Spaces:
Runtime error
Runtime error
File size: 7,966 Bytes
c19ca42 |
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 |
from enum import Enum
from typing import Any, List, NamedTuple
from functools import lru_cache
class UnetBlockType(Enum):
INPUT = "input"
OUTPUT = "output"
MIDDLE = "middle"
class TransformerID(NamedTuple):
block_type: UnetBlockType
# The id of the block the transformer is in. Not all blocks have cross attn.
block_id: int
# The index of transformer within the block.
# A block can have multiple transformers in SDXL.
block_index: int
# The call index of transformer if in a single step of diffusion.
transformer_index: int
class TransformerIDResult(NamedTuple):
input_ids: List[TransformerID]
output_ids: List[TransformerID]
middle_ids: List[TransformerID]
def get(self, idx: int) -> TransformerID:
return self.to_list()[idx]
def to_list(self) -> List[TransformerID]:
return sorted(
self.input_ids + self.output_ids + self.middle_ids,
key=lambda i: i.transformer_index,
)
class StableDiffusionVersion(Enum):
"""The version family of stable diffusion model."""
UNKNOWN = 0
SD1x = 1
SD2x = 2
SDXL = 3
@staticmethod
def detect_from_model_name(model_name: str) -> "StableDiffusionVersion":
"""Based on the model name provided, guess what stable diffusion version it is.
This might not be accurate without actually inspect the file content.
"""
if any(f"sd{v}" in model_name.lower() for v in ("14", "15", "16")):
return StableDiffusionVersion.SD1x
if "sd21" in model_name or "2.1" in model_name:
return StableDiffusionVersion.SD2x
if "xl" in model_name.lower():
return StableDiffusionVersion.SDXL
return StableDiffusionVersion.UNKNOWN
def encoder_block_num(self) -> int:
if self in (
StableDiffusionVersion.SD1x,
StableDiffusionVersion.SD2x,
StableDiffusionVersion.UNKNOWN,
):
return 12
else:
return 9 # SDXL
def controlnet_layer_num(self) -> int:
return self.encoder_block_num() + 1
@property
def transformer_block_num(self) -> int:
"""Number of blocks that has cross attn transformers in unet."""
if self in (
StableDiffusionVersion.SD1x,
StableDiffusionVersion.SD2x,
StableDiffusionVersion.UNKNOWN,
):
return 16
else:
return 11 # SDXL
@property
@lru_cache(maxsize=None)
def transformer_ids(self) -> List[TransformerID]:
"""id of blocks that have cross attention"""
if self in (
StableDiffusionVersion.SD1x,
StableDiffusionVersion.SD2x,
StableDiffusionVersion.UNKNOWN,
):
transformer_index = 0
input_ids = []
for block_id in [1, 2, 4, 5, 7, 8]:
input_ids.append(
TransformerID(UnetBlockType.INPUT, block_id, 0, transformer_index)
)
transformer_index += 1
middle_id = TransformerID(UnetBlockType.MIDDLE, 0, 0, transformer_index)
transformer_index += 1
output_ids = []
for block_id in [3, 4, 5, 6, 7, 8, 9, 10, 11]:
input_ids.append(
TransformerID(UnetBlockType.OUTPUT, block_id, 0, transformer_index)
)
transformer_index += 1
return TransformerIDResult(input_ids, output_ids, [middle_id])
else:
# SDXL
transformer_index = 0
input_ids = []
for block_id in [4, 5, 7, 8]:
block_indices = (
range(2) if block_id in [4, 5] else range(10)
) # transformer_depth
for index in block_indices:
input_ids.append(
TransformerID(
UnetBlockType.INPUT, block_id, index, transformer_index
)
)
transformer_index += 1
middle_ids = [
TransformerID(UnetBlockType.MIDDLE, 0, index, transformer_index)
for index in range(10)
]
transformer_index += 1
output_ids = []
for block_id in range(6):
block_indices = (
range(2) if block_id in [3, 4, 5] else range(10)
) # transformer_depth
for index in block_indices:
output_ids.append(
TransformerID(
UnetBlockType.OUTPUT, block_id, index, transformer_index
)
)
transformer_index += 1
return TransformerIDResult(input_ids, output_ids, middle_ids)
def is_compatible_with(self, other: "StableDiffusionVersion") -> bool:
"""Incompatible only when one of version is SDXL and other is not."""
return (
any(v == StableDiffusionVersion.UNKNOWN for v in [self, other])
or sum(v == StableDiffusionVersion.SDXL for v in [self, other]) != 1
)
class ControlModelType(Enum):
"""
The type of Control Models (supported or not).
"""
ControlNet = "ControlNet, Lvmin Zhang"
T2I_Adapter = "T2I_Adapter, Chong Mou"
T2I_StyleAdapter = "T2I_StyleAdapter, Chong Mou"
T2I_CoAdapter = "T2I_CoAdapter, Chong Mou"
MasaCtrl = "MasaCtrl, Mingdeng Cao"
GLIGEN = "GLIGEN, Yuheng Li"
AttentionInjection = "AttentionInjection, Lvmin Zhang" # A simple attention injection written by Lvmin
StableSR = "StableSR, Jianyi Wang"
PromptDiffusion = "PromptDiffusion, Zhendong Wang"
ControlLoRA = "ControlLoRA, Wu Hecong"
ReVision = "ReVision, Stability"
IPAdapter = "IPAdapter, Hu Ye"
Controlllite = "Controlllite, Kohya"
InstantID = "InstantID, Qixun Wang"
SparseCtrl = "SparseCtrl, Yuwei Guo"
def is_controlnet(self) -> bool:
"""Returns whether the control model should be treated as ControlNet."""
return self in (
ControlModelType.ControlNet,
ControlModelType.ControlLoRA,
ControlModelType.InstantID,
)
def allow_context_sharing(self) -> bool:
"""Returns whether this control model type allows the same PlugableControlModel
object map to multiple ControlNetUnit.
Both IPAdapter and Controlllite have unit specific input (clip/image) stored
on the model object during inference. Sharing the context means that the input
set earlier gets lost.
"""
return self not in (
ControlModelType.IPAdapter,
ControlModelType.Controlllite,
)
# Written by Lvmin
class AutoMachine(Enum):
"""
Lvmin's algorithm for Attention/AdaIn AutoMachine States.
"""
Read = "Read"
Write = "Write"
StyleAlign = "StyleAlign"
class HiResFixOption(Enum):
BOTH = "Both"
LOW_RES_ONLY = "Low res only"
HIGH_RES_ONLY = "High res only"
@staticmethod
def from_value(value: Any) -> "HiResFixOption":
if isinstance(value, str) and value.startswith("HiResFixOption."):
_, field = value.split(".")
return getattr(HiResFixOption, field)
if isinstance(value, str):
return HiResFixOption(value)
elif isinstance(value, int):
return [x for x in HiResFixOption][value]
else:
assert isinstance(value, HiResFixOption)
return value
class InputMode(Enum):
# Single image to a single ControlNet unit.
SIMPLE = "simple"
# Input is a directory. N generations. Each generation takes 1 input image
# from the directory.
BATCH = "batch"
# Input is a directory. 1 generation. Each generation takes N input image
# from the directory.
MERGE = "merge"
|