_channels = list(out_channels)
+        self.projects = nn.ModuleList([
+            operations.Conv2d(dim_in, oc, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype)
+            for oc in out_channels
+        ])
+        self.resize_layers = nn.ModuleList([
+            operations.ConvTranspose2d(out_channels[0], out_channels[0], kernel_size=4, stride=4, padding=0, device=device, dtype=dtype),
+            operations.ConvTranspose2d(out_channels[1], out_channels[1], kernel_size=2, stride=2, padding=0, device=device, dtype=dtype),
+            nn.Identity(),
+            operations.Conv2d(out_channels[3], out_channels[3], kernel_size=3, stride=2, padding=1, device=device, dtype=dtype),
+        ])
+
+        self.scratch = _make_scratch(out_channels, features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet1 = _make_fusion_block(features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet2 = _make_fusion_block(features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet3 = _make_fusion_block(features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet4 = _make_fusion_block(features, has_residual=False, device=device, dtype=dtype, operations=operations)
+
+        head_features_1 = features
+        head_features_2 = 32
+        self.scratch.output_conv1 = operations.Conv2d(
+            head_features_1, head_features_1 // 2, kernel_size=3, stride=1, padding=1,
+            device=device, dtype=dtype,
+        )
+        self.scratch.output_conv2 = nn.Sequential(
+            operations.Conv2d(head_features_1 // 2, head_features_2, kernel_size=3, stride=1, padding=1, device=device, dtype=dtype),
+            nn.ReLU(inplace=False),
+            operations.Conv2d(head_features_2, output_dim, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype),
+        )
+
+        if self.use_sky_head:
+            self.scratch.sky_output_conv2 = nn.Sequential(
+                operations.Conv2d(head_features_1 // 2, head_features_2, kernel_size=3, stride=1, padding=1, device=device, dtype=dtype),
+                nn.ReLU(inplace=False),
+                operations.Conv2d(head_features_2, 1, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype),
+            )
+
+    def forward(self, feats: List[torch.Tensor], H: int, W: int, patch_start_idx: int = 0, **_kwargs) -> dict:
+        # feats[i][0] is the patch-token tensor with shape (B, S, N_patch, C)
+        B, S, N, C = feats[0][0].shape
+        feats_flat = [feat[0].reshape(B * S, N, C) for feat in feats]
+
+        ph, pw = H // self.patch_size, W // self.patch_size
+        resized = []
+        for stage_idx, take_idx in enumerate(self.intermediate_layer_idx):
+            x = feats_flat[take_idx][:, patch_start_idx:]
+            x = self.norm(x)
+            x = x.permute(0, 2, 1).contiguous().reshape(B * S, C, ph, pw)
+            x = self.projects[stage_idx](x)
+            if self.pos_embed:
+                x = _add_pos_embed(x, W, H)
+            x = self.resize_layers[stage_idx](x)
+            resized.append(x)
+
+        l1_rn = self.scratch.layer1_rn(resized[0])
+        l2_rn = self.scratch.layer2_rn(resized[1])
+        l3_rn = self.scratch.layer3_rn(resized[2])
+        l4_rn = self.scratch.layer4_rn(resized[3])
+
+        out = self.scratch.refinenet4(l4_rn, size=l3_rn.shape[2:])
+        out = self.scratch.refinenet3(out, l3_rn, size=l2_rn.shape[2:])
+        out = self.scratch.refinenet2(out, l2_rn, size=l1_rn.shape[2:])
+        out = self.scratch.refinenet1(out, l1_rn)
+
+        h_out = int(ph * self.patch_size / self.down_ratio)
+        w_out = int(pw * self.patch_size / self.down_ratio)
+
+        fused = self.scratch.output_conv1(out)
+        fused = _custom_interpolate(fused, (h_out, w_out), mode="bilinear", align_corners=True)
+        if self.pos_embed:
+            fused = _add_pos_embed(fused, W, H)
+        feat = fused
+
+        main_logits = self.scratch.output_conv2(feat)
+        outs = {}
+        if self.has_conf:
+            fmap = main_logits.permute(0, 2, 3, 1)
+            pred = _apply_activation(fmap[..., :-1], self.activation)
+            conf = _apply_activation(fmap[..., -1], self.conf_activation)
+            outs[self.head_main] = pred.squeeze(-1).view(B, S, *pred.shape[1:-1])
+            outs[f"{self.head_main}_conf"] = conf.view(B, S, *conf.shape[1:])
+        else:
+            pred = _apply_activation(main_logits, self.activation)
+            outs[self.head_main] = pred.squeeze(1).view(B, S, *pred.shape[2:])
+
+        if self.use_sky_head:
+            sky_logits = self.scratch.sky_output_conv2(feat)
+            if self.sky_activation.lower() == "sigmoid":
+                sky = torch.sigmoid(sky_logits)
+            elif self.sky_activation.lower() == "relu":
+                sky = F.relu(sky_logits)
+            else:
+                sky = sky_logits
+            outs[self.sky_name] = sky.squeeze(1).view(B, S, *sky.shape[2:])
+
+        return outs
+
+
+# -----------------------------------------------------------------------------
+# DualDPT (depth + auxiliary "ray" head) -- used by DA3-Small / DA3-Base
+# -----------------------------------------------------------------------------
+
+
+class DualDPT(nn.Module):
+    """Two-head DPT used by DA3-Small / DA3-Base."""
+
+    def __init__(
+        self,
+        dim_in: int,
+        patch_size: int = 14,
+        output_dim: int = 2,
+        activation: str = "exp",
+        conf_activation: str = "expp1",
+        features: int = 256,
+        out_channels: Sequence[int] = (256, 512, 1024, 1024),
+        pos_embed: bool = True,
+        down_ratio: int = 1,
+        aux_pyramid_levels: int = 4,
+        aux_out1_conv_num: int = 5,
+        head_names: Tuple[str, str] = ("depth", "ray"),
+        device=None, dtype=None, operations=None,
+    ):
+        super().__init__()
+        self.patch_size = patch_size
+        self.activation = activation
+        self.conf_activation = conf_activation
+        self.pos_embed = pos_embed
+        self.down_ratio = down_ratio
+        self.aux_levels = aux_pyramid_levels
+        self.aux_out1_conv_num = aux_out1_conv_num
+        self.head_main, self.head_aux = head_names
+        self.intermediate_layer_idx: Tuple[int, int, int, int] = (0, 1, 2, 3)
+        # Toggle the auxiliary ray branch at runtime. Default off (mono path).
+        # DepthAnything3Net flips this on when running multi-view + ray-pose.
+        self.enable_aux: bool = False
+
+        self.norm = operations.LayerNorm(dim_in, device=device, dtype=dtype)
+        out_channels = list(out_channels)
+        self.projects = nn.ModuleList([
+            operations.Conv2d(dim_in, oc, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype)
+            for oc in out_channels
+        ])
+        self.resize_layers = nn.ModuleList([
+            operations.ConvTranspose2d(out_channels[0], out_channels[0], kernel_size=4, stride=4, padding=0, device=device, dtype=dtype),
+            operations.ConvTranspose2d(out_channels[1], out_channels[1], kernel_size=2, stride=2, padding=0, device=device, dtype=dtype),
+            nn.Identity(),
+            operations.Conv2d(out_channels[3], out_channels[3], kernel_size=3, stride=2, padding=1, device=device, dtype=dtype),
+        ])
+
+        self.scratch = _make_scratch(out_channels, features, device=device, dtype=dtype, operations=operations)
+        # Main fusion chain
+        self.scratch.refinenet1 = _make_fusion_block(features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet2 = _make_fusion_block(features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet3 = _make_fusion_block(features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet4 = _make_fusion_block(features, has_residual=False, device=device, dtype=dtype, operations=operations)
+        # Auxiliary fusion chain (separate copies)
+        self.scratch.refinenet1_aux = _make_fusion_block(features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet2_aux = _make_fusion_block(features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet3_aux = _make_fusion_block(features, device=device, dtype=dtype, operations=operations)
+        self.scratch.refinenet4_aux = _make_fusion_block(features, has_residual=False, device=device, dtype=dtype, operations=operations)
+
+        head_features_1 = features
+        head_features_2 = 32
+
+        # Main head neck + final projection
+        self.scratch.output_conv1 = operations.Conv2d(
+            head_features_1, head_features_1 // 2, kernel_size=3, stride=1, padding=1,
+            device=device, dtype=dtype,
+        )
+        self.scratch.output_conv2 = nn.Sequential(
+            operations.Conv2d(head_features_1 // 2, head_features_2, kernel_size=3, stride=1, padding=1, device=device, dtype=dtype),
+            nn.ReLU(inplace=False),
+            operations.Conv2d(head_features_2, output_dim, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype),
+        )
+
+        # Aux pre-head per level (multi-level pyramid)
+        self.scratch.output_conv1_aux = nn.ModuleList([
+            self._make_aux_out1_block(head_features_1, device=device, dtype=dtype, operations=operations)
+            for _ in range(self.aux_levels)
+        ])
+
+        # Aux final projection per level (includes LayerNorm permute path).
+        ln_seq = [Permute((0, 2, 3, 1)),
+                  operations.LayerNorm(head_features_2, device=device, dtype=dtype),
+                  Permute((0, 3, 1, 2))]
+        self.scratch.output_conv2_aux = nn.ModuleList([
+            nn.Sequential(
+                operations.Conv2d(head_features_1 // 2, head_features_2, kernel_size=3, stride=1, padding=1, device=device, dtype=dtype),
+                *ln_seq,
+                nn.ReLU(inplace=False),
+                operations.Conv2d(head_features_2, 7, kernel_size=1, stride=1, padding=0, device=device, dtype=dtype),
+            )
+            for _ in range(self.aux_levels)
+        ])
+
+    @staticmethod
+    def _make_aux_out1_block(in_ch: int, *, device=None, dtype=None, operations=None) -> nn.Sequential:
+        # aux_out1_conv_num=5 in all Apache-2.0 variants.
+        return nn.Sequential(
+            operations.Conv2d(in_ch, in_ch // 2, 3, 1, 1, device=device, dtype=dtype),
+            operations.Conv2d(in_ch // 2, in_ch, 3, 1, 1, device=device, dtype=dtype),
+            operations.Conv2d(in_ch, in_ch // 2, 3, 1, 1, device=device, dtype=dtype),
+            operations.Conv2d(in_ch // 2, in_ch, 3, 1, 1, device=device, dtype=dtype),
+            operations.Conv2d(in_ch, in_ch // 2, 3, 1, 1, device=device, dtype=dtype),
+        )
+
+    def forward(self, feats: List[torch.Tensor], H: int, W: int, patch_start_idx: int = 0, **_kwargs) -> dict:
+        B, S, N, C = feats[0][0].shape
+        feats_flat = [feat[0].reshape(B * S, N, C) for feat in feats]
+
+        ph, pw = H // self.patch_size, W // self.patch_size
+        resized = []
+        for stage_idx, take_idx in enumerate(self.intermediate_layer_idx):
+            x = feats_flat[take_idx][:, patch_start_idx:]
+            x = self.norm(x)
+            x = x.permute(0, 2, 1).contiguous().reshape(B * S, C, ph, pw)
+            x = self.projects[stage_idx](x)
+            if self.pos_embed:
+                x = _add_pos_embed(x, W, H)
+            x = self.resize_layers[stage_idx](x)
+            resized.append(x)
+
+        l1_rn = self.scratch.layer1_rn(resized[0])
+        l2_rn = self.scratch.layer2_rn(resized[1])
+        l3_rn = self.scratch.layer3_rn(resized[2])
+        l4_rn = self.scratch.layer4_rn(resized[3])
+
+        # Main pyramid (output_conv1 is applied inside the upstream `_fuse`,
+        # before interpolation -- replicate that order here).
+        m = self.scratch.refinenet4(l4_rn, size=l3_rn.shape[2:])
+        if self.enable_aux:
+            a4 = self.scratch.refinenet4_aux(l4_rn, size=l3_rn.shape[2:])
+            aux_pyr = [a4]
+        m = self.scratch.refinenet3(m, l3_rn, size=l2_rn.shape[2:])
+        if self.enable_aux:
+            aux_pyr.append(self.scratch.refinenet3_aux(aux_pyr[-1], l3_rn, size=l2_rn.shape[2:]))
+        m = self.scratch.refinenet2(m, l2_rn, size=l1_rn.shape[2:])
+        if self.enable_aux:
+            aux_pyr.append(self.scratch.refinenet2_aux(aux_pyr[-1], l2_rn, size=l1_rn.shape[2:]))
+        m = self.scratch.refinenet1(m, l1_rn)
+        if self.enable_aux:
+            aux_pyr.append(self.scratch.refinenet1_aux(aux_pyr[-1], l1_rn))
+        m = self.scratch.output_conv1(m)
+
+        h_out = int(ph * self.patch_size / self.down_ratio)
+        w_out = int(pw * self.patch_size / self.down_ratio)
+
+        m = _custom_interpolate(m, (h_out, w_out), mode="bilinear", align_corners=True)
+        if self.pos_embed:
+            m = _add_pos_embed(m, W, H)
+        main_logits = self.scratch.output_conv2(m)
+        fmap = main_logits.permute(0, 2, 3, 1)
+        depth_pred = _apply_activation(fmap[..., :-1], self.activation)
+        depth_conf = _apply_activation(fmap[..., -1], self.conf_activation)
+
+        outs = {
+            self.head_main: depth_pred.squeeze(-1).view(B, S, *depth_pred.shape[1:-1]),
+            f"{self.head_main}_conf": depth_conf.view(B, S, *depth_conf.shape[1:]),
+        }
+
+        if self.enable_aux:
+            # Auxiliary "ray" head (multi-level inside) -- only the last level
+            # is returned. Mirrors upstream ``DualDPT._fuse`` + ``_forward_impl``:
+            # each aux pyramid level goes through ``output_conv1_aux[i]``
+            # (5-layer conv stack that ends at ``features // 2`` channels),
+            # then the last level optionally gets a pos-embed and finally
+            # ``output_conv2_aux[-1]``.
+            aux_processed = [
+                self.scratch.output_conv1_aux[i](a) for i, a in enumerate(aux_pyr)
+            ]
+            last_aux = aux_processed[-1]
+            if self.pos_embed:
+                last_aux = _add_pos_embed(last_aux, W, H)
+            last_aux_logits = self.scratch.output_conv2_aux[-1](last_aux)
+            fmap_last = last_aux_logits.permute(0, 2, 3, 1)
+            # Channels: [ray(6), ray_conf(1)]; ray uses 'linear' activation.
+            aux_pred = fmap_last[..., :-1]
+            aux_conf = _apply_activation(fmap_last[..., -1], self.conf_activation)
+            outs[self.head_aux] = aux_pred.view(B, S, *aux_pred.shape[1:])
+            outs[f"{self.head_aux}_conf"] = aux_conf.view(B, S, *aux_conf.shape[1:])
+
+        return outs

vendor/ComfyUI/comfy/ldm/depth_anything_3/model.py

@@ -0,0 +1,236 @@
+from __future__ import annotations
+
+from typing import Dict, Optional, Sequence
+
+import torch
+import torch.nn as nn
+
+from comfy.image_encoders.dino2 import Dinov2Model
+
+from .camera import CameraDec, CameraEnc
+from .dpt import DPT, DualDPT
+from .ray_pose import get_extrinsic_from_camray
+from .transform import affine_inverse, pose_encoding_to_extri_intri
+
+
+_HEAD_REGISTRY = {
+    "dpt": DPT,
+    "dualdpt": DualDPT,
+}
+
+
+# Backbone presets (mirror the upstream DINOv2 ViT variants).
+_BACKBONE_PRESETS = {
+    "vits": dict(hidden_size=384,  num_hidden_layers=12, num_attention_heads=6,  use_swiglu_ffn=False),
+    "vitb": dict(hidden_size=768,  num_hidden_layers=12, num_attention_heads=12, use_swiglu_ffn=False),
+    "vitl": dict(hidden_size=1024, num_hidden_layers=24, num_attention_heads=16, use_swiglu_ffn=False),
+    "vitg": dict(hidden_size=1536, num_hidden_layers=40, num_attention_heads=24, use_swiglu_ffn=True),
+}
+
+
+def _build_backbone_config(
+    backbone_name: str,
+    *,
+    alt_start: int,
+    qknorm_start: int,
+    rope_start: int,
+    cat_token: bool,
+) -> dict:
+    if backbone_name not in _BACKBONE_PRESETS:
+        raise ValueError(f"Unknown DINOv2 backbone variant: {backbone_name!r}")
+    cfg = dict(_BACKBONE_PRESETS[backbone_name])
+    cfg.update(dict(
+        layer_norm_eps=1e-6,
+        patch_size=14,
+        image_size=518,
+        # No mask_token in DA3 weights; omit param to avoid load warnings.
+        use_mask_token=False,
+        alt_start=alt_start,
+        qknorm_start=qknorm_start,
+        rope_start=rope_start,
+        cat_token=cat_token,
+        rope_freq=100.0,
+    ))
+    return cfg
+
+
+class DepthAnything3Net(nn.Module):
+
+    PATCH_SIZE = 14
+
+    def __init__(
+        self,
+        # --- Backbone ---
+        backbone_name: str = "vitl",
+        out_layers: Sequence[int] = (4, 11, 17, 23),
+        alt_start: int = -1,
+        qknorm_start: int = -1,
+        rope_start: int = -1,
+        cat_token: bool = False,
+        # --- Head ---
+        head_type: str = "dpt",  # dpt or dualdpt
+        head_dim_in: int = 1024,
+        head_output_dim: int = 1,  # 1 = depth only, 2 = depth+conf
+        head_features: int = 256,
+        head_out_channels: Sequence[int] = (256, 512, 1024, 1024),
+        head_use_sky_head: bool = True, # ignored by DualDPT
+        head_pos_embed: Optional[bool] = None,  # default: True for DualDPT, False for DPT
+        # --- Camera (multi-view) ---
+        has_cam_enc: bool = False,
+        has_cam_dec: bool = False,
+        cam_dim_out: Optional[int] = None,  # CameraEnc dim_out (defaults to embed_dim)
+        cam_dec_dim_in: Optional[int] = None,  # CameraDec dim_in  (defaults to 2*embed_dim with cat_token)
+        # ComfyUI plumbing
+        device=None, dtype=None, operations=None,
+        **_ignored,
+    ):
+        super().__init__()
+        head_cls = _HEAD_REGISTRY[head_type.lower()]
+        self.head_type = head_type.lower()
+        self.has_sky = (self.head_type == "dpt") and head_use_sky_head
+        self.has_conf = head_output_dim > 1
+        self.out_layers = list(out_layers)
+
+        backbone_cfg = _build_backbone_config(
+            backbone_name,
+            alt_start=alt_start,
+            qknorm_start=qknorm_start,
+            rope_start=rope_start,
+            cat_token=cat_token,
+        )
+        self.backbone = Dinov2Model(backbone_cfg, dtype, device, operations)
+
+        head_kwargs = dict(
+            dim_in=head_dim_in,
+            patch_size=self.PATCH_SIZE,
+            output_dim=head_output_dim,
+            features=head_features,
+            out_channels=tuple(head_out_channels),
+            device=device, dtype=dtype, operations=operations,
+        )
+        if self.head_type == "dpt":
+            head_kwargs.update(
+                use_sky_head=head_use_sky_head,
+                pos_embed=(False if head_pos_embed is None else head_pos_embed),
+            )
+        else:  # dualdpt
+            head_kwargs.update(
+                pos_embed=(True if head_pos_embed is None else head_pos_embed),
+            )
+        self.head = head_cls(**head_kwargs)
+
+        # Built only if checkpoint has weights; cam_enc output dim == embed_dim.
+        embed_dim = backbone_cfg["hidden_size"]
+        if has_cam_enc:
+            self.cam_enc = CameraEnc(
+                dim_out=cam_dim_out if cam_dim_out is not None else embed_dim,
+                num_heads=max(1, embed_dim // 64),
+                device=device, dtype=dtype, operations=operations,
+            )
+        else:
+            self.cam_enc = None
+        if has_cam_dec:
+            default_dim = embed_dim * (2 if cat_token else 1)
+            self.cam_dec = CameraDec(
+                dim_in=cam_dec_dim_in if cam_dec_dim_in is not None else default_dim,
+                device=device, dtype=dtype, operations=operations,
+            )
+        else:
+            self.cam_dec = None
+
+        self.dtype = dtype
+
+    def forward(
+        self,
+        image: torch.Tensor,
+        extrinsics: Optional[torch.Tensor] = None,
+        intrinsics: Optional[torch.Tensor] = None,
+        *,
+        use_ray_pose: bool = False,
+        ref_view_strategy: str = "saddle_balanced",
+        export_feat_layers: Optional[Sequence[int]] = None,
+        **_unused,
+    ) -> Dict[str, torch.Tensor]:
+        """Run depth and optionally pose prediction."""
+        if image.ndim == 4:
+            image = image.unsqueeze(1)  # (B, 1, 3, H, W)
+        assert image.ndim == 5 and image.shape[2] == 3, \
+            f"image must be (B,3,H,W) or (B,S,3,H,W); got {tuple(image.shape)}"
+
+        B, S, _, H, W = image.shape
+        assert H % self.PATCH_SIZE == 0 and W % self.PATCH_SIZE == 0, \
+            f"image H,W must be multiples of {self.PATCH_SIZE}; got {(H, W)}"
+
+        # Camera-token preparation (multi-view path).
+        cam_token = None
+        if extrinsics is not None and intrinsics is not None and self.cam_enc is not None:
+            cam_token = self.cam_enc(extrinsics, intrinsics, (H, W))
+
+        # Toggle aux ray output on/off depending on what the caller asked for.
+        if isinstance(self.head, DualDPT):
+            self.head.enable_aux = bool(use_ray_pose)
+
+        feats, aux_feats = self.backbone.get_intermediate_layers_da3(
+            image, self.out_layers, cam_token=cam_token,
+            ref_view_strategy=ref_view_strategy,
+            export_feat_layers=export_feat_layers,
+        )
+        head_out = self.head(feats, H=H, W=W, patch_start_idx=0)
+
+        # Pose prediction.
+        out: Dict[str, torch.Tensor] = {}
+        if use_ray_pose and "ray" in head_out and "ray_conf" in head_out:
+            ray = head_out["ray"]
+            ray_conf = head_out["ray_conf"]
+            extr_c2w, focal, pp = get_extrinsic_from_camray(
+                ray, ray_conf, ray.shape[-3], ray.shape[-2],
+            )
+            # Match the upstream output: w2c, drop the homogeneous row.
+            extr_w2c = affine_inverse(extr_c2w)[:, :, :3, :]
+            # Build pixel-space intrinsics from the normalised focal/pp output.
+            intr = torch.eye(3, device=ray.device, dtype=ray.dtype)
+            intr = intr[None, None].expand(extr_c2w.shape[0], extr_c2w.shape[1], 3, 3).clone()
+            intr[:, :, 0, 0] = focal[:, :, 0] / 2 * W
+            intr[:, :, 1, 1] = focal[:, :, 1] / 2 * H
+            intr[:, :, 0, 2] = pp[:, :, 0] * W * 0.5
+            intr[:, :, 1, 2] = pp[:, :, 1] * H * 0.5
+            out["extrinsics"] = extr_w2c
+            out["intrinsics"] = intr
+        elif self.cam_dec is not None and S > 1:
+            # Decode the cam-token of the final out_layer into a pose encoding.
+            cam_feat = feats[-1][1]  # (B, S, dim_in_to_cam_dec)
+            pose_enc = self.cam_dec(cam_feat)
+            c2w_3x4, intr = pose_encoding_to_extri_intri(pose_enc, (H, W))
+            # Match the upstream output convention: w2c (world->camera), 3x4.
+            c2w_4x4 = torch.cat([
+                c2w_3x4,
+                torch.tensor([0, 0, 0, 1], device=c2w_3x4.device, dtype=c2w_3x4.dtype)
+                    .view(1, 1, 1, 4).expand(B, S, 1, 4),
+            ], dim=-2)
+            out["extrinsics"] = affine_inverse(c2w_4x4)[:, :, :3, :]
+            out["intrinsics"] = intr
+
+        # Flatten the views axis for per-pixel outputs (depth/conf/sky) so the
+        # per-image consumer keeps its (B*S, H, W) interface.
+        for k, v in head_out.items():
+            if k in ("ray", "ray_conf"):
+                # Keep multi-view shape for downstream pose work.
+                out[k] = v
+            elif v.ndim >= 3 and v.shape[0] == B and v.shape[1] == S:
+                out[k] = v.reshape(B * S, *v.shape[2:])
+            else:
+                out[k] = v
+
+        if export_feat_layers:
+            out["aux_features"] = self._reshape_aux_features(aux_feats, H, W)
+        return out
+
+    def _reshape_aux_features(self, aux_feats, H: int, W: int):
+        """Reshape (B, S, N, C) aux features into (B, S, h_p, w_p, C)."""
+        ph, pw = H // self.PATCH_SIZE, W // self.PATCH_SIZE
+        out = []
+        for f in aux_feats:
+            B, S, N, C = f.shape
+            assert N == ph * pw, f"aux feature seq mismatch: {N} != {ph}*{pw}"
+            out.append(f.reshape(B, S, ph, pw, C))
+        return out

vendor/ComfyUI/comfy/ldm/depth_anything_3/preprocess.py

@@ -0,0 +1,128 @@
+"""Input/output preprocessing helpers for Depth Anything 3."""
+
+from __future__ import annotations
+
+from typing import Tuple
+
+import torch
+
+import comfy.utils
+
+PATCH_SIZE = 14
+
+# ImageNet normalization constants used during DA3 training.
+_IMAGENET_MEAN = torch.tensor([0.485, 0.456, 0.406])
+_IMAGENET_STD = torch.tensor([0.229, 0.224, 0.225])
+
+
+def _round_to_patch(x: int, patch: int = PATCH_SIZE) -> int:
+    down = (x // patch) * patch
+    up = down + patch
+    return up if abs(up - x) <= abs(x - down) else down
+
+
+def compute_target_size(orig_h: int, orig_w: int, process_res: int, method: str = "upper_bound_resize") -> Tuple[int, int]:
+    """Compute (target_h, target_w) for a single image.
+    upper_bound_resize: scale longest side to process_res, then round each dim to nearest multiple of 14 (default upstream method).
+    lower_bound_resize: scale shortest side to process_res, then round."""
+
+    if method == "upper_bound_resize":
+        longest = max(orig_h, orig_w)
+        scale = process_res / float(longest)
+    elif method == "lower_bound_resize":
+        shortest = min(orig_h, orig_w)
+        scale = process_res / float(shortest)
+    else:
+        raise ValueError(f"Unsupported process_res_method: {method}")
+
+    new_w = max(1, _round_to_patch(int(round(orig_w * scale))))
+    new_h = max(1, _round_to_patch(int(round(orig_h * scale))))
+    return new_h, new_w
+
+
+def preprocess_image(image: torch.Tensor, process_res: int = 504, method: str = "upper_bound_resize") -> torch.Tensor:
+    assert image.ndim == 4 and image.shape[-1] == 3, f"expected (B,H,W,3) IMAGE; got {tuple(image.shape)}"
+    B, H, W, _ = image.shape
+    target_h, target_w = compute_target_size(H, W, process_res, method)
+
+    # (B, H, W, 3) -> (B, 3, H, W)
+    x = image.movedim(-1, 1).contiguous()
+    if (target_h, target_w) != (H, W):
+        # Upstream uses cv2 INTER_CUBIC (upscale) / INTER_AREA (downscale).
+        # Lanczos in ``common_upscale`` is anti-aliased and produces the
+        # closest pixel-wise match in a sweep across {bilinear, bicubic,
+        # area, lanczos, bislerp}. Used in both directions for simplicity.
+        x = comfy.utils.common_upscale(x.float(), target_w, target_h, "lanczos", "disabled",)
+    x = x.clamp(0.0, 1.0)
+
+    mean = _IMAGENET_MEAN.to(device=x.device, dtype=x.dtype).view(1, 3, 1, 1)
+    std = _IMAGENET_STD.to(device=x.device, dtype=x.dtype).view(1, 3, 1, 1)
+    x = (x - mean) / std
+    return x
+
+
+# -----------------------------------------------------------------------------
+# Output post-processing (sky-aware clipping for Mono/Metric variants)
+# -----------------------------------------------------------------------------
+
+
+def compute_non_sky_mask(sky_prediction: torch.Tensor, threshold: float = 0.3) -> torch.Tensor:
+    """Boolean mask: True for non-sky pixels (sky probability < threshold)."""
+    return sky_prediction < threshold
+
+
+def apply_sky_aware_clip(depth: torch.Tensor, sky: torch.Tensor, threshold: float = 0.3, quantile: float = 0.99) -> torch.Tensor:
+    """Clips sky regions to the 99th percentile of non-sky depth. Returns a new depth tensor."""
+    non_sky = compute_non_sky_mask(sky, threshold=threshold)
+    if non_sky.sum() <= 10 or (~non_sky).sum() <= 10:
+        return depth.clone()
+
+    non_sky_depth = depth[non_sky]
+    if non_sky_depth.numel() > 100_000:
+        idx = torch.randint(0, non_sky_depth.numel(), (100_000,), device=non_sky_depth.device)
+        sampled = non_sky_depth[idx]
+    else:
+        sampled = non_sky_depth
+
+    max_depth = torch.quantile(sampled, quantile)
+    out = depth.clone()
+    out[~non_sky] = max_depth
+    return out
+
+
+def normalize_depth_v2_style(depth: torch.Tensor, sky: torch.Tensor | None = None, low_quantile: float = 0.01, high_quantile: float = 0.99) -> torch.Tensor:
+    """V2-style normalization computes percentile bounds over non-sky pixels (when available), then maps depth into [0, 1] with near = white (1.0)."""
+    if sky is not None:
+        mask = compute_non_sky_mask(sky)
+        if mask.any():
+            valid = depth[mask]
+        else:
+            valid = depth.flatten()
+    else:
+        valid = depth.flatten()
+
+    if valid.numel() > 100_000:
+        idx = torch.randint(0, valid.numel(), (100_000,), device=valid.device)
+        sample = valid[idx]
+    else:
+        sample = valid
+
+    lo = torch.quantile(sample, low_quantile)
+    hi = torch.quantile(sample, high_quantile)
+    rng = (hi - lo).clamp(min=1e-6)
+    norm = ((depth - lo) / rng).clamp(0.0, 1.0)
+    # Nearer pixels are brighter (1.0)
+    norm = 1.0 - norm
+    if sky is not None:
+        # Sky pixels become black (far / unknown)
+        sky_mask = ~compute_non_sky_mask(sky)
+        norm = torch.where(sky_mask, torch.zeros_like(norm), norm)
+    return norm
+
+
+def normalize_depth_min_max(depth: torch.Tensor) -> torch.Tensor:
+    """Simple per-frame min/max normalization with near=1.0 convention."""
+    lo = depth.amin(dim=(-2, -1), keepdim=True)
+    hi = depth.amax(dim=(-2, -1), keepdim=True)
+    rng = (hi - lo).clamp(min=1e-6)
+    return 1.0 - ((depth - lo) / rng).clamp(0.0, 1.0)

vendor/ComfyUI/comfy/ldm/depth_anything_3/ray_pose.py

@@ -0,0 +1,272 @@
+"""Ray-to-pose conversion for the multi-view path of Depth Anything 3."""
+
+from __future__ import annotations
+
+from typing import Optional, Tuple
+
+import torch
+
+
+# qr/svd use fp32: CUDA often has no fp16/bf16 kernels for these ops.
+
+
+def _ql_decomposition(A: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]:
+    """Decompose A = Q @ L with Q orthogonal and L lower-triangular.
+    Implemented in terms of QR by reversing the columns/rows; the standard
+    trick from the upstream reference. Inputs A are (3, 3)."""
+    P = torch.tensor([[0, 0, 1], [0, 1, 0], [1, 0, 0]], device=A.device, dtype=A.dtype)
+    A_tilde = A @ P
+    # CUDA QR is not implemented for fp16/bf16; upcast just for this call.
+    Q_tilde, R_tilde = torch.linalg.qr(A_tilde.float())
+    Q_tilde = Q_tilde.to(A.dtype)
+    R_tilde = R_tilde.to(A.dtype)
+    Q = Q_tilde @ P
+    L = P @ R_tilde @ P
+    d = torch.diag(L)
+    sign = torch.sign(d)
+    Q = Q * sign[None, :]  # scale columns of Q
+    L = L * sign[:, None]  # scale rows of L
+    return Q, L
+
+
+def _homogenize_points(points: torch.Tensor) -> torch.Tensor:
+    return torch.cat([points, torch.ones_like(points[..., :1])], dim=-1)
+
+
+# -----------------------------------------------------------------------------
+# Weighted-LSQ + RANSAC homography (batched)
+# -----------------------------------------------------------------------------
+
+
+def _find_homography_weighted_lsq(src_pts: torch.Tensor, dst_pts: torch.Tensor, confident_weight: torch.Tensor,) -> torch.Tensor:
+    """Solve a single H with weighted least-squares (DLT)."""
+    N = src_pts.shape[0]
+    if N < 4:
+        raise ValueError("At least 4 points are required to compute a homography.")
+    w = confident_weight.sqrt().unsqueeze(1)  # (N, 1)
+    x = src_pts[:, 0:1]
+    y = src_pts[:, 1:2]
+    u = dst_pts[:, 0:1]
+    v = dst_pts[:, 1:2]
+    zeros = torch.zeros_like(x)
+    A1 = torch.cat([-x * w, -y * w, -w, zeros, zeros, zeros, x * u * w, y * u * w, u * w], dim=1)
+    A2 = torch.cat([zeros, zeros, zeros, -x * w, -y * w, -w, x * v * w, y * v * w, v * w], dim=1)
+    A = torch.cat([A1, A2], dim=0)        # (2N, 9)
+    # CUDA SVD is not implemented for fp16/bf16; upcast just for this call.
+    _, _, Vh = torch.linalg.svd(A.float())
+    Vh = Vh.to(A.dtype)
+    H = Vh[-1].reshape(3, 3)
+    return H / H[-1, -1]
+
+
+def _find_homography_weighted_lsq_batched(src_pts_batch: torch.Tensor, dst_pts_batch: torch.Tensor, confident_weight_batch: torch.Tensor) -> torch.Tensor:
+    """Batched DLT solver. Inputs (B, K, 2) / (B, K); output (B, 3, 3)."""
+    B, K, _ = src_pts_batch.shape
+    w = confident_weight_batch.sqrt().unsqueeze(2)
+    x = src_pts_batch[:, :, 0:1]
+    y = src_pts_batch[:, :, 1:2]
+    u = dst_pts_batch[:, :, 0:1]
+    v = dst_pts_batch[:, :, 1:2]
+    zeros = torch.zeros_like(x)
+    A1 = torch.cat([-x * w, -y * w, -w, zeros, zeros, zeros, x * u * w, y * u * w, u * w], dim=2)
+    A2 = torch.cat([zeros, zeros, zeros, -x * w, -y * w, -w, x * v * w, y * v * w, v * w], dim=2)
+    A = torch.cat([A1, A2], dim=1)        # (B, 2K, 9)
+    # CUDA SVD is not implemented for fp16/bf16; upcast just for this call.
+    _, _, Vh = torch.linalg.svd(A.float())
+    Vh = Vh.to(A.dtype)
+    H = Vh[:, -1].reshape(B, 3, 3)
+    return H / H[:, 2:3, 2:3]
+
+
+def _ransac_find_homography_weighted_batched(
+    src_pts: torch.Tensor,                # (B, N, 2)
+    dst_pts: torch.Tensor,                # (B, N, 2)
+    confident_weight: torch.Tensor,       # (B, N)
+    n_sample: int,
+    n_iter: int = 100,
+    reproj_threshold: float = 3.0,
+    num_sample_for_ransac: int = 8,
+    max_inlier_num: int = 10000,
+    rand_sample_iters_idx: Optional[torch.Tensor] = None,
+) -> torch.Tensor:
+    """Batched weighted-RANSAC homography estimator. Returns (B, 3, 3) homography matrices."""
+    B, N, _ = src_pts.shape
+    assert N >= 4
+    device = src_pts.device
+
+    sorted_idx = torch.argsort(confident_weight, descending=True, dim=1)
+    candidate_idx = sorted_idx[:, :n_sample]                  # (B, n_sample)
+
+    if rand_sample_iters_idx is None:
+        rand_sample_iters_idx = torch.stack(
+            [torch.randperm(n_sample, device=device)[:num_sample_for_ransac]
+             for _ in range(n_iter)],
+            dim=0,
+        )
+
+    rand_idx = candidate_idx[:, rand_sample_iters_idx]        # (B, n_iter, k)
+    b_idx = (
+        torch.arange(B, device=device)
+        .view(B, 1, 1)
+        .expand(B, n_iter, num_sample_for_ransac)
+    )
+    src_b = src_pts[b_idx, rand_idx]
+    dst_b = dst_pts[b_idx, rand_idx]
+    w_b = confident_weight[b_idx, rand_idx]
+
+    cB, cN = src_b.shape[:2]
+    H_batch = _find_homography_weighted_lsq_batched(
+        src_b.flatten(0, 1), dst_b.flatten(0, 1), w_b.flatten(0, 1),
+    ).unflatten(0, (cB, cN))                                  # (B, n_iter, 3, 3)
+
+    src_homo = torch.cat([src_pts, torch.ones(B, N, 1, device=device, dtype=src_pts.dtype)], dim=2)
+    proj = torch.bmm(
+        src_homo.unsqueeze(1).expand(B, n_iter, N, 3).reshape(-1, N, 3),
+        H_batch.reshape(-1, 3, 3).transpose(1, 2),
+    )                                                          # (B*n_iter, N, 3)
+    proj_xy = (proj[:, :, :2] / proj[:, :, 2:3]).reshape(B, n_iter, N, 2)
+    err = ((proj_xy - dst_pts.unsqueeze(1)) ** 2).sum(-1).sqrt()  # (B, n_iter, N)
+    inlier_mask = err < reproj_threshold
+    score = (inlier_mask * confident_weight.unsqueeze(1)).sum(dim=2)
+    best_idx = torch.argmax(score, dim=1)
+    best_inlier_mask = inlier_mask[torch.arange(B, device=device), best_idx]
+
+    # Refit with the inlier set (per-batch, since the inlier counts vary).
+    H_inlier_list = []
+    for b in range(B):
+        mask = best_inlier_mask[b]
+        in_src = src_pts[b][mask]
+        in_dst = dst_pts[b][mask]
+        in_w = confident_weight[b][mask]
+        if in_src.shape[0] < 4:
+            # Fall back to identity when RANSAC fails to find enough inliers.
+            H_inlier_list.append(torch.eye(3, device=device, dtype=src_pts.dtype))
+            continue
+        sorted_w = torch.argsort(in_w, descending=True)
+        if len(sorted_w) > max_inlier_num:
+            keep = max(int(len(sorted_w) * 0.95), max_inlier_num)
+            sorted_w = sorted_w[:keep][torch.randperm(keep, device=device)[:max_inlier_num]]
+        H_inlier_list.append(
+            _find_homography_weighted_lsq(in_src[sorted_w], in_dst[sorted_w], in_w[sorted_w])
+        )
+    return torch.stack(H_inlier_list, dim=0)
+
+
+# -----------------------------------------------------------------------------
+# Camera-ray utilities
+# -----------------------------------------------------------------------------
+
+
+def _unproject_identity(num_y: int, num_x: int, B: int, S: int, device, dtype) -> torch.Tensor:
+    """Camera-space unit rays for an identity intrinsic on a 2x2 image plane."""
+    dx = 1.0 / num_x
+    dy = 1.0 / num_y
+    # Centered camera-space coords directly (skip the K^-1 step since it's
+    # just a translation by -1 on x and y when K is identity-with-center=1).
+    y = torch.linspace(-(1 - dy), (1 - dy), num_y, device=device, dtype=dtype)
+    x = torch.linspace(-(1 - dx), (1 - dx), num_x, device=device, dtype=dtype)
+    yy, xx = torch.meshgrid(y, x, indexing="ij")
+    grid = torch.stack((xx, yy), dim=-1)            # (h, w, 2)
+    grid = grid.unsqueeze(0).unsqueeze(0).expand(B, S, num_y, num_x, 2)
+    return torch.cat([grid, torch.ones_like(grid[..., :1])], dim=-1)
+
+
+def _camray_to_caminfo(
+    camray: torch.Tensor,  # (B, S, h, w, 6)
+    confidence: Optional[torch.Tensor] = None,  # (B, S, h, w)
+    reproj_threshold: float = 0.2,
+) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
+    """Convert per-pixel camera rays to per-view (R, T, focal, principal)."""
+    if confidence is None:
+        confidence = torch.ones_like(camray[..., 0])
+    B, S, h, w, _ = camray.shape
+    device = camray.device
+    dtype = camray.dtype
+
+    rays_target = camray[..., :3]                           # (B, S, h, w, 3)
+    rays_origin = _unproject_identity(h, w, B, S, device, dtype)
+
+    # Flatten (B*S, h*w, *) for the RANSAC routine.
+    rays_target = rays_target.flatten(0, 1).flatten(1, 2)
+    rays_origin = rays_origin.flatten(0, 1).flatten(1, 2)
+    weights = confidence.flatten(0, 1).flatten(1, 2).clone()
+
+    # Project to 2D in homogeneous form (the upstream calls this "perspective division").
+    z_thresh = 1e-4
+    mask = (rays_target[:, :, 2].abs() > z_thresh) & (rays_origin[:, :, 2].abs() > z_thresh)
+    weights = torch.where(mask, weights, torch.zeros_like(weights))
+    src = rays_origin.clone()
+    dst = rays_target.clone()
+    src[..., 0] = torch.where(mask, src[..., 0] / src[..., 2], src[..., 0])
+    src[..., 1] = torch.where(mask, src[..., 1] / src[..., 2], src[..., 1])
+    dst[..., 0] = torch.where(mask, dst[..., 0] / dst[..., 2], dst[..., 0])
+    dst[..., 1] = torch.where(mask, dst[..., 1] / dst[..., 2], dst[..., 1])
+    src = src[..., :2]
+    dst = dst[..., :2]
+
+    N = src.shape[1]
+    n_iter = 100
+    sample_ratio = 0.3
+    num_sample_for_ransac = 8
+    n_sample = max(num_sample_for_ransac, int(N * sample_ratio))
+    rand_idx = torch.stack(
+        [torch.randperm(n_sample, device=device)[:num_sample_for_ransac] for _ in range(n_iter)],
+        dim=0,
+    )
+
+    # Chunk along the view axis to keep peak memory predictable.
+    chunk = 2
+    A_list = []
+    for i in range(0, src.shape[0], chunk):
+        A = _ransac_find_homography_weighted_batched(
+            src[i:i + chunk], dst[i:i + chunk], weights[i:i + chunk],
+            n_sample=n_sample, n_iter=n_iter,
+            num_sample_for_ransac=num_sample_for_ransac,
+            reproj_threshold=reproj_threshold,
+            rand_sample_iters_idx=rand_idx,
+            max_inlier_num=8000,
+        )
+        # Flip sign on dets that come out < 0 (so that the QL produces a
+        # right-handed rotation). ``det`` lacks fp16/bf16 CUDA kernels, so
+        # do the comparison in fp32.
+        flip = torch.linalg.det(A.float()) < 0
+        A = torch.where(flip[:, None, None], -A, A)
+        A_list.append(A)
+    A = torch.cat(A_list, dim=0)                            # (B*S, 3, 3)
+
+    R_list, f_list, pp_list = [], [], []
+    for i in range(A.shape[0]):
+        R, L = _ql_decomposition(A[i])
+        L = L / L[2][2]
+        f_list.append(torch.stack((L[0][0], L[1][1])))
+        pp_list.append(torch.stack((L[2][0], L[2][1])))
+        R_list.append(R)
+    R = torch.stack(R_list).reshape(B, S, 3, 3)
+    focal = torch.stack(f_list).reshape(B, S, 2)
+    pp = torch.stack(pp_list).reshape(B, S, 2)
+
+    # Translation: confidence-weighted average of camray direction(s).
+    cf = confidence.flatten(0, 1).flatten(1, 2)
+    T = (camray.flatten(0, 1).flatten(1, 2)[..., 3:] * cf.unsqueeze(-1)).sum(dim=1)
+    T = T / cf.sum(dim=-1, keepdim=True)
+    T = T.reshape(B, S, 3)
+
+    # Match upstream output convention: focal -> 1/focal, pp + 1.
+    return R, T, 1.0 / focal, pp + 1.0
+
+
+def get_extrinsic_from_camray(
+    camray: torch.Tensor,  # (B, S, h, w, 6)
+    conf: torch.Tensor,  # (B, S, h, w, 1) or (B, S, h, w)
+    patch_size_y: int,
+    patch_size_x: int,
+) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+    """Wrap a 4x4 extrinsic + per-view focal + principal-point output."""
+    if conf.ndim == 5 and conf.shape[-1] == 1:
+        conf = conf.squeeze(-1)
+    R, T, focal, pp = _camray_to_caminfo(camray, confidence=conf)
+    extr = torch.cat([R, T.unsqueeze(-1)], dim=-1)           # (B, S, 3, 4)
+    homo_row = torch.tensor([0, 0, 0, 1], dtype=R.dtype, device=R.device)
+    homo_row = homo_row.view(1, 1, 1, 4).expand(R.shape[0], R.shape[1], 1, 4)
+    extr = torch.cat([extr, homo_row], dim=-2)               # (B, S, 4, 4)
+    return extr, focal, pp

vendor/ComfyUI/comfy/ldm/depth_anything_3/reference_view_selector.py

@@ -0,0 +1,87 @@
+"""Reference-view selection for the multi-view path of Depth Anything 3."""
+
+from __future__ import annotations
+
+from typing import Literal
+
+import torch
+
+
+RefViewStrategy = Literal["first", "middle", "saddle_balanced", "saddle_sim_range"]
+
+
+# Per the upstream constants module: ``THRESH_FOR_REF_SELECTION = 3``.
+# Reference selection only runs when there are at least this many views.
+THRESH_FOR_REF_SELECTION: int = 3
+
+
+def select_reference_view(x: torch.Tensor, strategy: RefViewStrategy = "saddle_balanced") -> torch.Tensor:
+    """Pick a reference view index per batch element."""
+    B, S, _, _ = x.shape
+    if S <= 1:
+        return torch.zeros(B, dtype=torch.long, device=x.device)
+    if strategy == "first":
+        return torch.zeros(B, dtype=torch.long, device=x.device)
+    if strategy == "middle":
+        return torch.full((B,), S // 2, dtype=torch.long, device=x.device)
+
+    # Feature-based strategies: normalised cls/cam token per view.
+    img_class_feat = x[:, :, 0] / x[:, :, 0].norm(dim=-1, keepdim=True)  # (B,S,C)
+
+    if strategy == "saddle_balanced":
+        sim = torch.matmul(img_class_feat, img_class_feat.transpose(1, 2))  # (B,S,S)
+        sim_no_diag = sim - torch.eye(S, device=sim.device).unsqueeze(0)
+        sim_score = sim_no_diag.sum(dim=-1) / (S - 1)               # (B,S)
+        feat_norm = x[:, :, 0].norm(dim=-1)                          # (B,S)
+        feat_var = img_class_feat.var(dim=-1)                        # (B,S)
+
+        def _normalize(metric):
+            mn = metric.min(dim=1, keepdim=True).values
+            mx = metric.max(dim=1, keepdim=True).values
+            return (metric - mn) / (mx - mn + 1e-8)
+
+        sim_n, norm_n, var_n = _normalize(sim_score), _normalize(feat_norm), _normalize(feat_var)
+        balance = (sim_n - 0.5).abs() + (norm_n - 0.5).abs() + (var_n - 0.5).abs()
+        return balance.argmin(dim=1)
+
+    if strategy == "saddle_sim_range":
+        sim = torch.matmul(img_class_feat, img_class_feat.transpose(1, 2))
+        sim_no_diag = sim - torch.eye(S, device=sim.device).unsqueeze(0)
+        sim_max = sim_no_diag.max(dim=-1).values
+        sim_min = sim_no_diag.min(dim=-1).values
+        return (sim_max - sim_min).argmax(dim=1)
+
+    raise ValueError(
+        f"Unknown reference view selection strategy: {strategy!r}. "
+        f"Must be one of: 'first', 'middle', 'saddle_balanced', 'saddle_sim_range'"
+    )
+
+
+def reorder_by_reference(x: torch.Tensor, b_idx: torch.Tensor) -> torch.Tensor:
+    """Reorder x so the reference view is at position 0 in axis S."""
+    B, S = x.shape[0], x.shape[1]
+    if S <= 1:
+        return x
+    positions = torch.arange(S, device=x.device).unsqueeze(0).expand(B, -1)
+    b_idx_exp = b_idx.unsqueeze(1)
+    reorder = torch.where(
+        (positions > 0) & (positions <= b_idx_exp),
+        positions - 1,
+        positions,
+    )
+    reorder[:, 0] = b_idx
+    batch = torch.arange(B, device=x.device).unsqueeze(1)
+    return x[batch, reorder]
+
+
+def restore_original_order(x: torch.Tensor, b_idx: torch.Tensor) -> torch.Tensor:
+    """Inverse of reorder_by_reference."""
+    B, S = x.shape[0], x.shape[1]
+    if S <= 1:
+        return x
+    target_positions = torch.arange(S, device=x.device).unsqueeze(0).expand(B, -1)
+    b_idx_exp = b_idx.unsqueeze(1)
+    restore = torch.where(target_positions < b_idx_exp, target_positions + 1, target_positions)
+    restore = torch.scatter(restore, dim=1, index=b_idx_exp, src=torch.zeros_like(b_idx_exp))
+    batch = torch.arange(B, device=x.device).unsqueeze(1)
+    return x[batch, restore]

vendor/ComfyUI/comfy/ldm/depth_anything_3/transform.py

@@ -0,0 +1,160 @@
+"""Geometry / camera transform helpers for Depth Anything 3."""
+
+from __future__ import annotations
+
+from typing import Tuple
+
+import torch
+import torch.nn.functional as F
+
+
+# -----------------------------------------------------------------------------
+# Affine 4x4 helpers
+# -----------------------------------------------------------------------------
+
+
+def as_homogeneous(ext: torch.Tensor) -> torch.Tensor:
+    """Promote (...,3,4) extrinsics to (...,4,4) homogeneous form. No-op when the input is already ``(...,4,4)``."""
+    if ext.shape[-2:] == (4, 4):
+        return ext
+    if ext.shape[-2:] == (3, 4):
+        ones = torch.zeros_like(ext[..., :1, :4])
+        ones[..., 0, 3] = 1.0
+        return torch.cat([ext, ones], dim=-2)
+    raise ValueError(f"Invalid affine shape: {ext.shape}")
+
+
+def affine_inverse(A: torch.Tensor) -> torch.Tensor:
+    """Inverse of an affine matrix ``[R|T; 0 0 0 1]``."""
+    R = A[..., :3, :3]
+    T = A[..., :3, 3:]
+    P = A[..., 3:, :]
+    return torch.cat([torch.cat([R.mT, -R.mT @ T], dim=-1), P], dim=-2)
+
+
+# -----------------------------------------------------------------------------
+# Quaternion <-> rotation matrix (xyzw / scalar-last)
+# -----------------------------------------------------------------------------
+
+
+def _sqrt_positive_part(x: torch.Tensor) -> torch.Tensor:
+    """sqrt(max(0, x)) with a zero subgradient where x == 0."""
+    ret = torch.zeros_like(x)
+    positive_mask = x > 0
+    if torch.is_grad_enabled():
+        ret[positive_mask] = torch.sqrt(x[positive_mask])
+    else:
+        ret = torch.where(positive_mask, torch.sqrt(x), ret)
+    return ret
+
+
+def standardize_quaternion(quaternions: torch.Tensor) -> torch.Tensor:
+    """Force the real part of a unit quaternion (xyzw) to be non-negative."""
+    return torch.where(quaternions[..., 3:4] < 0, -quaternions, quaternions)
+
+
+def quat_to_mat(quaternions: torch.Tensor) -> torch.Tensor:
+    """Convert quaternions (xyzw) to (...,3,3) rotation matrices."""
+    i, j, k, r = torch.unbind(quaternions, -1)
+    two_s = 2.0 / (quaternions * quaternions).sum(-1)
+    o = torch.stack(
+        (
+            1 - two_s * (j * j + k * k),
+            two_s * (i * j - k * r),
+            two_s * (i * k + j * r),
+            two_s * (i * j + k * r),
+            1 - two_s * (i * i + k * k),
+            two_s * (j * k - i * r),
+            two_s * (i * k - j * r),
+            two_s * (j * k + i * r),
+            1 - two_s * (i * i + j * j),
+        ),
+        -1,
+    )
+    return o.reshape(quaternions.shape[:-1] + (3, 3))
+
+
+def mat_to_quat(matrix: torch.Tensor) -> torch.Tensor:
+    """Convert (...,3,3) rotation matrices to quaternions (xyzw)."""
+    if matrix.size(-1) != 3 or matrix.size(-2) != 3:
+        raise ValueError(f"Invalid rotation matrix shape {matrix.shape}.")
+
+    batch_dim = matrix.shape[:-2]
+    m00, m01, m02, m10, m11, m12, m20, m21, m22 = torch.unbind(
+        matrix.reshape(batch_dim + (9,)), dim=-1
+    )
+
+    q_abs = _sqrt_positive_part(
+        torch.stack(
+            [
+                1.0 + m00 + m11 + m22,
+                1.0 + m00 - m11 - m22,
+                1.0 - m00 + m11 - m22,
+                1.0 - m00 - m11 + m22,
+            ],
+            dim=-1,
+        )
+    )
+
+    quat_by_rijk = torch.stack(
+        [
+            torch.stack([q_abs[..., 0] ** 2, m21 - m12, m02 - m20, m10 - m01], dim=-1),
+            torch.stack([m21 - m12, q_abs[..., 1] ** 2, m10 + m01, m02 + m20], dim=-1),
+            torch.stack([m02 - m20, m10 + m01, q_abs[..., 2] ** 2, m12 + m21], dim=-1),
+            torch.stack([m10 - m01, m20 + m02, m21 + m12, q_abs[..., 3] ** 2], dim=-1),
+        ],
+        dim=-2,
+    )
+
+    flr = torch.tensor(0.1).to(dtype=q_abs.dtype, device=q_abs.device)
+    quat_candidates = quat_by_rijk / (2.0 * q_abs[..., None].max(flr))
+
+    out = quat_candidates[F.one_hot(q_abs.argmax(dim=-1), num_classes=4) > 0.5, :].reshape(
+        batch_dim + (4,)
+    )
+    # Reorder rijk -> xyzw (i.e. ijkr).
+    out = out[..., [1, 2, 3, 0]]
+    return standardize_quaternion(out)
+
+
+# -----------------------------------------------------------------------------
+# Pose-encoding <-> extrinsics + intrinsics
+# -----------------------------------------------------------------------------
+
+
+def extri_intri_to_pose_encoding(extrinsics: torch.Tensor, intrinsics: torch.Tensor, image_size_hw: Tuple[int, int]) -> torch.Tensor:
+    """Pack (extr, intr, image_size) into the 9-D pose-encoding vector.
+    extrinsics: camera-to-world (c2w) (B,S,4,4) matrices,
+    intrinsics: pixel-space (B,S,3,3) matrices,
+    image_size_hw: is a (H, W) pair.
+    """
+    R = extrinsics[..., :3, :3]
+    T = extrinsics[..., :3, 3]
+    quat = mat_to_quat(R)
+    H, W = image_size_hw
+    fov_h = 2 * torch.atan((H / 2) / intrinsics[..., 1, 1])
+    fov_w = 2 * torch.atan((W / 2) / intrinsics[..., 0, 0])
+    return torch.cat([T, quat, fov_h[..., None], fov_w[..., None]], dim=-1).float()
+
+
+def pose_encoding_to_extri_intri(pose_encoding: torch.Tensor, image_size_hw: Tuple[int, int]) -> Tuple[torch.Tensor, torch.Tensor]:
+    """Inverse of extri_intri_to_pose_encoding."""
+    T = pose_encoding[..., :3]
+    quat = pose_encoding[..., 3:7]
+    fov_h = pose_encoding[..., 7]
+    fov_w = pose_encoding[..., 8]
+    # Normalize to unit quaternion. CameraDec outputs raw values; a near-zero
+    # quaternion causes two_s = 2/norm² → inf in quat_to_mat → NaN extrinsics.
+    quat = quat / quat.norm(dim=-1, keepdim=True).clamp(min=1e-6)
+    R = quat_to_mat(quat)
+    extrinsics = torch.cat([R, T[..., None]], dim=-1)
+    H, W = image_size_hw
+    fy = (H / 2.0) / torch.clamp(torch.tan(fov_h / 2.0), 1e-6)
+    fx = (W / 2.0) / torch.clamp(torch.tan(fov_w / 2.0), 1e-6)
+    intrinsics = torch.zeros(pose_encoding.shape[:2] + (3, 3), device=pose_encoding.device, dtype=pose_encoding.dtype)
+    intrinsics[..., 0, 0] = fx
+    intrinsics[..., 1, 1] = fy
+    intrinsics[..., 0, 2] = W / 2
+    intrinsics[..., 1, 2] = H / 2
+    intrinsics[..., 2, 2] = 1.0
+    return extrinsics, intrinsics