daquanzhou
merge github repos and lfs track ckpt/path/safetensors/pt
613c9ab
raw
history blame
41.6 kB
import math
from typing import Iterable, Tuple, Union
import re
import torch
from einops import rearrange, repeat
from torch import Tensor, nn
from comfy.ldm.modules.attention import FeedForward, SpatialTransformer
from comfy.model_patcher import ModelPatcher
from comfy.ldm.modules.diffusionmodules import openaimodel
from comfy.ldm.modules.diffusionmodules.openaimodel import SpatialTransformer
from comfy.controlnet import broadcast_image_to
from comfy.utils import repeat_to_batch_size
import comfy.ops
import comfy.model_management
from .context import ContextFuseMethod, ContextOptions, get_context_weights, get_context_windows
from .utils_motion import CrossAttentionMM, MotionCompatibilityError, extend_to_batch_size, prepare_mask_batch
from .utils_model import BetaSchedules, ModelTypeSD
from .logger import logger
def zero_module(module):
# Zero out the parameters of a module and return it.
for p in module.parameters():
p.detach().zero_()
return module
class AnimateDiffFormat:
ANIMATEDIFF = "AnimateDiff"
HOTSHOTXL = "HotshotXL"
ANIMATELCM = "AnimateLCM"
class AnimateDiffVersion:
V1 = "v1"
V2 = "v2"
V3 = "v3"
class AnimateDiffInfo:
def __init__(self, sd_type: str, mm_format: str, mm_version: str, mm_name: str):
self.sd_type = sd_type
self.mm_format = mm_format
self.mm_version = mm_version
self.mm_name = mm_name
def get_string(self):
return f"{self.mm_name}:{self.mm_version}:{self.mm_format}:{self.sd_type}"
def is_hotshotxl(mm_state_dict: dict[str, Tensor]) -> bool:
# use pos_encoder naming to determine if hotshotxl model
for key in mm_state_dict.keys():
if key.endswith("pos_encoder.positional_encoding"):
return True
return False
def is_animatelcm(mm_state_dict: dict[str, Tensor]) -> bool:
# use lack of ANY pos_encoder keys to determine if animatelcm model
for key in mm_state_dict.keys():
if "pos_encoder" in key:
return False
return True
def get_down_block_max(mm_state_dict: dict[str, Tensor]) -> int:
# keep track of biggest down_block count in module
biggest_block = 0
for key in mm_state_dict.keys():
if "down_blocks" in key:
try:
block_int = key.split(".")[1]
block_num = int(block_int)
if block_num > biggest_block:
biggest_block = block_num
except ValueError:
pass
return biggest_block
def has_mid_block(mm_state_dict: dict[str, Tensor]):
# check if keys contain mid_block
for key in mm_state_dict.keys():
if key.startswith("mid_block."):
return True
return False
def get_position_encoding_max_len(mm_state_dict: dict[str, Tensor], mm_name: str, mm_format: str) -> Union[int, None]:
# use pos_encoder.pe entries to determine max length - [1, {max_length}, {320|640|1280}]
for key in mm_state_dict.keys():
if key.endswith("pos_encoder.pe"):
return mm_state_dict[key].size(1) # get middle dim
# AnimateLCM models should have no pos_encoder entries, and assumed to be 64
if mm_format == AnimateDiffFormat.ANIMATELCM:
return 64
raise MotionCompatibilityError(f"No pos_encoder.pe found in mm_state_dict - {mm_name} is not a valid AnimateDiff motion module!")
_regex_hotshotxl_module_num = re.compile(r'temporal_attentions\.(\d+)\.')
def find_hotshot_module_num(key: str) -> Union[int, None]:
found = _regex_hotshotxl_module_num.search(key)
if found:
return int(found.group(1))
return None
def normalize_ad_state_dict(mm_state_dict: dict[str, Tensor], mm_name: str) -> Tuple[dict[str, Tensor], AnimateDiffInfo]:
# from pathlib import Path
# with open(Path(__file__).parent.parent.parent / f"keys_{mm_name}.txt", "w") as afile:
# for key, value in mm_state_dict.items():
# afile.write(f"{key}:\t{value.shape}\n")
# remove all non-temporal keys (in case model has extra stuff in it)
for key in list(mm_state_dict.keys()):
if "temporal" not in key:
del mm_state_dict[key]
# determine what SD model the motion module is intended for
sd_type: str = None
down_block_max = get_down_block_max(mm_state_dict)
if down_block_max == 3:
sd_type = ModelTypeSD.SD1_5
elif down_block_max == 2:
sd_type = ModelTypeSD.SDXL
else:
raise ValueError(f"'{mm_name}' is not a valid SD1.5 nor SDXL motion module - contained {down_block_max} downblocks.")
# determine the model's format
mm_format = AnimateDiffFormat.ANIMATEDIFF
if is_hotshotxl(mm_state_dict):
mm_format = AnimateDiffFormat.HOTSHOTXL
if is_animatelcm(mm_state_dict):
mm_format = AnimateDiffFormat.ANIMATELCM
# determine the model's version
mm_version = AnimateDiffVersion.V1
if has_mid_block(mm_state_dict):
mm_version = AnimateDiffVersion.V2
elif sd_type==ModelTypeSD.SD1_5 and get_position_encoding_max_len(mm_state_dict, mm_name, mm_format)==32:
mm_version = AnimateDiffVersion.V3
info = AnimateDiffInfo(sd_type=sd_type, mm_format=mm_format, mm_version=mm_version, mm_name=mm_name)
# convert to AnimateDiff format, if needed
if mm_format == AnimateDiffFormat.HOTSHOTXL:
# HotshotXL is AD-based architecture applied to SDXL instead of SD1.5
# By renaming the keys, no code needs to be adapted at all
#
# reformat temporal_attentions:
# HSXL: temporal_attentions.#.
# AD: motion_modules.#.temporal_transformer.
# HSXL: pos_encoder.positional_encoding
# AD: pos_encoder.pe
for key in list(mm_state_dict.keys()):
module_num = find_hotshot_module_num(key)
if module_num is not None:
new_key = key.replace(f"temporal_attentions.{module_num}",
f"motion_modules.{module_num}.temporal_transformer", 1)
new_key = new_key.replace("pos_encoder.positional_encoding", "pos_encoder.pe")
mm_state_dict[new_key] = mm_state_dict[key]
del mm_state_dict[key]
# return adjusted mm_state_dict and info
return mm_state_dict, info
class BlockType:
UP = "up"
DOWN = "down"
MID = "mid"
class AnimateDiffModel(nn.Module):
def __init__(self, mm_state_dict: dict[str, Tensor], mm_info: AnimateDiffInfo):
super().__init__()
self.mm_info = mm_info
self.down_blocks: Iterable[MotionModule] = nn.ModuleList([])
self.up_blocks: Iterable[MotionModule] = nn.ModuleList([])
self.mid_block: Union[MotionModule, None] = None
self.encoding_max_len = get_position_encoding_max_len(mm_state_dict, mm_info.mm_name, mm_info.mm_format)
self.has_position_encoding = self.encoding_max_len is not None
# determine ops to use (to support fp8 properly)
if comfy.model_management.unet_manual_cast(comfy.model_management.unet_dtype(), comfy.model_management.get_torch_device()) is None:
ops = comfy.ops.disable_weight_init
else:
ops = comfy.ops.manual_cast
# SDXL has 3 up/down blocks, SD1.5 has 4 up/down blocks
if mm_info.sd_type == ModelTypeSD.SDXL:
layer_channels = (320, 640, 1280)
else:
layer_channels = (320, 640, 1280, 1280)
# fill out down/up blocks and middle block, if present
for c in layer_channels:
self.down_blocks.append(MotionModule(c, temporal_position_encoding=self.has_position_encoding,
temporal_position_encoding_max_len=self.encoding_max_len, block_type=BlockType.DOWN, ops=ops))
for c in reversed(layer_channels):
self.up_blocks.append(MotionModule(c, temporal_position_encoding=self.has_position_encoding,
temporal_position_encoding_max_len=self.encoding_max_len, block_type=BlockType.UP, ops=ops))
if has_mid_block(mm_state_dict):
self.mid_block = MotionModule(1280, temporal_position_encoding=self.has_position_encoding,
temporal_position_encoding_max_len=self.encoding_max_len, block_type=BlockType.MID, ops=ops)
self.AD_video_length: int = 24
def get_device_debug(self):
return self.down_blocks[0].motion_modules[0].temporal_transformer.proj_in.weight.device
def is_length_valid_for_encoding_max_len(self, length: int):
if self.encoding_max_len is None:
return True
return length <= self.encoding_max_len
def get_best_beta_schedule(self, log=False) -> str:
to_return = None
if self.mm_info.sd_type == ModelTypeSD.SD1_5:
if self.mm_info.mm_format == AnimateDiffFormat.ANIMATELCM:
to_return = BetaSchedules.LCM # while LCM_100 is the intended schedule, I find LCM to have much less flicker
else:
to_return = BetaSchedules.SQRT_LINEAR
elif self.mm_info.sd_type == ModelTypeSD.SDXL:
if self.mm_info.mm_format == AnimateDiffFormat.HOTSHOTXL:
to_return = BetaSchedules.LINEAR
else:
to_return = BetaSchedules.LINEAR_ADXL
if to_return is not None:
if log: logger.info(f"[Autoselect]: '{to_return}' beta_schedule for {self.mm_info.get_string()}")
else:
to_return = BetaSchedules.USE_EXISTING
if log: logger.info(f"[Autoselect]: could not find beta_schedule for {self.mm_info.get_string()}, defaulting to '{to_return}'")
return to_return
def cleanup(self):
pass
def inject(self, model: ModelPatcher):
unet: openaimodel.UNetModel = model.model.diffusion_model
# inject input (down) blocks
# SD15 mm contains 4 downblocks, each with 2 TemporalTransformers - 8 in total
# SDXL mm contains 3 downblocks, each with 2 TemporalTransformers - 6 in total
self._inject(unet.input_blocks, self.down_blocks)
# inject output (up) blocks
# SD15 mm contains 4 upblocks, each with 3 TemporalTransformers - 12 in total
# SDXL mm contains 3 upblocks, each with 3 TemporalTransformers - 9 in total
self._inject(unet.output_blocks, self.up_blocks)
# inject mid block, if needed (encapsulate in list to make structure compatible)
if self.mid_block is not None:
self._inject([unet.middle_block], [self.mid_block])
del unet
def _inject(self, unet_blocks: nn.ModuleList, mm_blocks: nn.ModuleList):
# Rules for injection:
# For each component list in a unet block:
# if SpatialTransformer exists in list, place next block after last occurrence
# elif ResBlock exists in list, place next block after first occurrence
# else don't place block
injection_count = 0
unet_idx = 0
# details about blocks passed in
per_block = len(mm_blocks[0].motion_modules)
injection_goal = len(mm_blocks) * per_block
# only stop injecting when modules exhausted
while injection_count < injection_goal:
# figure out which VanillaTemporalModule from mm to inject
mm_blk_idx, mm_vtm_idx = injection_count // per_block, injection_count % per_block
# figure out layout of unet block components
st_idx = -1 # SpatialTransformer index
res_idx = -1 # first ResBlock index
# first, figure out indeces of relevant blocks
for idx, component in enumerate(unet_blocks[unet_idx]):
if type(component) == SpatialTransformer:
st_idx = idx
elif type(component).__name__ == "ResBlock" and res_idx < 0:
res_idx = idx
# if SpatialTransformer exists, inject right after
if st_idx >= 0:
#logger.info(f"AD: injecting after ST({st_idx})")
unet_blocks[unet_idx].insert(st_idx+1, mm_blocks[mm_blk_idx].motion_modules[mm_vtm_idx])
injection_count += 1
# otherwise, if only ResBlock exists, inject right after
elif res_idx >= 0:
#logger.info(f"AD: injecting after Res({res_idx})")
unet_blocks[unet_idx].insert(res_idx+1, mm_blocks[mm_blk_idx].motion_modules[mm_vtm_idx])
injection_count += 1
# increment unet_idx
unet_idx += 1
def eject(self, model: ModelPatcher):
unet: openaimodel.UNetModel = model.model.diffusion_model
# remove from input blocks (downblocks)
self._eject(unet.input_blocks)
# remove from output blocks (upblocks)
self._eject(unet.output_blocks)
# remove from middle block (encapsulate in list to make compatible)
self._eject([unet.middle_block])
del unet
def _eject(self, unet_blocks: nn.ModuleList):
# eject all VanillaTemporalModule objects from all blocks
for block in unet_blocks:
idx_to_pop = []
for idx, component in enumerate(block):
if type(component) == VanillaTemporalModule:
idx_to_pop.append(idx)
# pop in backwards order, as to not disturb what the indeces refer to
for idx in sorted(idx_to_pop, reverse=True):
block.pop(idx)
def set_video_length(self, video_length: int, full_length: int):
self.AD_video_length = video_length
for block in self.down_blocks:
block.set_video_length(video_length, full_length)
for block in self.up_blocks:
block.set_video_length(video_length, full_length)
if self.mid_block is not None:
self.mid_block.set_video_length(video_length, full_length)
def set_scale(self, multival: Union[float, Tensor]):
if multival is None:
multival = 1.0
if type(multival) == Tensor:
self._set_scale_multiplier(1.0)
self._set_scale_mask(multival)
else:
self._set_scale_multiplier(multival)
self._set_scale_mask(None)
def set_effect(self, multival: Union[float, Tensor]):
for block in self.down_blocks:
block.set_effect(multival)
for block in self.up_blocks:
block.set_effect(multival)
if self.mid_block is not None:
self.mid_block.set_effect(multival)
def set_sub_idxs(self, sub_idxs: list[int]):
for block in self.down_blocks:
block.set_sub_idxs(sub_idxs)
for block in self.up_blocks:
block.set_sub_idxs(sub_idxs)
if self.mid_block is not None:
self.mid_block.set_sub_idxs(sub_idxs)
def set_view_options(self, view_options: ContextOptions):
for block in self.down_blocks:
block.set_view_options(view_options)
for block in self.up_blocks:
block.set_view_options(view_options)
if self.mid_block is not None:
self.mid_block.set_view_options(view_options)
def reset(self):
self._reset_sub_idxs()
self._reset_scale_multiplier()
self._reset_temp_vars()
def _set_scale_multiplier(self, multiplier: Union[float, None]):
for block in self.down_blocks:
block.set_scale_multiplier(multiplier)
for block in self.up_blocks:
block.set_scale_multiplier(multiplier)
if self.mid_block is not None:
self.mid_block.set_scale_multiplier(multiplier)
def _set_scale_mask(self, mask: Tensor):
for block in self.down_blocks:
block.set_scale_mask(mask)
for block in self.up_blocks:
block.set_scale_mask(mask)
if self.mid_block is not None:
self.mid_block.set_scale_mask(mask)
def _reset_temp_vars(self):
for block in self.down_blocks:
block.reset_temp_vars()
for block in self.up_blocks:
block.reset_temp_vars()
if self.mid_block is not None:
self.mid_block.reset_temp_vars()
def _reset_scale_multiplier(self):
self._set_scale_multiplier(None)
def _reset_sub_idxs(self):
self.set_sub_idxs(None)
class MotionModule(nn.Module):
def __init__(self,
in_channels,
temporal_position_encoding=True,
temporal_position_encoding_max_len=24,
block_type: str=BlockType.DOWN,
ops=comfy.ops.disable_weight_init
):
super().__init__()
if block_type == BlockType.MID:
# mid blocks contain only a single VanillaTemporalModule
self.motion_modules: Iterable[VanillaTemporalModule] = nn.ModuleList([get_motion_module(in_channels, temporal_position_encoding, temporal_position_encoding_max_len, ops=ops)])
else:
# down blocks contain two VanillaTemporalModules
self.motion_modules: Iterable[VanillaTemporalModule] = nn.ModuleList(
[
get_motion_module(in_channels, temporal_position_encoding, temporal_position_encoding_max_len, ops=ops),
get_motion_module(in_channels, temporal_position_encoding, temporal_position_encoding_max_len, ops=ops)
]
)
# up blocks contain one additional VanillaTemporalModule
if block_type == BlockType.UP:
self.motion_modules.append(get_motion_module(in_channels, temporal_position_encoding, temporal_position_encoding_max_len, ops=ops))
def set_video_length(self, video_length: int, full_length: int):
for motion_module in self.motion_modules:
motion_module.set_video_length(video_length, full_length)
def set_scale_multiplier(self, multiplier: Union[float, None]):
for motion_module in self.motion_modules:
motion_module.set_scale_multiplier(multiplier)
def set_scale_mask(self, mask: Tensor):
for motion_module in self.motion_modules:
motion_module.set_scale_mask(mask)
def set_effect(self, multival: Union[float, Tensor]):
for motion_module in self.motion_modules:
motion_module.set_effect(multival)
def set_sub_idxs(self, sub_idxs: list[int]):
for motion_module in self.motion_modules:
motion_module.set_sub_idxs(sub_idxs)
def set_view_options(self, view_options: ContextOptions):
for motion_module in self.motion_modules:
motion_module.set_view_options(view_options=view_options)
def reset_temp_vars(self):
for motion_module in self.motion_modules:
motion_module.reset_temp_vars()
def get_motion_module(in_channels, temporal_position_encoding, temporal_position_encoding_max_len, ops=comfy.ops.disable_weight_init):
return VanillaTemporalModule(in_channels=in_channels, temporal_position_encoding=temporal_position_encoding, temporal_position_encoding_max_len=temporal_position_encoding_max_len, ops=ops)
class VanillaTemporalModule(nn.Module):
def __init__(
self,
in_channels,
num_attention_heads=8,
num_transformer_block=1,
attention_block_types=("Temporal_Self", "Temporal_Self"),
cross_frame_attention_mode=None,
temporal_position_encoding=True,
temporal_position_encoding_max_len=24,
temporal_attention_dim_div=1,
zero_initialize=True,
ops=comfy.ops.disable_weight_init,
):
super().__init__()
self.video_length = 16
self.full_length = 16
self.sub_idxs = None
self.view_options = None
self.effect = None
self.temp_effect_mask: Tensor = None
self.prev_input_tensor_batch = 0
self.temporal_transformer = TemporalTransformer3DModel(
in_channels=in_channels,
num_attention_heads=num_attention_heads,
attention_head_dim=in_channels
// num_attention_heads
// temporal_attention_dim_div,
num_layers=num_transformer_block,
attention_block_types=attention_block_types,
cross_frame_attention_mode=cross_frame_attention_mode,
temporal_position_encoding=temporal_position_encoding,
temporal_position_encoding_max_len=temporal_position_encoding_max_len,
ops=ops
)
if zero_initialize:
self.temporal_transformer.proj_out = zero_module(
self.temporal_transformer.proj_out
)
def set_video_length(self, video_length: int, full_length: int):
self.video_length = video_length
self.full_length = full_length
self.temporal_transformer.set_video_length(video_length, full_length)
def set_scale_multiplier(self, multiplier: Union[float, None]):
self.temporal_transformer.set_scale_multiplier(multiplier)
def set_scale_mask(self, mask: Tensor):
self.temporal_transformer.set_scale_mask(mask)
def set_effect(self, multival: Union[float, Tensor]):
if type(multival) == Tensor:
self.effect = multival
elif multival is not None and math.isclose(multival, 1.0):
self.effect = None
else:
self.effect = multival
self.temp_effect_mask = None
def set_sub_idxs(self, sub_idxs: list[int]):
self.sub_idxs = sub_idxs
self.temporal_transformer.set_sub_idxs(sub_idxs)
def set_view_options(self, view_options: ContextOptions):
self.view_options = view_options
def reset_temp_vars(self):
self.set_effect(None)
self.set_view_options(None)
self.temporal_transformer.reset_temp_vars()
def get_effect_mask(self, input_tensor: Tensor):
batch, channel, height, width = input_tensor.shape
batched_number = batch // self.video_length
full_batched_idxs = list(range(self.video_length))*batched_number
# if there is a cached temp_effect_mask and it is valid for current input, return it
if batch == self.prev_input_tensor_batch and self.temp_effect_mask is not None:
if self.sub_idxs is not None:
return self.temp_effect_mask[self.sub_idxs*batched_number]
return self.temp_effect_mask[full_batched_idxs]
# clear any existing mask
del self.temp_effect_mask
self.temp_effect_mask = None
# recalculate temp mask
self.prev_input_tensor_batch = batch
# make sure mask matches expected dimensions
mask = prepare_mask_batch(self.effect, shape=(self.full_length, 1, height, width))
# make sure mask is as long as full_length - clone last element of list if too short
self.temp_effect_mask = extend_to_batch_size(mask, self.full_length).to(
dtype=input_tensor.dtype, device=input_tensor.device)
# return finalized mask
if self.sub_idxs is not None:
return self.temp_effect_mask[self.sub_idxs*batched_number]
return self.temp_effect_mask[full_batched_idxs]
def forward(self, input_tensor: Tensor, encoder_hidden_states=None, attention_mask=None):
if self.effect is None:
return self.temporal_transformer(input_tensor, encoder_hidden_states, attention_mask, self.view_options)
# return weighted average of input_tensor and AD output
if type(self.effect) != Tensor:
effect = self.effect
# do nothing if effect is 0
if math.isclose(effect, 0.0):
return input_tensor
else:
effect = self.get_effect_mask(input_tensor)
return input_tensor*(1.0-effect) + self.temporal_transformer(input_tensor, encoder_hidden_states, attention_mask, self.view_options)*effect
class TemporalTransformer3DModel(nn.Module):
def __init__(
self,
in_channels,
num_attention_heads,
attention_head_dim,
num_layers,
attention_block_types=(
"Temporal_Self",
"Temporal_Self",
),
dropout=0.0,
norm_num_groups=32,
cross_attention_dim=768,
activation_fn="geglu",
attention_bias=False,
upcast_attention=False,
cross_frame_attention_mode=None,
temporal_position_encoding=False,
temporal_position_encoding_max_len=24,
ops=comfy.ops.disable_weight_init,
):
super().__init__()
self.video_length = 16
self.full_length = 16
self.raw_scale_mask: Union[Tensor, None] = None
self.temp_scale_mask: Union[Tensor, None] = None
self.sub_idxs: Union[list[int], None] = None
self.prev_hidden_states_batch = 0
inner_dim = num_attention_heads * attention_head_dim
self.norm = ops.GroupNorm(
num_groups=norm_num_groups, num_channels=in_channels, eps=1e-6, affine=True
)
self.proj_in = ops.Linear(in_channels, inner_dim)
self.transformer_blocks: Iterable[TemporalTransformerBlock] = nn.ModuleList(
[
TemporalTransformerBlock(
dim=inner_dim,
num_attention_heads=num_attention_heads,
attention_head_dim=attention_head_dim,
attention_block_types=attention_block_types,
dropout=dropout,
norm_num_groups=norm_num_groups,
cross_attention_dim=cross_attention_dim,
activation_fn=activation_fn,
attention_bias=attention_bias,
upcast_attention=upcast_attention,
cross_frame_attention_mode=cross_frame_attention_mode,
temporal_position_encoding=temporal_position_encoding,
temporal_position_encoding_max_len=temporal_position_encoding_max_len,
ops=ops,
)
for d in range(num_layers)
]
)
self.proj_out = ops.Linear(inner_dim, in_channels)
def set_video_length(self, video_length: int, full_length: int):
self.video_length = video_length
self.full_length = full_length
def set_scale_multiplier(self, multiplier: Union[float, None]):
for block in self.transformer_blocks:
block.set_scale_multiplier(multiplier)
def set_scale_mask(self, mask: Tensor):
self.raw_scale_mask = mask
self.temp_scale_mask = None
def set_sub_idxs(self, sub_idxs: list[int]):
self.sub_idxs = sub_idxs
for block in self.transformer_blocks:
block.set_sub_idxs(sub_idxs)
def reset_temp_vars(self):
del self.temp_scale_mask
self.temp_scale_mask = None
self.prev_hidden_states_batch = 0
def get_scale_mask(self, hidden_states: Tensor) -> Union[Tensor, None]:
# if no raw mask, return None
if self.raw_scale_mask is None:
return None
shape = hidden_states.shape
batch, channel, height, width = shape
# if temp mask already calculated, return it
if self.temp_scale_mask != None:
# check if hidden_states batch matches
if batch == self.prev_hidden_states_batch:
if self.sub_idxs is not None:
return self.temp_scale_mask[:, self.sub_idxs, :]
return self.temp_scale_mask
# if does not match, reset cached temp_scale_mask and recalculate it
del self.temp_scale_mask
self.temp_scale_mask = None
# otherwise, calculate temp mask
self.prev_hidden_states_batch = batch
mask = prepare_mask_batch(self.raw_scale_mask, shape=(self.full_length, 1, height, width))
mask = repeat_to_batch_size(mask, self.full_length)
# if mask not the same amount length as full length, make it match
if self.full_length != mask.shape[0]:
mask = broadcast_image_to(mask, self.full_length, 1)
# reshape mask to attention K shape (h*w, latent_count, 1)
batch, channel, height, width = mask.shape
# first, perform same operations as on hidden_states,
# turning (b, c, h, w) -> (b, h*w, c)
mask = mask.permute(0, 2, 3, 1).reshape(batch, height*width, channel)
# then, make it the same shape as attention's k, (h*w, b, c)
mask = mask.permute(1, 0, 2)
# make masks match the expected length of h*w
batched_number = shape[0] // self.video_length
if batched_number > 1:
mask = torch.cat([mask] * batched_number, dim=0)
# cache mask and set to proper device
self.temp_scale_mask = mask
# move temp_scale_mask to proper dtype + device
self.temp_scale_mask = self.temp_scale_mask.to(dtype=hidden_states.dtype, device=hidden_states.device)
# return subset of masks, if needed
if self.sub_idxs is not None:
return self.temp_scale_mask[:, self.sub_idxs, :]
return self.temp_scale_mask
def forward(self, hidden_states, encoder_hidden_states=None, attention_mask=None, view_options: ContextOptions=None):
batch, channel, height, width = hidden_states.shape
residual = hidden_states
scale_mask = self.get_scale_mask(hidden_states)
# add some casts for fp8 purposes - does not affect speed otherwise
hidden_states = self.norm(hidden_states).to(hidden_states.dtype)
inner_dim = hidden_states.shape[1]
hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(
batch, height * width, inner_dim
)
hidden_states = self.proj_in(hidden_states).to(hidden_states.dtype)
# Transformer Blocks
for block in self.transformer_blocks:
hidden_states = block(
hidden_states,
encoder_hidden_states=encoder_hidden_states,
attention_mask=attention_mask,
video_length=self.video_length,
scale_mask=scale_mask,
view_options=view_options
)
# output
hidden_states = self.proj_out(hidden_states)
hidden_states = (
hidden_states.reshape(batch, height, width, inner_dim)
.permute(0, 3, 1, 2)
.contiguous()
)
output = hidden_states + residual
return output
class TemporalTransformerBlock(nn.Module):
def __init__(
self,
dim,
num_attention_heads,
attention_head_dim,
attention_block_types=(
"Temporal_Self",
"Temporal_Self",
),
dropout=0.0,
norm_num_groups=32,
cross_attention_dim=768,
activation_fn="geglu",
attention_bias=False,
upcast_attention=False,
cross_frame_attention_mode=None,
temporal_position_encoding=False,
temporal_position_encoding_max_len=24,
ops=comfy.ops.disable_weight_init,
):
super().__init__()
attention_blocks = []
norms = []
for block_name in attention_block_types:
attention_blocks.append(
VersatileAttention(
attention_mode=block_name.split("_")[0],
context_dim=cross_attention_dim # called context_dim for ComfyUI impl
if block_name.endswith("_Cross")
else None,
query_dim=dim,
heads=num_attention_heads,
dim_head=attention_head_dim,
dropout=dropout,
#bias=attention_bias, # remove for Comfy CrossAttention
#upcast_attention=upcast_attention, # remove for Comfy CrossAttention
cross_frame_attention_mode=cross_frame_attention_mode,
temporal_position_encoding=temporal_position_encoding,
temporal_position_encoding_max_len=temporal_position_encoding_max_len,
ops=ops,
)
)
norms.append(ops.LayerNorm(dim))
self.attention_blocks: Iterable[VersatileAttention] = nn.ModuleList(attention_blocks)
self.norms = nn.ModuleList(norms)
self.ff = FeedForward(dim, dropout=dropout, glu=(activation_fn == "geglu"), operations=ops)
self.ff_norm = ops.LayerNorm(dim)
def set_scale_multiplier(self, multiplier: Union[float, None]):
for block in self.attention_blocks:
block.set_scale_multiplier(multiplier)
def set_sub_idxs(self, sub_idxs: list[int]):
for block in self.attention_blocks:
block.set_sub_idxs(sub_idxs)
def forward(
self,
hidden_states: Tensor,
encoder_hidden_states: Tensor=None,
attention_mask: Tensor=None,
video_length: int=None,
scale_mask: Tensor=None,
view_options: ContextOptions=None,
):
# make view_options None if context_length > video_length, or if equal and equal not allowed
if view_options:
if view_options.context_length > video_length:
view_options = None
elif view_options.context_length == video_length and not view_options.use_on_equal_length:
view_options = None
if not view_options:
for attention_block, norm in zip(self.attention_blocks, self.norms):
norm_hidden_states = norm(hidden_states).to(hidden_states.dtype)
hidden_states = (
attention_block(
norm_hidden_states,
encoder_hidden_states=encoder_hidden_states
if attention_block.is_cross_attention
else None,
attention_mask=attention_mask,
video_length=video_length,
scale_mask=scale_mask
) + hidden_states
)
else:
# views idea gotten from diffusers AnimateDiff FreeNoise implementation:
# https://github.com/arthur-qiu/FreeNoise-AnimateDiff/blob/main/animatediff/models/motion_module.py
# apply sliding context windows (views)
views = get_context_windows(num_frames=video_length, opts=view_options)
hidden_states = rearrange(hidden_states, "(b f) d c -> b f d c", f=video_length)
value_final = torch.zeros_like(hidden_states)
count_final = torch.zeros_like(hidden_states)
# bias_final = [0.0] * video_length
batched_conds = hidden_states.size(1) // video_length
for sub_idxs in views:
sub_hidden_states = rearrange(hidden_states[:, sub_idxs], "b f d c -> (b f) d c")
for attention_block, norm in zip(self.attention_blocks, self.norms):
norm_hidden_states = norm(sub_hidden_states).to(sub_hidden_states.dtype)
sub_hidden_states = (
attention_block(
norm_hidden_states,
encoder_hidden_states=encoder_hidden_states # do these need to be changed for sub_idxs too?
if attention_block.is_cross_attention
else None,
attention_mask=attention_mask,
video_length=len(sub_idxs),
scale_mask=scale_mask[:, sub_idxs, :] if scale_mask is not None else scale_mask
) + sub_hidden_states
)
sub_hidden_states = rearrange(sub_hidden_states, "(b f) d c -> b f d c", f=len(sub_idxs))
# if view_options.fuse_method == ContextFuseMethod.RELATIVE:
# for pos, idx in enumerate(sub_idxs):
# # bias is the influence of a specific index in relation to the whole context window
# bias = 1 - abs(idx - (sub_idxs[0] + sub_idxs[-1]) / 2) / ((sub_idxs[-1] - sub_idxs[0] + 1e-2) / 2)
# bias = max(1e-2, bias)
# # take weighted averate relative to total bias of current idx
# bias_total = bias_final[idx]
# prev_weight = torch.tensor([bias_total / (bias_total + bias)],
# dtype=value_final.dtype, device=value_final.device).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
# #prev_weight = torch.cat([prev_weight]*value_final.shape[1], dim=1)
# new_weight = torch.tensor([bias / (bias_total + bias)],
# dtype=value_final.dtype, device=value_final.device).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
# #new_weight = torch.cat([new_weight]*value_final.shape[1], dim=1)
# test = value_final[:, idx:idx+1, :, :]
# value_final[:, idx:idx+1, :, :] = value_final[:, idx:idx+1, :, :] * prev_weight + sub_hidden_states[:, pos:pos+1, : ,:] * new_weight
# bias_final[idx] = bias_total + bias
# else:
weights = get_context_weights(len(sub_idxs), view_options.fuse_method) * batched_conds
weights_tensor = torch.Tensor(weights).to(device=hidden_states.device).unsqueeze(0).unsqueeze(-1).unsqueeze(-1)
value_final[:, sub_idxs] += sub_hidden_states * weights_tensor
count_final[:, sub_idxs] += weights_tensor
# get weighted average of sub_hidden_states, if fuse method requires it
# if view_options.fuse_method != ContextFuseMethod.RELATIVE:
hidden_states = value_final / count_final
hidden_states = rearrange(hidden_states, "b f d c -> (b f) d c")
del value_final
del count_final
# del bias_final
hidden_states = self.ff(self.ff_norm(hidden_states)) + hidden_states
output = hidden_states
return output
class PositionalEncoding(nn.Module):
def __init__(self, d_model, dropout=0.0, max_len=24):
super().__init__()
self.dropout = nn.Dropout(p=dropout)
position = torch.arange(max_len).unsqueeze(1)
div_term = torch.exp(
torch.arange(0, d_model, 2) * (-math.log(10000.0) / d_model)
)
pe = torch.zeros(1, max_len, d_model)
pe[0, :, 0::2] = torch.sin(position * div_term)
pe[0, :, 1::2] = torch.cos(position * div_term)
self.register_buffer("pe", pe)
self.sub_idxs = None
def set_sub_idxs(self, sub_idxs: list[int]):
self.sub_idxs = sub_idxs
def forward(self, x):
#if self.sub_idxs is not None:
# x = x + self.pe[:, self.sub_idxs]
#else:
x = x + self.pe[:, : x.size(1)]
return self.dropout(x)
class VersatileAttention(CrossAttentionMM):
def __init__(
self,
attention_mode=None,
cross_frame_attention_mode=None,
temporal_position_encoding=False,
temporal_position_encoding_max_len=24,
ops=comfy.ops.disable_weight_init,
*args,
**kwargs,
):
super().__init__(operations=ops, *args, **kwargs)
assert attention_mode == "Temporal"
self.attention_mode = attention_mode
self.is_cross_attention = kwargs["context_dim"] is not None
self.pos_encoder = (
PositionalEncoding(
kwargs["query_dim"],
dropout=0.0,
max_len=temporal_position_encoding_max_len,
)
if (temporal_position_encoding and attention_mode == "Temporal")
else None
)
def extra_repr(self):
return f"(Module Info) Attention_Mode: {self.attention_mode}, Is_Cross_Attention: {self.is_cross_attention}"
def set_scale_multiplier(self, multiplier: Union[float, None]):
if multiplier is None or math.isclose(multiplier, 1.0):
self.scale = 1.0
else:
self.scale = multiplier
def set_sub_idxs(self, sub_idxs: list[int]):
if self.pos_encoder != None:
self.pos_encoder.set_sub_idxs(sub_idxs)
def forward(
self,
hidden_states: Tensor,
encoder_hidden_states=None,
attention_mask=None,
video_length=None,
scale_mask=None,
):
if self.attention_mode != "Temporal":
raise NotImplementedError
d = hidden_states.shape[1]
hidden_states = rearrange(
hidden_states, "(b f) d c -> (b d) f c", f=video_length
)
if self.pos_encoder is not None:
hidden_states = self.pos_encoder(hidden_states).to(hidden_states.dtype)
encoder_hidden_states = (
repeat(encoder_hidden_states, "b n c -> (b d) n c", d=d)
if encoder_hidden_states is not None
else encoder_hidden_states
)
hidden_states = super().forward(
hidden_states,
encoder_hidden_states,
value=None,
mask=attention_mask,
scale_mask=scale_mask,
)
hidden_states = rearrange(hidden_states, "(b d) f c -> (b f) d c", d=d)
return hidden_states