File size: 12,027 Bytes
4f10e5a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
02a5e06
 
 
4f10e5a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
a867cc7
 
 
4f10e5a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
02a5e06
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4f10e5a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d901250
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
4f10e5a
 
 
 
 
02a5e06
4f10e5a
 
 
 
 
 
 
 
 
 
a867cc7
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
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
"""
Condition Encoder β€” reads config to select the right VAE, encodes all
conditions, and produces concat-ready latents.

Usage:
    config = ActionConditioningConfig(backbone="wan", vae_path="path/to/vae.pt")
    cond_enc = ConditionEncoder(config, device="cuda")

    encoded = cond_enc.encode(
        obs_image=obs,            # (B, 3, H, W)
        actions=delta_actions,    # (B, T, action_dim)
        masked_traj=traj_imgs,    # (B, 3, T, H, W)
    )
    # encoded.concat_latent: (B, C_concat, T', H', W') β€” ready for input_concat
    # encoded.action_tokens: (B, T, embed_dim) β€” for adaln or cross_attn
"""

from dataclasses import dataclass
from typing import Optional

import torch
import torch.nn as nn

from .config import ActionConditioningConfig
from .model_paths import resolve_vae_ckpt_path
# from .encoders.action_encoder import ActionFFNEncoder

# Prepended to visual_latent on channel dim when `noisy_latent` is passed (must match pipeline expectations).
VISUAL_CONDITION_MASK_CHANNELS = 4


class ActionFFNEncoder(nn.Module):
    def __init__(self, action_dim: int, embed_dim: int, num_layers: int = 2):
        super().__init__()
        layers = [nn.Linear(action_dim, embed_dim), nn.GELU()]
        for _ in range(max(0, num_layers - 2)):
            layers += [nn.Linear(embed_dim, embed_dim), nn.GELU()]
        self.mlp = nn.Sequential(*layers)
        self.norm = nn.LayerNorm(embed_dim)
    def forward(self, actions: torch.Tensor) -> torch.Tensor:
        return self.norm(self.mlp(actions))


@dataclass
class EncodedConditions:
    """All encoded conditions, ready for the DiT wrapper."""

    action_tokens: Optional[torch.Tensor] = None   # (B, T, embed_dim)
    obs_latent: Optional[torch.Tensor] = None       # (B, C, 1, H', W')
    traj_latent: Optional[torch.Tensor] = None      # (B, C, T', H', W')
    history_latent: Optional[torch.Tensor] = None    # (B, C, K', H', W')
    visual_latent: Optional[torch.Tensor] = None     # (B, C, T_vis, H', W') β€” unified visual condition
    concat_latent: Optional[torch.Tensor] = None     # (B, C_total, T', H', W') β€” all input_concat streams merged


def _build_vae(config: ActionConditioningConfig):
    # model-name driven
    if config.vae_model_name == "wan_video_vae":
        from diffsynth.models.wan_video_vae import WanVideoVAE
        vae = WanVideoVAE(z_dim=config.vae_z_dim)
    elif config.vae_model_name == "wan_video_vae38":
        from diffsynth.models.wan_video_vae import WanVideoVAE38
        vae = WanVideoVAE38(z_dim=config.vae_z_dim)
    else:
        raise ValueError(f"Unknown vae_model_name: {config.vae_model_name!r}")

    # optional load from explicit path or model_root-derived path
    vae_ckpt_path = resolve_vae_ckpt_path(config)
    missing_keys = []
    unexpected_keys = []
    if vae_ckpt_path is not None:
        from diffsynth.core.loader.file import load_state_dict
        state_dict = load_state_dict(vae_ckpt_path, device="cpu")
        if hasattr(vae, "state_dict_converter"):
            converter = vae.state_dict_converter()
            state_dict = converter.from_civitai(state_dict)
        incompatible = vae.load_state_dict(state_dict, strict=False)
        if incompatible is not None:
            missing_keys = list(getattr(incompatible, "missing_keys", []))
            unexpected_keys = list(getattr(incompatible, "unexpected_keys", []))
    else:
        raise ValueError(
            "VAE checkpoint path is not resolved. Please set config.vae_path "
            "or config.model_root with a valid Wan VAE file."
        )

    vae.eval()
    for p in vae.parameters():
        p.requires_grad = False
    # Attach debug metadata for verification in tests/scripts.
    vae._loaded_ckpt_path = vae_ckpt_path
    vae._missing_keys = missing_keys
    vae._unexpected_keys = unexpected_keys
    return vae


class ConditionEncoder(nn.Module):
    """
    Reads config.backbone to select the right VAE, encodes all visual
    conditions through it, and produces concat-ready latents.

    The VAE is frozen. Only the action_encoder is trainable.
    """

    def __init__(self, config: ActionConditioningConfig, device: torch.device = torch.device("cuda")):
        super().__init__()
        self.config = config
        self.device = device

        # Build frozen VAE (not registered as submodule to avoid saving its weights)
        self._vae = _build_vae(config)

        # Trainable action encoder
        self.action_encoder = ActionFFNEncoder(
            action_dim=config.action_dim,
            embed_dim=config.action_embed_dim,
            num_layers=config.action_num_layers,
        )

    @property
    def vae(self):
        return self._vae

    def to(self, *args, **kwargs):
        """Override to also move the frozen VAE."""
        super().to(*args, **kwargs)
        self._vae = self._vae.to(*args, **kwargs)
        return self

    # ------------------------------------------------------------------
    # VAE encoding helpers
    # ------------------------------------------------------------------

    @torch.no_grad()
    def _vae_encode_image(self, image: torch.Tensor) -> torch.Tensor:
        """(B, 3, H, W) β†’ (B, C, 1, H', W')"""
        videos = [img.unsqueeze(1) for img in image]
        return self._vae.encode(videos, self.device)

    @torch.no_grad()
    def _vae_encode_video(self, video: torch.Tensor) -> torch.Tensor:
        """(B, 3, T, H, W) β†’ (B, C, T', H', W')"""
        videos = [v for v in video]
        return self._vae.encode(videos, self.device)

    # ------------------------------------------------------------------
    # Main encode method
    # ------------------------------------------------------------------

    def encode(
        self,
        obs_image: Optional[torch.Tensor] = None,
        actions: Optional[torch.Tensor] = None,
        masked_traj: Optional[torch.Tensor] = None,
        history: Optional[torch.Tensor] = None,
        noisy_latent: Optional[torch.Tensor] = None,
    ) -> EncodedConditions:
        """
        Encode all conditions and assemble concat_latent.

        Parameters
        ----------
        obs_image : (B, 3, H, W) β€” observation frame
        actions : (B, T, action_dim) β€” delta action vectors
        masked_traj : (B, 3, T, H, W) β€” rendered trajectory images
        history : (B, 3, K, H, W) β€” past frames
        noisy_latent : (B, C, T', H', W') β€” the noisy latent (needed for
            shape reference when broadcasting obs to T' frames)

        Returns
        -------
        EncodedConditions
        """
        result = EncodedConditions()
        cfg = self.config

        # --- Action tokens (trainable) ---
        if actions is not None:
            result.action_tokens = self.action_encoder(actions)

        # --- VAE encode visual conditions (frozen) ---
        # Past context rule:
        # 1) force history length to cfg.history_frames
        # 2) past stream is [observation + history] (total cfg.history_frames + 1 frames)
        # 3) if history is missing, repeat observation to fill all past frames
        if obs_image is not None and (cfg.obs_injection is not None or cfg.history_injection is not None):
            target_past_len = cfg.history_frames + 1
            obs_seq = obs_image.unsqueeze(2)  # (B, 3, 1, H, W)
            if history is None:
                past_video = obs_seq.repeat(1, 1, target_past_len, 1, 1)
            else:
                # history: (B, 3, K, H, W) -> normalize to exactly cfg.history_frames
                if history.shape[2] >= cfg.history_frames:
                    history_norm = history[:, :, :cfg.history_frames]
                else:
                    pad_len = cfg.history_frames - history.shape[2]
                    pad_hist = obs_seq.repeat(1, 1, pad_len, 1, 1)
                    history_norm = torch.cat([history, pad_hist], dim=2)
                past_video = torch.cat([obs_seq, history_norm], dim=2)
            result.history_latent = self._vae_encode_video(past_video)
            result.obs_latent = None
        elif obs_image is not None and cfg.obs_injection is not None:
            result.obs_latent = self._vae_encode_image(obs_image)

        if cfg.traj_injection is not None:
            if masked_traj is not None:
                result.traj_latent = self._vae_encode_video(masked_traj)
            elif noisy_latent is not None:
                # Placeholder trajectory latent for true no-traj ablation.
                # Keeps visual_latent temporal length aligned with noisy_latent.
                B, C, T_lat, H_lat, W_lat = noisy_latent.shape

                # history_latent is usually obs + history compressed to T=1
                T_hist = result.history_latent.shape[2] if result.history_latent is not None else 0
                T_traj = T_lat - T_hist

                if T_traj < 0:
                    raise ValueError(
                        f"history latent is longer than noisy latent: "
                        f"T_hist={T_hist}, T_lat={T_lat}"
                    )

                result.traj_latent = torch.zeros(
                    B, C, T_traj, H_lat, W_lat,
                    device=noisy_latent.device,
                    dtype=noisy_latent.dtype,
                )

        if history is not None and cfg.history_injection is not None and result.history_latent is None:
            result.history_latent = self._vae_encode_video(history)

        # --- Assemble unified visual condition ---
        # Merge observation into history first, then concat with trajectory.
        # This keeps one "past context" stream + one "future trajectory" stream.
        visual_parts = []

        history_merged = result.history_latent
        if result.obs_latent is not None:
            if history_merged is None:
                history_merged = result.obs_latent
            else:
                history_merged = torch.cat([result.obs_latent, history_merged], dim=2)

        if result.traj_latent is not None:
            visual_parts.append(result.traj_latent)
        if history_merged is not None:
            visual_parts.append(history_merged)

        if visual_parts:
            result.visual_latent = torch.cat(visual_parts, dim=2)

        # Wan DiT does `torch.cat([latents, y], dim=1)` β€” T and spatial sizes must match latents.
        # Past-only history often yields T_vis=1 while generation uses T_lat=(num_frames-1)//4+1.
        if result.visual_latent is not None and noisy_latent is not None:
            T_lat = noisy_latent.shape[2]
            Bv, Cv, Tv, Hv, Wv = result.visual_latent.shape
            if Tv != T_lat:
                if Tv < T_lat:
                    pad_t = T_lat - Tv
                    tail = torch.zeros(
                        Bv, Cv, pad_t, Hv, Wv,
                        device=result.visual_latent.device,
                        dtype=result.visual_latent.dtype,
                    )
                    result.visual_latent = torch.cat([result.visual_latent, tail], dim=2)
                else:
                    result.visual_latent = result.visual_latent[:, :, :T_lat, :, :]

        # --- Build condition mask (4-channel, per sub-frame) ---
        if result.visual_latent is not None and noisy_latent is not None:
            B = noisy_latent.shape[0]
            _, C_vis, T_vis, H_lat, W_lat = result.visual_latent.shape  # ← 用 visual_latent ηš„η»΄εΊ¦

            mask = torch.zeros(B, VISUAL_CONDITION_MASK_CHANNELS, T_vis, H_lat, W_lat,  # ← T_vis θ€ŒδΈζ˜― T_lat
                               device=result.visual_latent.device,
                               dtype=result.visual_latent.dtype)

            n_traj_pixels = masked_traj.shape[2] if masked_traj is not None else 0
            real_start_lat = (n_traj_pixels + 3) // 4
            mask[:, :, real_start_lat:] = 1.0

            result.visual_latent = torch.cat([mask, result.visual_latent], dim=1)
            result.concat_latent = result.visual_latent

        return result