ComfyUI/comfy/image_encoders/dino2.py
2026-05-13 10:59:31 +02:00

528 lines
24 KiB
Python

import math
import torch
import torch.nn.functional as F
from comfy.text_encoders.bert import BertAttention
import comfy.model_management
from comfy.ldm.modules.attention import optimized_attention_for_device
from comfy.ldm.depth_anything_3.reference_view_selector import (
select_reference_view, reorder_by_reference, restore_original_order,
THRESH_FOR_REF_SELECTION,
)
class Dino2AttentionOutput(torch.nn.Module):
def __init__(self, input_dim, output_dim, layer_norm_eps, dtype, device, operations):
super().__init__()
self.dense = operations.Linear(input_dim, output_dim, dtype=dtype, device=device)
def forward(self, x):
return self.dense(x)
class Dino2AttentionBlock(torch.nn.Module):
def __init__(self, embed_dim, heads, layer_norm_eps, dtype, device, operations,
qk_norm=False):
super().__init__()
self.heads = heads
self.head_dim = embed_dim // heads
self.attention = BertAttention(embed_dim, heads, dtype, device, operations)
self.output = Dino2AttentionOutput(embed_dim, embed_dim, layer_norm_eps, dtype, device, operations)
if qk_norm:
self.q_norm = operations.LayerNorm(self.head_dim, dtype=dtype, device=device)
self.k_norm = operations.LayerNorm(self.head_dim, dtype=dtype, device=device)
else:
self.q_norm = None
self.k_norm = None
def forward(self, x, mask, optimized_attention, pos=None, rope=None):
# Fast path used by the existing CLIP-vision DINOv2 (no DA3 extensions).
if self.q_norm is None and rope is None:
return self.output(self.attention(x, mask, optimized_attention))
# DA3 path: do QKV manually so we can apply per-head QK-norm and 2D RoPE.
attn = self.attention
B, N, C = x.shape
h = self.heads
d = self.head_dim
q = attn.query(x).view(B, N, h, d).transpose(1, 2)
k = attn.key(x).view(B, N, h, d).transpose(1, 2)
v = attn.value(x).view(B, N, h, d).transpose(1, 2)
if self.q_norm is not None:
q = self.q_norm(q)
k = self.k_norm(k)
if rope is not None and pos is not None:
q = rope(q, pos)
k = rope(k, pos)
out = F.scaled_dot_product_attention(q, k, v, attn_mask=mask)
out = out.transpose(1, 2).reshape(B, N, C)
return self.output(out)
class LayerScale(torch.nn.Module):
def __init__(self, dim, dtype, device, operations):
super().__init__()
self.lambda1 = torch.nn.Parameter(torch.empty(dim, device=device, dtype=dtype))
def forward(self, x):
return x * comfy.model_management.cast_to_device(self.lambda1, x.device, x.dtype)
class Dinov2MLP(torch.nn.Module):
def __init__(self, hidden_size: int, dtype, device, operations):
super().__init__()
mlp_ratio = 4
hidden_features = int(hidden_size * mlp_ratio)
self.fc1 = operations.Linear(hidden_size, hidden_features, bias = True, device=device, dtype=dtype)
self.fc2 = operations.Linear(hidden_features, hidden_size, bias = True, device=device, dtype=dtype)
def forward(self, hidden_state: torch.Tensor) -> torch.Tensor:
hidden_state = self.fc1(hidden_state)
hidden_state = torch.nn.functional.gelu(hidden_state)
hidden_state = self.fc2(hidden_state)
return hidden_state
class SwiGLUFFN(torch.nn.Module):
def __init__(self, dim, dtype, device, operations):
super().__init__()
in_features = out_features = dim
hidden_features = int(dim * 4)
hidden_features = (int(hidden_features * 2 / 3) + 7) // 8 * 8
self.weights_in = operations.Linear(in_features, 2 * hidden_features, bias=True, device=device, dtype=dtype)
self.weights_out = operations.Linear(hidden_features, out_features, bias=True, device=device, dtype=dtype)
def forward(self, x):
x = self.weights_in(x)
x1, x2 = x.chunk(2, dim=-1)
x = torch.nn.functional.silu(x1) * x2
return self.weights_out(x)
class Dino2Block(torch.nn.Module):
def __init__(self, dim, num_heads, layer_norm_eps, dtype, device, operations, use_swiglu_ffn,
qk_norm=False):
super().__init__()
self.attention = Dino2AttentionBlock(dim, num_heads, layer_norm_eps, dtype, device, operations,
qk_norm=qk_norm)
self.layer_scale1 = LayerScale(dim, dtype, device, operations)
self.layer_scale2 = LayerScale(dim, dtype, device, operations)
if use_swiglu_ffn:
self.mlp = SwiGLUFFN(dim, dtype, device, operations)
else:
self.mlp = Dinov2MLP(dim, dtype, device, operations)
self.norm1 = operations.LayerNorm(dim, eps=layer_norm_eps, dtype=dtype, device=device)
self.norm2 = operations.LayerNorm(dim, eps=layer_norm_eps, dtype=dtype, device=device)
def forward(self, x, optimized_attention, pos=None, rope=None, attn_mask=None):
x = x + self.layer_scale1(self.attention(self.norm1(x), attn_mask, optimized_attention,
pos=pos, rope=rope))
x = x + self.layer_scale2(self.mlp(self.norm2(x)))
return x
# -----------------------------------------------------------------------------
# 2D Rotary position embedding (DA3 extension)
# -----------------------------------------------------------------------------
class _PositionGetter:
"""Cache (h, w) -> flat (y, x) position grid used to feed ``rope``."""
def __init__(self):
self._cache: dict = {}
def __call__(self, batch_size: int, height: int, width: int, device) -> torch.Tensor:
key = (height, width, device)
if key not in self._cache:
y = torch.arange(height, device=device)
x = torch.arange(width, device=device)
self._cache[key] = torch.cartesian_prod(y, x)
cached = self._cache[key]
return cached.view(1, height * width, 2).expand(batch_size, -1, -1).clone()
class RotaryPositionEmbedding2D(torch.nn.Module):
"""2D RoPE used by DA3-Small/Base. No learnable parameters."""
def __init__(self, frequency: float = 100.0):
super().__init__()
self.base_frequency = frequency
self._freq_cache: dict = {}
def _components(self, dim: int, seq_len: int, device, dtype):
key = (dim, seq_len, device, dtype)
if key not in self._freq_cache:
exp = torch.arange(0, dim, 2, device=device).float() / dim
inv_freq = 1.0 / (self.base_frequency ** exp)
pos = torch.arange(seq_len, device=device, dtype=inv_freq.dtype)
ang = torch.einsum("i,j->ij", pos, inv_freq)
ang = ang.to(dtype)
ang = torch.cat((ang, ang), dim=-1)
self._freq_cache[key] = (ang.cos().to(dtype), ang.sin().to(dtype))
return self._freq_cache[key]
@staticmethod
def _rotate(x: torch.Tensor) -> torch.Tensor:
d = x.shape[-1]
x1, x2 = x[..., : d // 2], x[..., d // 2:]
return torch.cat((-x2, x1), dim=-1)
def _apply_1d(self, tokens, positions, cos_c, sin_c):
cos = F.embedding(positions, cos_c)[:, None, :, :]
sin = F.embedding(positions, sin_c)[:, None, :, :]
return (tokens * cos) + (self._rotate(tokens) * sin)
def forward(self, tokens: torch.Tensor, positions: torch.Tensor) -> torch.Tensor:
feature_dim = tokens.size(-1) // 2
max_pos = int(positions.max()) + 1
cos_c, sin_c = self._components(feature_dim, max_pos, tokens.device, tokens.dtype)
v, h = tokens.chunk(2, dim=-1)
v = self._apply_1d(v, positions[..., 0], cos_c, sin_c)
h = self._apply_1d(h, positions[..., 1], cos_c, sin_c)
return torch.cat((v, h), dim=-1)
class Dino2Encoder(torch.nn.Module):
def __init__(self, dim, num_heads, layer_norm_eps, num_layers, dtype, device, operations, use_swiglu_ffn,
qknorm_start: int = -1, rope: "RotaryPositionEmbedding2D | None" = None,
rope_start: int = -1):
super().__init__()
self.layer = torch.nn.ModuleList([
Dino2Block(
dim, num_heads, layer_norm_eps, dtype, device, operations,
use_swiglu_ffn=use_swiglu_ffn,
qk_norm=(qknorm_start != -1 and i >= qknorm_start),
)
for i in range(num_layers)
])
self.rope = rope
self.rope_start = rope_start
def forward(self, x, intermediate_output=None):
# Backward-compat path used by ``ClipVisionModel`` (no DA3 extensions).
optimized_attention = optimized_attention_for_device(x.device, False, small_input=True)
if intermediate_output is not None:
if intermediate_output < 0:
intermediate_output = len(self.layer) + intermediate_output
intermediate = None
for i, layer in enumerate(self.layer):
x = layer(x, optimized_attention)
if i == intermediate_output:
intermediate = x.clone()
return x, intermediate
class Dino2PatchEmbeddings(torch.nn.Module):
def __init__(self, dim, num_channels=3, patch_size=14, image_size=518, dtype=None, device=None, operations=None):
super().__init__()
self.projection = operations.Conv2d(
in_channels=num_channels,
out_channels=dim,
kernel_size=patch_size,
stride=patch_size,
bias=True,
dtype=dtype,
device=device
)
def forward(self, pixel_values):
return self.projection(pixel_values).flatten(2).transpose(1, 2)
class Dino2Embeddings(torch.nn.Module):
def __init__(self, dim, dtype, device, operations,
patch_size: int = 14, image_size: int = 518,
use_mask_token: bool = True,
num_camera_tokens: int = 0):
super().__init__()
self.patch_size = patch_size
self.image_size = image_size
self.patch_embeddings = Dino2PatchEmbeddings(dim, patch_size=patch_size, image_size=image_size, dtype=dtype, device=device, operations=operations)
self.position_embeddings = torch.nn.Parameter(torch.empty(1, (image_size // patch_size) ** 2 + 1, dim, dtype=dtype, device=device))
self.cls_token = torch.nn.Parameter(torch.empty(1, 1, dim, dtype=dtype, device=device))
if use_mask_token:
self.mask_token = torch.nn.Parameter(torch.empty(1, dim, dtype=dtype, device=device))
else:
self.mask_token = None
if num_camera_tokens > 0:
# DA3 stores (ref_token, src_token) pairs that get injected at the
# alt-attn boundary; see ``Dinov2Model._inject_camera_token``.
self.camera_token = torch.nn.Parameter(torch.empty(1, num_camera_tokens, dim, dtype=dtype, device=device))
else:
self.camera_token = None
def _interpolate_pos_encoding(self, x: torch.Tensor, h: int, w: int) -> torch.Tensor:
previous_dtype = x.dtype
npatch = x.shape[1] - 1
N = self.position_embeddings.shape[1] - 1
pos_embed = comfy.model_management.cast_to_device(self.position_embeddings, x.device, x.dtype).float()
if npatch == N and w == h:
return pos_embed
class_pos_embed = pos_embed[:, 0]
patch_pos_embed = pos_embed[:, 1:]
dim = x.shape[-1]
ph = h // self.patch_size # patch grid height
pw = w // self.patch_size # patch grid width
M = int(math.sqrt(N))
assert N == M * M
# Historical 0.1 offset preserves bicubic resample compatibility with
# the original DINOv2 release; see the upstream PR for context.
# ``scale_factor`` is interpreted as (height_scale, width_scale) by
# ``F.interpolate`` so we must put the height scale FIRST. Earlier
# revisions of this function had it swapped which only worked for
# square inputs (e.g. CLIP-vision square crops); non-square inputs
# like DA3-Small / DA3-Base multi-view paths exposed the bug.
sh = float(ph + 0.1) / M
sw = float(pw + 0.1) / M
patch_pos_embed = F.interpolate(
patch_pos_embed.reshape(1, M, M, dim).permute(0, 3, 1, 2),
scale_factor=(sh, sw), mode="bicubic", antialias=False,
)
assert (ph, pw) == patch_pos_embed.shape[-2:]
patch_pos_embed = patch_pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
return torch.cat((class_pos_embed.unsqueeze(0), patch_pos_embed), dim=1).to(previous_dtype)
def forward(self, pixel_values):
_, _, H, W = pixel_values.shape
x = self.patch_embeddings(pixel_values)
# TODO: mask_token?
x = torch.cat((self.cls_token.to(device=x.device, dtype=x.dtype).expand(x.shape[0], -1, -1), x), dim=1)
x = x + self._interpolate_pos_encoding(x, H, W)
return x
class Dinov2Model(torch.nn.Module):
"""DINOv2 vision backbone.
Supports two operating modes:
* **CLIP-vision DINOv2** (default): vanilla DINOv2-ViT used for
``ClipVisionModel`` and SigLIP-style image encoding.
* **Depth Anything 3** extensions (opt-in via config keys): 2D RoPE,
QK-norm, alternating local/global attention, camera-token injection,
``cat_token`` output and multi-layer feature extraction. These are
enabled when the corresponding fields (``alt_start``, ``qknorm_start``,
``rope_start``, ``cat_token``) are set in ``config_dict``. When all of
them are at their disabled defaults this module behaves identically to
the historical ``Dinov2Model``.
"""
def __init__(self, config_dict, dtype, device, operations):
super().__init__()
num_layers = config_dict["num_hidden_layers"]
dim = config_dict["hidden_size"]
heads = config_dict["num_attention_heads"]
layer_norm_eps = config_dict["layer_norm_eps"]
use_swiglu_ffn = config_dict["use_swiglu_ffn"]
patch_size = config_dict.get("patch_size", 14)
image_size = config_dict.get("image_size", 518)
use_mask_token = config_dict.get("use_mask_token", True)
# DA3 extensions (all default to disabled).
self.alt_start = config_dict.get("alt_start", -1)
self.qknorm_start = config_dict.get("qknorm_start", -1)
self.rope_start = config_dict.get("rope_start", -1)
self.cat_token = config_dict.get("cat_token", False)
rope_freq = config_dict.get("rope_freq", 100.0)
self.embed_dim = dim
self.patch_size = patch_size
self.num_register_tokens = 0
self.patch_start_idx = 1
if self.rope_start != -1 and rope_freq > 0:
self.rope = RotaryPositionEmbedding2D(frequency=rope_freq)
self._position_getter = _PositionGetter()
else:
self.rope = None
self._position_getter = None
# camera_token shape: (1, 2, dim) -> (ref_token, src_token).
num_cam_tokens = 2 if self.alt_start != -1 else 0
self.embeddings = Dino2Embeddings(
dim, dtype, device, operations,
patch_size=patch_size, image_size=image_size,
use_mask_token=use_mask_token, num_camera_tokens=num_cam_tokens,
)
self.encoder = Dino2Encoder(
dim, heads, layer_norm_eps, num_layers, dtype, device, operations,
use_swiglu_ffn=use_swiglu_ffn,
qknorm_start=self.qknorm_start,
rope=self.rope, rope_start=self.rope_start,
)
self.layernorm = operations.LayerNorm(dim, eps=layer_norm_eps, dtype=dtype, device=device)
# ------------------------------------------------------------------
# CLIP-vision-style forward (no DA3 extensions, no multi-layer output).
# Kept for backward compatibility with ``ClipVisionModel.encode_image``.
# ------------------------------------------------------------------
def forward(self, pixel_values, attention_mask=None, intermediate_output=None):
x = self.embeddings(pixel_values)
x, i = self.encoder(x, intermediate_output=intermediate_output)
x = self.layernorm(x)
pooled_output = x[:, 0, :]
return x, i, pooled_output, None
# ------------------------------------------------------------------
# Depth Anything 3 forward
# ------------------------------------------------------------------
def _prepare_rope_positions(self, B, S, H, W, device):
if self.rope is None:
return None, None
ph, pw = H // self.patch_size, W // self.patch_size
pos = self._position_getter(B * S, ph, pw, device=device)
# Shift so the cls/cam token at position 0 is reserved for "no diff".
pos = pos + 1
cls_pos = torch.zeros(B * S, self.patch_start_idx, 2, device=device, dtype=pos.dtype)
# Per-view local: real grid positions for patches, 0 for cls token.
pos_local = torch.cat([cls_pos, pos], dim=1)
# Global (across views): same grid positions; cls token still at 0,
# but patches share the same positions in every view.
pos_global = torch.cat([cls_pos, torch.zeros_like(pos) + 1], dim=1)
return pos_local, pos_global
def _inject_camera_token(self, x: torch.Tensor, B: int, S: int,
cam_token: "torch.Tensor | None") -> torch.Tensor:
# x: (B, S, N, C). Replace token at index 0 with the camera token.
if cam_token is not None:
inj = cam_token
else:
ct = comfy.model_management.cast_to_device(self.embeddings.camera_token, x.device, x.dtype)
ref_token = ct[:, :1].expand(B, -1, -1)
src_token = ct[:, 1:].expand(B, max(S - 1, 0), -1)
inj = torch.cat([ref_token, src_token], dim=1)
x = x.clone()
x[:, :, 0] = inj
return x
def get_intermediate_layers(self, pixel_values, out_layers, cam_token=None,
ref_view_strategy="saddle_balanced",
export_feat_layers=None):
"""Multi-layer DINOv2 feature extraction used by Depth Anything 3.
Args:
pixel_values: ``(B, S, 3, H, W)`` views or ``(B, 3, H, W)``.
out_layers: indices into ``self.encoder.layer``.
cam_token: optional ``(B, S, dim)`` camera token to inject at
``alt_start``. If ``None`` and the model has its own
``camera_token`` parameter, that is used.
ref_view_strategy: when ``S >= 3`` and ``cam_token is None``,
pick a reference view via this strategy and move it to
position 0 right before the first alt-attention block.
The original view order is restored on the way out.
export_feat_layers: optional iterable of layer indices whose
local attention outputs to also return as auxiliary
features (``(B, S, N_patch, C)`` after final norm). Used
by the multi-view path to expose intermediate features
to the nested-architecture wrapper.
Returns:
``(layer_outputs, aux_outputs)`` where ``layer_outputs`` is a
list of ``(patch_tokens, cls_or_cam_token)`` tuples (one per
``out_layers`` entry) and ``aux_outputs`` is a list of
``(B, S, N_patch, C)`` features for ``export_feat_layers``
(empty list when not requested).
"""
if pixel_values.ndim == 4:
pixel_values = pixel_values.unsqueeze(1)
assert pixel_values.ndim == 5 and pixel_values.shape[2] == 3, \
f"expected (B,3,H,W) or (B,S,3,H,W); got {tuple(pixel_values.shape)}"
B, S, _, H, W = pixel_values.shape
# Patch + cls + (interpolated) pos embed for each view.
x = pixel_values.reshape(B * S, 3, H, W)
x = self.embeddings(x) # (B*S, 1+N, C)
x = x.reshape(B, S, x.shape[-2], x.shape[-1]) # (B, S, 1+N, C)
pos_local, pos_global = self._prepare_rope_positions(B, S, H, W, x.device)
# ``optimized_attention`` is only used by blocks without QK-norm/RoPE
# (vanilla DINOv2 path); enabling-aware blocks fall through to SDPA.
optimized_attention = optimized_attention_for_device(x.device, False, small_input=True)
out_set = set(out_layers)
export_set = set(export_feat_layers) if export_feat_layers else set()
outputs: list[torch.Tensor] = []
aux_outputs: list[torch.Tensor] = []
local_x = x
b_idx = None
for i, blk in enumerate(self.encoder.layer):
apply_rope = self.rope is not None and i >= self.rope_start
block_rope = self.rope if apply_rope else None
l_pos = pos_local if apply_rope else None
g_pos = pos_global if apply_rope else None
# Reference-view selection threshold: matches the upstream constant
# ``THRESH_FOR_REF_SELECTION = 3``. Skipped when a user-supplied
# cam_token is provided (camera info already pins the geometry).
if (self.alt_start != -1 and i == self.alt_start - 1
and S >= THRESH_FOR_REF_SELECTION and cam_token is None):
b_idx = select_reference_view(x, strategy=ref_view_strategy)
x = reorder_by_reference(x, b_idx)
local_x = reorder_by_reference(local_x, b_idx)
if self.alt_start != -1 and i == self.alt_start:
x = self._inject_camera_token(x, B, S, cam_token)
if self.alt_start != -1 and i >= self.alt_start and (i % 2 == 1):
# Global attention across views: flatten S into the seq dim.
t = x.reshape(B, S * x.shape[-2], x.shape[-1])
p = g_pos.reshape(B, S * g_pos.shape[-2], g_pos.shape[-1]) if g_pos is not None else None
t = blk(t, optimized_attention=optimized_attention, pos=p, rope=block_rope)
x = t.reshape(B, S, x.shape[-2], x.shape[-1])
else:
# Per-view local attention.
t = x.reshape(B * S, x.shape[-2], x.shape[-1])
p = l_pos.reshape(B * S, l_pos.shape[-2], l_pos.shape[-1]) if l_pos is not None else None
t = blk(t, optimized_attention=optimized_attention, pos=p, rope=block_rope)
x = t.reshape(B, S, x.shape[-2], x.shape[-1])
local_x = x
if i in out_set:
if self.cat_token:
out_x = torch.cat([local_x, x], dim=-1)
else:
out_x = x
# Restore original view order on the way out so heads see views
# in the user's expected order.
if b_idx is not None and self.alt_start != -1:
out_x = restore_original_order(out_x, b_idx)
outputs.append(out_x)
if i in export_set:
aux = x
if b_idx is not None and self.alt_start != -1:
aux = restore_original_order(aux, b_idx)
aux_outputs.append(aux)
# Apply final norm. When ``cat_token`` is set, only the right half
# ("global" features) is normalised; the left half is left as-is to
# match the upstream DA3 head signature.
normed: list[torch.Tensor] = []
cls_tokens: list[torch.Tensor] = []
for out_x in outputs:
cls_tokens.append(out_x[:, :, 0])
if out_x.shape[-1] == self.embed_dim:
normed.append(self.layernorm(out_x))
elif out_x.shape[-1] == self.embed_dim * 2:
left = out_x[..., :self.embed_dim]
right = self.layernorm(out_x[..., self.embed_dim:])
normed.append(torch.cat([left, right], dim=-1))
else:
raise ValueError(f"Unexpected token width: {out_x.shape[-1]}")
# Drop cls/cam token from the patch sequence.
normed = [o[..., 1 + self.num_register_tokens:, :] for o in normed]
# Final layernorm + drop cls token from auxiliary features too.
aux_normed = [self.layernorm(o)[..., 1 + self.num_register_tokens:, :]
for o in aux_outputs]
return list(zip(normed, cls_tokens)), aux_normed