ComfyUI/comfy/ldm/sam3d_body/model/dinov3.py
2026-05-26 02:15:15 +03:00

251 lines
10 KiB
Python

# DINOv3 ViT-H+ backbone for SAM 3D Body.
#
# Single-file consolidation of the inference path. SAM 3D Body only ships a
# `dinov3_vith16plus` checkpoint, so the architecture is hardcoded rather
# than reconstructed from Hydra-flavoured configs.
#
# Adapted from facebookresearch/dinov3 (DINOv3 License Agreement). Trimmed
# to what's actually exercised at inference: no multi-crop training path,
# no DINOHead, no causal blocks, no rmsnorm/Mlp variants, no rope shift /
# jitter / rescale (training-time augmentations).
#TODO: Unify with TRELLIS2
import math
from typing import Optional, Tuple
import torch
import torch.nn.functional as F
from comfy.ldm.modules.attention import optimized_attention
from torch import Tensor, nn
# DINOv3 ViT-H+ architecture constants.
EMBED_DIM = 1280
DEPTH = 32
NUM_HEADS = 20
FFN_RATIO = 6.0
PATCH_SIZE = 16
LAYERSCALE_INIT = 1.0e-5
N_STORAGE_TOKENS = 4
LAYERNORM_EPS = 1e-5 # "layernormbf16" preset uses 1e-5
ROPE_BASE = 100.0
# RoPE (axial sin/cos, no learnable weights)
def _rotate_half(x: Tensor) -> Tensor:
x1, x2 = x.chunk(2, dim=-1)
return torch.cat([-x2, x1], dim=-1)
def _apply_rope(x: Tensor, sin: Tensor, cos: Tensor) -> Tensor:
return x * cos + _rotate_half(x) * sin
class RopePositionEmbedding(nn.Module):
"""Axial RoPE for 2D patch grids; periods buffer is deterministic."""
def __init__(self, embed_dim: int, num_heads: int, dtype=torch.float32, device=None):
super().__init__()
assert embed_dim % (4 * num_heads) == 0
D_head = embed_dim // num_heads
# Periods are persistent so they round-trip through state_dict, but the
# values are deterministic from D_head/base; load_state_dict will
# overwrite this with the saved buffer either way.
periods = ROPE_BASE ** (
2 * torch.arange(D_head // 4, dtype=dtype, device=device) / (D_head // 2)
)
self.register_buffer("periods", periods, persistent=True)
self._dtype = dtype
def forward(self, H: int, W: int) -> Tuple[Tensor, Tensor]:
device, dtype = self.periods.device, self._dtype
coords_h = torch.arange(0.5, H, device=device, dtype=dtype) / H
coords_w = torch.arange(0.5, W, device=device, dtype=dtype) / W
coords = torch.stack(torch.meshgrid(coords_h, coords_w, indexing="ij"), dim=-1)
coords = 2.0 * coords.flatten(0, 1) - 1.0 # [HW, 2] in [-1, +1]
angles = 2 * math.pi * coords[:, :, None] / self.periods[None, None, :]
angles = angles.flatten(1, 2).tile(2) # [HW, D_head]
return torch.sin(angles), torch.cos(angles)
def _apply_rope_to_qk(q: Tensor, k: Tensor, rope: Tuple[Tensor, Tensor]):
"""Apply RoPE only to the patch-token slice (skip CLS + storage tokens)."""
sin, cos = rope
rope_dtype = sin.dtype
q_dtype, k_dtype = q.dtype, k.dtype
q = q.to(rope_dtype)
k = k.to(rope_dtype)
prefix = q.shape[-2] - sin.shape[-2]
q_pre, q_rope = q[..., :prefix, :], q[..., prefix:, :]
k_pre, k_rope = k[..., :prefix, :], k[..., prefix:, :]
q = torch.cat([q_pre, _apply_rope(q_rope, sin, cos)], dim=-2)
k = torch.cat([k_pre, _apply_rope(k_rope, sin, cos)], dim=-2)
return q.to(q_dtype), k.to(k_dtype)
# Layers
class LayerScale(nn.Module):
def __init__(self, dim: int, init_values: float, device=None, dtype=None):
super().__init__()
self.gamma = nn.Parameter(
torch.full((dim,), init_values, device=device, dtype=dtype)
)
def forward(self, x: Tensor) -> Tensor:
return x * self.gamma
class SwiGLUFFN(nn.Module):
"""w3(silu(w1(x)) * w2(x))."""
def __init__(self, in_features: int, hidden_features: int, align_to: int = 8,
device=None, dtype=None, operations=None):
super().__init__()
ops = operations if operations is not None else nn
d = int(hidden_features * 2 / 3)
h = d + (-d % align_to)
self.w1 = ops.Linear(in_features, h, bias=True, device=device, dtype=dtype)
self.w2 = ops.Linear(in_features, h, bias=True, device=device, dtype=dtype)
self.w3 = ops.Linear(h, in_features, bias=True, device=device, dtype=dtype)
def forward(self, x: Tensor) -> Tensor:
return self.w3(F.silu(self.w1(x)) * self.w2(x))
class SelfAttention(nn.Module):
def __init__(self, dim: int, num_heads: int, device=None, dtype=None, operations=None):
super().__init__()
ops = operations if operations is not None else nn
self.num_heads = num_heads
# DINOv3's `mask_k_bias` zeroes the K third of qkv.bias. The mask is
# deterministic from out_features, so the loader applies it in-place
# once after load_state_dict (see `apply_dinov3_qkv_bias_mask`) and the
# forward stays a plain F.linear.
self.qkv = ops.Linear(dim, dim * 3, bias=True, device=device, dtype=dtype)
self.proj = ops.Linear(dim, dim, bias=True, device=device, dtype=dtype)
def forward(self, x: Tensor, rope: Optional[Tuple[Tensor, Tensor]] = None) -> Tensor:
B, N, C = x.shape
qkv = self.qkv(x).reshape(B, N, 3, self.num_heads, C // self.num_heads)
q, k, v = qkv.unbind(2)
q, k, v = (t.transpose(1, 2) for t in (q, k, v))
if rope is not None:
q, k = _apply_rope_to_qk(q, k, rope)
# low_precision_attention=False forces attention_sage (when enabled
# globally in comfy) to fall back to pytorch SDPA. SAM 3D Body's
# regression heads (camera projection, MHR rig math) are sensitive
# to attention output precision; sage's int8/fp8 path drifts the
# keypoints and mesh visibly.
x = optimized_attention(
q, k, v, self.num_heads, skip_reshape=True,
low_precision_attention=False,
)
return self.proj(x)
class Block(nn.Module):
def __init__(self, dim: int, num_heads: int, ffn_ratio: float,
device=None, dtype=None, operations=None):
super().__init__()
ops = operations if operations is not None else nn
self.norm1 = ops.LayerNorm(dim, eps=LAYERNORM_EPS, device=device, dtype=dtype)
self.attn = SelfAttention(dim, num_heads, device=device, dtype=dtype, operations=operations)
self.ls1 = LayerScale(dim, LAYERSCALE_INIT, device=device, dtype=dtype)
self.norm2 = ops.LayerNorm(dim, eps=LAYERNORM_EPS, device=device, dtype=dtype)
self.mlp = SwiGLUFFN(dim, int(dim * ffn_ratio), device=device, dtype=dtype, operations=operations)
self.ls2 = LayerScale(dim, LAYERSCALE_INIT, device=device, dtype=dtype)
def forward(self, x: Tensor, rope=None) -> Tensor:
x = x + self.ls1(self.attn(self.norm1(x), rope=rope))
x = x + self.ls2(self.mlp(self.norm2(x)))
return x
class PatchEmbed(nn.Module):
def __init__(self, in_chans=3, embed_dim=EMBED_DIM, patch_size=PATCH_SIZE,
device=None, dtype=None, operations=None):
super().__init__()
ops = operations if operations is not None else nn
self.proj = ops.Conv2d(
in_chans, embed_dim,
kernel_size=patch_size, stride=patch_size,
device=device, dtype=dtype,
)
# Encoder + wrapper
class _DinoEncoder(nn.Module):
"""Inner ViT module. Held under `Dinov3Backbone.encoder` so state_dict
keys (`backbone.encoder.*`) match the upstream layout."""
def __init__(self, device=None, dtype=None, operations=None):
super().__init__()
ops = operations if operations is not None else nn
self.patch_size = PATCH_SIZE
self.embed_dim = EMBED_DIM
self.patch_embed = PatchEmbed(
embed_dim=EMBED_DIM, patch_size=PATCH_SIZE,
device=device, dtype=dtype, operations=operations,
)
self.cls_token = nn.Parameter(torch.empty(1, 1, EMBED_DIM, device=device, dtype=dtype))
self.storage_tokens = nn.Parameter(
torch.empty(1, N_STORAGE_TOKENS, EMBED_DIM, device=device, dtype=dtype)
)
# The released config sets pos_embed_rope_dtype="fp32"; periods stays
# in fp32 regardless of the backbone weight dtype.
self.rope_embed = RopePositionEmbedding(EMBED_DIM, NUM_HEADS, dtype=torch.float32, device=device)
self.blocks = nn.ModuleList([
Block(EMBED_DIM, NUM_HEADS, FFN_RATIO, device=device, dtype=dtype, operations=operations)
for _ in range(DEPTH)
])
self.norm = ops.LayerNorm(EMBED_DIM, eps=LAYERNORM_EPS, device=device, dtype=dtype)
def forward(self, x: Tensor) -> Tensor:
x = self.patch_embed.proj(x) # (B, embed_dim, H, W)
B, _, H, W = x.shape
x = x.flatten(2).transpose(1, 2) # (B, H*W, embed_dim)
# Prepend CLS + storage tokens.
x = torch.cat([
self.cls_token.expand(B, -1, -1),
self.storage_tokens.expand(B, -1, -1),
x,
], dim=1)
rope = self.rope_embed(H=H, W=W)
for blk in self.blocks:
x = blk(x, rope)
x = self.norm(x)
# Drop CLS + storage tokens; reshape patch grid to (B, C, H, W).
x = x[:, 1 + N_STORAGE_TOKENS :]
return x.reshape(B, H, W, EMBED_DIM).permute(0, 3, 1, 2).contiguous()
class Dinov3Backbone(nn.Module):
"""Public backbone interface used by SAM3DBody."""
def __init__(self, device=None, dtype=None, operations=None):
super().__init__()
self.encoder = _DinoEncoder(device=device, dtype=dtype, operations=operations)
self.patch_size = PATCH_SIZE
self.embed_dim = self.embed_dims = EMBED_DIM
def forward(self, x: Tensor) -> Tensor:
return self.encoder(x)
def apply_dinov3_qkv_bias_mask(backbone: "Dinov3Backbone") -> None:
"""Zero the K third of every block's qkv.bias in-place.
Implements DINOv3's `mask_k_bias` once at load time so the per-block forward
stays a plain F.linear instead of cloning + slicing the bias every call.
"""
for blk in backbone.encoder.blocks:
qkv = blk.attn.qkv
if qkv.bias is not None:
o = qkv.out_features
qkv.bias.data[o // 3 : 2 * o // 3] = 0