File size: 10,922 Bytes
b47a1ce
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1dae740
 
b47a1ce
 
 
 
 
 
 
 
 
 
 
1dae740
 
b47a1ce
 
1dae740
 
 
b47a1ce
1dae740
 
 
b47a1ce
 
 
 
1dae740
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
from __future__ import annotations

import math
from typing import Optional

import torch
import torch.nn.functional as F
from torch import nn


def latent_patch_tokens(latents: torch.Tensor, patch_size: int) -> torch.Tensor:
    if latents.ndim != 5:
        raise ValueError("latents must have shape (T,B,C,H,W)")
    if patch_size <= 0:
        raise ValueError("patch_size must be positive")
    T, B, C, H, W = latents.shape
    if H % patch_size != 0 or W % patch_size != 0:
        raise ValueError(f"latent H,W=({H},{W}) must be divisible by patch_size={patch_size}")
    flat = latents.reshape(T * B, C, H, W)
    patches = F.unfold(flat, kernel_size=patch_size, stride=patch_size).transpose(1, 2).contiguous()
    return patches.reshape(T, B, patches.shape[1], C * patch_size * patch_size)


def spatial_pool_tokens(
    tokens: torch.Tensor,
    pool_h: int,
    pool_w: int,
    src_h: int,
    src_w: int,
) -> torch.Tensor:
    """2D adaptive average pool on a flattened (src_h*src_w, D) token grid.
    Preserves 2D spatial layout. Returns (pool_h*pool_w, D)."""
    if tokens.ndim != 2:
        raise ValueError("tokens must have shape (N, D)")
    D = tokens.shape[-1]
    spatial = tokens.reshape(src_h, src_w, D).permute(2, 0, 1).unsqueeze(0)
    pooled = F.adaptive_avg_pool2d(spatial, (pool_h, pool_w))
    return pooled.squeeze(0).permute(1, 2, 0).reshape(-1, D)


class SpatialConv2DMemoryProjector(nn.Module):
    """Project latent maps to DiT hidden tokens while preserving the HxW grid."""

    projects_spatial_latents = True

    def __init__(
        self,
        latent_channels: int,
        dit_hidden_size: int,
        mid_channels: int,
        kernel_size: int = 3,
    ):
        super().__init__()
        kernel_size = int(kernel_size)
        if kernel_size <= 0 or kernel_size % 2 == 0:
            raise ValueError("kernel_size must be a positive odd integer")
        self.latent_channels = int(latent_channels)
        self.dit_hidden_size = int(dit_hidden_size)
        self.mid_channels = int(mid_channels)
        self.kernel_size = kernel_size
        self.out_features = self.dit_hidden_size
        self.proj_in = nn.Conv2d(self.latent_channels, self.mid_channels, kernel_size=1)
        self.proj_spatial = nn.Conv2d(
            self.mid_channels,
            self.dit_hidden_size,
            kernel_size=kernel_size,
            padding=kernel_size // 2,
        )

    def forward(self, latents: torch.Tensor) -> torch.Tensor:
        if latents.ndim != 5:
            raise ValueError("latents must have shape (T,B,C,H,W)")
        T, B, C, H, W = latents.shape
        if C != self.latent_channels:
            raise ValueError(f"expected {self.latent_channels} latent channels, got {C}")
        x = latents.reshape(T * B, C, H, W)
        x = self.proj_spatial(self.proj_in(x))
        x = x.reshape(T, B, self.dit_hidden_size, H, W)
        return x.permute(1, 0, 3, 4, 2).reshape(B, T, H * W, self.dit_hidden_size).contiguous()


class CausalConv3DDynamicCompressor(nn.Module):
    """Dynamic memory compressor: delta preprocessing + causal Conv3D on raw latents.

    Replaces ShortTermLatentCompressor (slot cross-attention).
    - Operates directly on (T, C, H, W) raw latents
    - Delta: inp[0]=latent[0], inp[t]=latent[t]-latent[t-1]
    - Causal padding prepends temporal zeros and right-aligns fixed outputs
    - Zero-padded to max_source_frames for fixed output shape
    - No slot cross-attention, no chunking
    """

    def __init__(
        self,
        latent_channels: int,
        dit_hidden_size: int,
        patch_size: int = 2,
        conv_kernel_t: int = 3,
        conv_stride_t: int = 2,
        max_source_frames: int = 8,
        exclude_latest_local_frames: int = 4,
    ):
        super().__init__()
        self.latent_channels = latent_channels
        self.dit_hidden_size = dit_hidden_size
        self.patch_size = patch_size
        self.conv_kernel_t = conv_kernel_t
        self.conv_stride_t = conv_stride_t
        self.max_source_frames = max_source_frames
        self.exclude_latest_local_frames = int(exclude_latest_local_frames)
        self.causal_pad = self._temporal_left_pad()
        self.conv3d = nn.Conv3d(
            latent_channels, dit_hidden_size,
            kernel_size=(conv_kernel_t, patch_size, patch_size),
            stride=(conv_stride_t, patch_size, patch_size),
            padding=0,
        )
        self.out_norm = nn.LayerNorm(dit_hidden_size)
        self._init_temporal_as_delta()

    def _init_temporal_as_delta(self) -> None:
        with torch.no_grad():
            self.conv3d.weight.zero_()
            k_t, p = self.conv_kernel_t, self.patch_size
            D_out, D_in = self.conv3d.weight.shape[:2]
            scale = 1.0 / (p * p)
            # Delta preprocessing happens in forward. Initialize every output
            # channel to read a patch-averaged current delta, repeating latent
            # channels across the wider DiT hidden dimension.
            for d in range(D_out):
                self.conv3d.weight[d, d % D_in, k_t - 1, :, :] = scale
            if self.conv3d.bias is not None:
                nn.init.zeros_(self.conv3d.bias)

    def _temporal_output_count(self) -> int:
        return math.ceil(self.max_source_frames / self.conv_stride_t)

    def _temporal_left_pad(self) -> int:
        t_out = self._temporal_output_count()
        latest_output_end = (t_out - 1) * self.conv_stride_t + self.conv_kernel_t - 1
        latest_source = self.max_source_frames - 1
        return max(0, latest_output_end - latest_source)

    def _output_time_indices(self, device: torch.device) -> torch.Tensor:
        t_out = self._temporal_output_count()
        return (
            torch.arange(t_out, device=device, dtype=torch.long) * self.conv_stride_t
            + self.conv_kernel_t
            - 1
            - self.causal_pad
        )

    def tokens_per_target(self, H: int, W: int) -> int:
        p = self.patch_size
        T_out = self._temporal_output_count()
        return T_out * (H // p) * (W // p)

    def forward(
        self,
        latents: torch.Tensor,
        frame_indices: torch.Tensor,
        pose: Optional[torch.Tensor],
        target_frame_indices: torch.Tensor,
        source_is_generated: Optional[torch.Tensor] = None,
        exclude_latest_local_frames: Optional[int] = None,
    ) -> tuple[torch.Tensor, torch.Tensor]:
        del pose, source_is_generated
        if latents.ndim != 5:
            raise ValueError("latents must have shape (T_src,B,C,H,W)")
        exclude_latest_local_frames = (
            self.exclude_latest_local_frames
            if exclude_latest_local_frames is None
            else int(exclude_latest_local_frames)
        )
        T_src, B, C, H, W = latents.shape
        p = self.patch_size
        if H % p != 0 or W % p != 0:
            raise ValueError(f"latent H,W=({H},{W}) must be divisible by patch_size={p}")
        if frame_indices.shape != (T_src, B):
            raise ValueError("frame_indices must have shape (T_src,B)")
        if target_frame_indices.ndim == 1:
            target_frame_indices = target_frame_indices[:, None].expand(-1, B)
        if target_frame_indices.ndim != 2 or target_frame_indices.shape[1] != B:
            raise ValueError("target_frame_indices must have shape (T_tgt,B)")

        device = latents.device
        frame_indices = frame_indices.to(device=device)
        target_frame_indices = target_frame_indices.to(device=device)
        T_tgt = target_frame_indices.shape[0]
        n_spatial = (H // p) * (W // p)
        T_out = self._temporal_output_count()
        num_slots = T_out * n_spatial
        output_time_idx = self._output_time_indices(device)
        if T_src == 0:
            out_tokens = latents.new_zeros((B, T_tgt, num_slots, self.dit_hidden_size))
            out_mask = torch.zeros((B, T_tgt, num_slots), device=device, dtype=torch.bool)
            return out_tokens, out_mask

        source_frames = frame_indices.transpose(0, 1).contiguous()
        target_frames = target_frame_indices.transpose(0, 1).contiguous()
        valid = source_frames[:, None, :] < (target_frames[:, :, None] - int(exclude_latest_local_frames))
        valid_flat = valid.reshape(B * T_tgt, T_src)
        source_frames_flat = source_frames[:, None, :].expand(B, T_tgt, T_src).reshape(B * T_tgt, T_src)

        topk = min(int(self.max_source_frames), T_src)
        rank = source_frames_flat.to(dtype=torch.float64).masked_fill(~valid_flat, -float("inf"))
        top = torch.topk(rank, k=topk, dim=1, largest=True, sorted=True)
        selected_idx = top.indices.flip(dims=(1,))
        selected_valid = torch.isfinite(top.values).flip(dims=(1,))
        if topk < self.max_source_frames:
            pad_count = self.max_source_frames - topk
            selected_idx = torch.cat([
                torch.zeros((B * T_tgt, pad_count), device=device, dtype=torch.long),
                selected_idx,
            ], dim=1)
            selected_valid = torch.cat([
                torch.zeros((B * T_tgt, pad_count), device=device, dtype=torch.bool),
                selected_valid,
            ], dim=1)

        selected_idx_clamped = selected_idx.to(device=device, dtype=torch.long).clamp(min=0, max=max(0, T_src - 1))
        has_valid = selected_valid.any(dim=1)
        batch_ids = torch.arange(B, device=device, dtype=torch.long).repeat_interleave(T_tgt)
        latents_by_batch = latents.permute(1, 0, 2, 3, 4).contiguous()
        latents_per_query = latents_by_batch.index_select(0, batch_ids)
        gather_idx = selected_idx_clamped.reshape(B * T_tgt, self.max_source_frames, 1, 1, 1).expand(
            -1, -1, C, H, W
        )
        chunk = torch.gather(latents_per_query, 1, gather_idx)
        chunk = torch.where(
            selected_valid[:, :, None, None, None],
            chunk,
            torch.zeros((), device=device, dtype=latents.dtype),
        )

        inp = chunk.clone()
        inp[:, 1:] = chunk[:, 1:] - chunk[:, :-1]
        x = inp.permute(0, 2, 1, 3, 4)
        x = F.pad(x, (0, 0, 0, 0, self.causal_pad, 0))
        x = self.conv3d(x)
        x = self.out_norm(x.permute(0, 2, 3, 4, 1))
        tokens_flat = x.reshape(B * T_tgt, num_slots, self.dit_hidden_size)
        tokens_flat = torch.where(has_valid[:, None, None], tokens_flat, torch.zeros_like(tokens_flat))
        out_tokens = tokens_flat.reshape(B, T_tgt, num_slots, self.dit_hidden_size)

        clamped_time_idx = output_time_idx.clamp(min=0, max=self.max_source_frames - 1)
        temporal_mask = (
            (output_time_idx >= 0)
            & (output_time_idx < self.max_source_frames)
            & selected_valid.index_select(1, clamped_time_idx)
        )
        out_mask = temporal_mask[:, :, None].expand(B * T_tgt, T_out, n_spatial).reshape(B, T_tgt, num_slots)
        return out_tokens, out_mask