751 lines
32 KiB
Python
751 lines
32 KiB
Python
import torch
|
|
import contextlib
|
|
import os
|
|
import math
|
|
|
|
import comfy.utils
|
|
import comfy.model_management
|
|
from comfy.clip_vision import clip_preprocess
|
|
from comfy.ldm.modules.attention import optimized_attention
|
|
import folder_paths
|
|
|
|
from torch import nn
|
|
from PIL import Image
|
|
import torch.nn.functional as F
|
|
import torchvision.transforms as TT
|
|
|
|
# set the models directory backward compatible
|
|
GLOBAL_MODELS_DIR = os.path.join(folder_paths.models_dir, "ipadapter")
|
|
MODELS_DIR = GLOBAL_MODELS_DIR if os.path.isdir(GLOBAL_MODELS_DIR) else os.path.join(os.path.dirname(os.path.realpath(__file__)), "models")
|
|
if "ipadapter" not in folder_paths.folder_names_and_paths:
|
|
folder_paths.folder_names_and_paths["ipadapter"] = ([MODELS_DIR], folder_paths.supported_pt_extensions)
|
|
else:
|
|
folder_paths.folder_names_and_paths["ipadapter"][1].update(folder_paths.supported_pt_extensions)
|
|
|
|
class MLPProjModelImport(torch.nn.Module):
|
|
"""SD model with image prompt"""
|
|
def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024):
|
|
super().__init__()
|
|
|
|
self.proj = torch.nn.Sequential(
|
|
torch.nn.Linear(clip_embeddings_dim, clip_embeddings_dim),
|
|
torch.nn.GELU(),
|
|
torch.nn.Linear(clip_embeddings_dim, cross_attention_dim),
|
|
torch.nn.LayerNorm(cross_attention_dim)
|
|
)
|
|
|
|
def forward(self, image_embeds):
|
|
clip_extra_context_tokens = self.proj(image_embeds)
|
|
return clip_extra_context_tokens
|
|
|
|
class ImageProjModelImport(nn.Module):
|
|
def __init__(self, cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4):
|
|
super().__init__()
|
|
|
|
self.cross_attention_dim = cross_attention_dim
|
|
self.clip_extra_context_tokens = clip_extra_context_tokens
|
|
self.proj = nn.Linear(clip_embeddings_dim, self.clip_extra_context_tokens * cross_attention_dim)
|
|
self.norm = nn.LayerNorm(cross_attention_dim)
|
|
|
|
def forward(self, image_embeds):
|
|
embeds = image_embeds
|
|
clip_extra_context_tokens = self.proj(embeds).reshape(-1, self.clip_extra_context_tokens, self.cross_attention_dim)
|
|
clip_extra_context_tokens = self.norm(clip_extra_context_tokens)
|
|
return clip_extra_context_tokens
|
|
|
|
class To_KVImport(nn.Module):
|
|
def __init__(self, state_dict):
|
|
super().__init__()
|
|
|
|
self.to_kvs = nn.ModuleDict()
|
|
for key, value in state_dict.items():
|
|
self.to_kvs[key.replace(".weight", "").replace(".", "_")] = nn.Linear(value.shape[1], value.shape[0], bias=False)
|
|
self.to_kvs[key.replace(".weight", "").replace(".", "_")].weight.data = value
|
|
|
|
def FeedForward(dim, mult=4):
|
|
inner_dim = int(dim * mult)
|
|
return nn.Sequential(
|
|
nn.LayerNorm(dim),
|
|
nn.Linear(dim, inner_dim, bias=False),
|
|
nn.GELU(),
|
|
nn.Linear(inner_dim, dim, bias=False),
|
|
)
|
|
|
|
|
|
class PerceiverAttention(nn.Module):
|
|
def __init__(self, *, dim, dim_head=64, heads=8):
|
|
super().__init__()
|
|
self.scale = dim_head**-0.5
|
|
self.dim_head = dim_head
|
|
self.heads = heads
|
|
inner_dim = dim_head * heads
|
|
|
|
self.norm1 = nn.LayerNorm(dim)
|
|
self.norm2 = nn.LayerNorm(dim)
|
|
|
|
self.to_q = nn.Linear(dim, inner_dim, bias=False)
|
|
self.to_kv = nn.Linear(dim, inner_dim * 2, bias=False)
|
|
self.to_out = nn.Linear(inner_dim, dim, bias=False)
|
|
|
|
|
|
def forward(self, x, latents):
|
|
"""
|
|
Args:
|
|
x (torch.Tensor): image features
|
|
shape (b, n1, D)
|
|
latent (torch.Tensor): latent features
|
|
shape (b, n2, D)
|
|
"""
|
|
x = self.norm1(x)
|
|
latents = self.norm2(latents)
|
|
|
|
b, l, _ = latents.shape
|
|
|
|
q = self.to_q(latents)
|
|
kv_input = torch.cat((x, latents), dim=-2)
|
|
k, v = self.to_kv(kv_input).chunk(2, dim=-1)
|
|
|
|
q = reshape_tensor(q, self.heads)
|
|
k = reshape_tensor(k, self.heads)
|
|
v = reshape_tensor(v, self.heads)
|
|
|
|
# attention
|
|
scale = 1 / math.sqrt(math.sqrt(self.dim_head))
|
|
weight = (q * scale) @ (k * scale).transpose(-2, -1) # More stable with f16 than dividing afterwards
|
|
weight = torch.softmax(weight.float(), dim=-1).type(weight.dtype)
|
|
out = weight @ v
|
|
|
|
out = out.permute(0, 2, 1, 3).reshape(b, l, -1)
|
|
|
|
return self.to_out(out)
|
|
|
|
def reshape_tensor(x, heads):
|
|
bs, length, width = x.shape
|
|
#(bs, length, width) --> (bs, length, n_heads, dim_per_head)
|
|
x = x.view(bs, length, heads, -1)
|
|
# (bs, length, n_heads, dim_per_head) --> (bs, n_heads, length, dim_per_head)
|
|
x = x.transpose(1, 2)
|
|
# (bs, n_heads, length, dim_per_head) --> (bs*n_heads, length, dim_per_head)
|
|
x = x.reshape(bs, heads, length, -1)
|
|
return x
|
|
|
|
def set_model_patch_replace(model, patch_kwargs, key):
|
|
to = model.model_options["transformer_options"]
|
|
if "patches_replace" not in to:
|
|
to["patches_replace"] = {}
|
|
if "attn2" not in to["patches_replace"]:
|
|
to["patches_replace"]["attn2"] = {}
|
|
if key not in to["patches_replace"]["attn2"]:
|
|
patch = CrossAttentionPatchImport(**patch_kwargs)
|
|
to["patches_replace"]["attn2"][key] = patch
|
|
else:
|
|
to["patches_replace"]["attn2"][key].set_new_condition(**patch_kwargs)
|
|
|
|
def image_add_noise(image, noise):
|
|
image = image.permute([0,3,1,2])
|
|
torch.manual_seed(0) # use a fixed random for reproducible results
|
|
transforms = TT.Compose([
|
|
TT.CenterCrop(min(image.shape[2], image.shape[3])),
|
|
TT.Resize((224, 224), interpolation=TT.InterpolationMode.BICUBIC, antialias=True),
|
|
TT.ElasticTransform(alpha=75.0, sigma=noise*3.5), # shuffle the image
|
|
TT.RandomVerticalFlip(p=1.0), # flip the image to change the geometry even more
|
|
TT.RandomHorizontalFlip(p=1.0),
|
|
])
|
|
image = transforms(image.cpu())
|
|
image = image.permute([0,2,3,1])
|
|
image = image + ((0.25*(1-noise)+0.05) * torch.randn_like(image) ) # add further random noise
|
|
return image
|
|
|
|
def zeroed_hidden_states(clip_vision, batch_size):
|
|
image = torch.zeros([batch_size, 224, 224, 3])
|
|
comfy.model_management.load_model_gpu(clip_vision.patcher)
|
|
pixel_values = clip_preprocess(image.to(clip_vision.load_device))
|
|
|
|
if clip_vision.dtype != torch.float32:
|
|
precision_scope = torch.autocast
|
|
else:
|
|
precision_scope = lambda a, b: contextlib.nullcontext(a)
|
|
|
|
with precision_scope(comfy.model_management.get_autocast_device(clip_vision.load_device), torch.float32):
|
|
outputs = clip_vision.model(pixel_values, intermediate_output=-2)
|
|
|
|
# we only need the penultimate hidden states
|
|
outputs = outputs[1].to(comfy.model_management.intermediate_device())
|
|
|
|
return outputs
|
|
|
|
def min_(tensor_list):
|
|
# return the element-wise min of the tensor list.
|
|
x = torch.stack(tensor_list)
|
|
mn = x.min(axis=0)[0]
|
|
return torch.clamp(mn, min=0)
|
|
|
|
def max_(tensor_list):
|
|
# return the element-wise max of the tensor list.
|
|
x = torch.stack(tensor_list)
|
|
mx = x.max(axis=0)[0]
|
|
return torch.clamp(mx, max=1)
|
|
|
|
# From https://github.com/Jamy-L/Pytorch-Contrast-Adaptive-Sharpening/
|
|
def contrast_adaptive_sharpening(image, amount):
|
|
img = F.pad(image, pad=(1, 1, 1, 1)).cpu()
|
|
|
|
a = img[..., :-2, :-2]
|
|
b = img[..., :-2, 1:-1]
|
|
c = img[..., :-2, 2:]
|
|
d = img[..., 1:-1, :-2]
|
|
e = img[..., 1:-1, 1:-1]
|
|
f = img[..., 1:-1, 2:]
|
|
g = img[..., 2:, :-2]
|
|
h = img[..., 2:, 1:-1]
|
|
i = img[..., 2:, 2:]
|
|
|
|
# Computing contrast
|
|
cross = (b, d, e, f, h)
|
|
mn = min_(cross)
|
|
mx = max_(cross)
|
|
|
|
diag = (a, c, g, i)
|
|
mn2 = min_(diag)
|
|
mx2 = max_(diag)
|
|
mx = mx + mx2
|
|
mn = mn + mn2
|
|
|
|
# Computing local weight
|
|
inv_mx = torch.reciprocal(mx)
|
|
amp = inv_mx * torch.minimum(mn, (2 - mx))
|
|
|
|
# scaling
|
|
amp = torch.sqrt(amp)
|
|
w = - amp * (amount * (1/5 - 1/8) + 1/8)
|
|
div = torch.reciprocal(1 + 4*w)
|
|
|
|
output = ((b + d + f + h)*w + e) * div
|
|
output = output.clamp(0, 1)
|
|
output = torch.nan_to_num(output)
|
|
|
|
return (output)
|
|
|
|
class IPAdapterImport(nn.Module):
|
|
def __init__(self, ipadapter_model, cross_attention_dim=1024, output_cross_attention_dim=1024, clip_embeddings_dim=1024, clip_extra_context_tokens=4, is_sdxl=False, is_plus=False, is_full=False):
|
|
super().__init__()
|
|
|
|
self.clip_embeddings_dim = clip_embeddings_dim
|
|
self.cross_attention_dim = cross_attention_dim
|
|
self.output_cross_attention_dim = output_cross_attention_dim
|
|
self.clip_extra_context_tokens = clip_extra_context_tokens
|
|
self.is_sdxl = is_sdxl
|
|
self.is_full = is_full
|
|
|
|
self.image_proj_model = self.init_proj() if not is_plus else self.init_proj_plus()
|
|
self.image_proj_model.load_state_dict(ipadapter_model["image_proj"])
|
|
self.ip_layers = To_KVImport(ipadapter_model["ip_adapter"])
|
|
|
|
def init_proj(self):
|
|
image_proj_model = ImageProjModelImport(
|
|
cross_attention_dim=self.cross_attention_dim,
|
|
clip_embeddings_dim=self.clip_embeddings_dim,
|
|
clip_extra_context_tokens=self.clip_extra_context_tokens
|
|
)
|
|
return image_proj_model
|
|
|
|
def init_proj_plus(self):
|
|
if self.is_full:
|
|
image_proj_model = MLPProjModelImport(
|
|
cross_attention_dim=self.cross_attention_dim,
|
|
clip_embeddings_dim=self.clip_embeddings_dim
|
|
)
|
|
else:
|
|
image_proj_model = ResamplerImport(
|
|
dim=self.cross_attention_dim,
|
|
depth=4,
|
|
dim_head=64,
|
|
heads=20 if self.is_sdxl else 12,
|
|
num_queries=self.clip_extra_context_tokens,
|
|
embedding_dim=self.clip_embeddings_dim,
|
|
output_dim=self.output_cross_attention_dim,
|
|
ff_mult=4
|
|
)
|
|
return image_proj_model
|
|
|
|
@torch.inference_mode()
|
|
def get_image_embeds(self, clip_embed, clip_embed_zeroed):
|
|
image_prompt_embeds = self.image_proj_model(clip_embed)
|
|
uncond_image_prompt_embeds = self.image_proj_model(clip_embed_zeroed)
|
|
return image_prompt_embeds, uncond_image_prompt_embeds
|
|
|
|
class CrossAttentionPatchImport:
|
|
# forward for patching
|
|
def __init__(self, weight, ipadapter, device, dtype, number, cond, uncond, weight_type, mask=None, sigma_start=0.0, sigma_end=1.0, unfold_batch=False):
|
|
self.weights = [weight]
|
|
self.ipadapters = [ipadapter]
|
|
self.conds = [cond]
|
|
self.unconds = [uncond]
|
|
self.device = 'cuda' if 'cuda' in device.type else 'cpu'
|
|
self.dtype = dtype if 'cuda' in self.device else torch.bfloat16
|
|
self.number = number
|
|
self.weight_type = [weight_type]
|
|
self.masks = [mask]
|
|
self.sigma_start = [sigma_start]
|
|
self.sigma_end = [sigma_end]
|
|
self.unfold_batch = [unfold_batch]
|
|
|
|
self.k_key = str(self.number*2+1) + "_to_k_ip"
|
|
self.v_key = str(self.number*2+1) + "_to_v_ip"
|
|
|
|
def set_new_condition(self, weight, ipadapter, device, dtype, number, cond, uncond, weight_type, mask=None, sigma_start=0.0, sigma_end=1.0, unfold_batch=False):
|
|
self.weights.append(weight)
|
|
self.ipadapters.append(ipadapter)
|
|
self.conds.append(cond)
|
|
self.unconds.append(uncond)
|
|
self.masks.append(mask)
|
|
self.device = 'cuda' if 'cuda' in device.type else 'cpu'
|
|
self.dtype = dtype if 'cuda' in self.device else torch.bfloat16
|
|
self.weight_type.append(weight_type)
|
|
self.sigma_start.append(sigma_start)
|
|
self.sigma_end.append(sigma_end)
|
|
self.unfold_batch.append(unfold_batch)
|
|
|
|
def __call__(self, n, context_attn2, value_attn2, extra_options):
|
|
org_dtype = n.dtype
|
|
cond_or_uncond = extra_options["cond_or_uncond"]
|
|
sigma = extra_options["sigmas"][0].item() if 'sigmas' in extra_options else 999999999.9
|
|
|
|
# extra options for AnimateDiff
|
|
ad_params = extra_options['ad_params'] if "ad_params" in extra_options else None
|
|
|
|
with torch.autocast(device_type=self.device, dtype=self.dtype):
|
|
q = n
|
|
k = context_attn2
|
|
v = value_attn2
|
|
b = q.shape[0]
|
|
qs = q.shape[1]
|
|
batch_prompt = b // len(cond_or_uncond)
|
|
out = optimized_attention(q, k, v, extra_options["n_heads"])
|
|
_, _, lh, lw = extra_options["original_shape"]
|
|
|
|
for weight, cond, uncond, ipadapter, mask, weight_type, sigma_start, sigma_end, unfold_batch in zip(self.weights, self.conds, self.unconds, self.ipadapters, self.masks, self.weight_type, self.sigma_start, self.sigma_end, self.unfold_batch):
|
|
if sigma > sigma_start or sigma < sigma_end:
|
|
continue
|
|
|
|
if unfold_batch and cond.shape[0] > 1:
|
|
# Check AnimateDiff context window
|
|
if ad_params is not None and ad_params["sub_idxs"] is not None:
|
|
# if images length matches or exceeds full_length get sub_idx images
|
|
if cond.shape[0] >= ad_params["full_length"]:
|
|
cond = torch.Tensor(cond[ad_params["sub_idxs"]])
|
|
uncond = torch.Tensor(uncond[ad_params["sub_idxs"]])
|
|
# otherwise, need to do more to get proper sub_idxs masks
|
|
else:
|
|
# check if images length matches full_length - if not, make it match
|
|
if cond.shape[0] < ad_params["full_length"]:
|
|
cond = torch.cat((cond, cond[-1:].repeat((ad_params["full_length"]-cond.shape[0], 1, 1))), dim=0)
|
|
uncond = torch.cat((uncond, uncond[-1:].repeat((ad_params["full_length"]-uncond.shape[0], 1, 1))), dim=0)
|
|
# if we have too many remove the excess (should not happen, but just in case)
|
|
if cond.shape[0] > ad_params["full_length"]:
|
|
cond = cond[:ad_params["full_length"]]
|
|
uncond = uncond[:ad_params["full_length"]]
|
|
cond = cond[ad_params["sub_idxs"]]
|
|
uncond = uncond[ad_params["sub_idxs"]]
|
|
|
|
# if we don't have enough reference images repeat the last one until we reach the right size
|
|
if cond.shape[0] < batch_prompt:
|
|
cond = torch.cat((cond, cond[-1:].repeat((batch_prompt-cond.shape[0], 1, 1))), dim=0)
|
|
uncond = torch.cat((uncond, uncond[-1:].repeat((batch_prompt-uncond.shape[0], 1, 1))), dim=0)
|
|
# if we have too many remove the exceeding
|
|
elif cond.shape[0] > batch_prompt:
|
|
cond = cond[:batch_prompt]
|
|
uncond = uncond[:batch_prompt]
|
|
|
|
k_cond = ipadapter.ip_layers.to_kvs[self.k_key](cond)
|
|
k_uncond = ipadapter.ip_layers.to_kvs[self.k_key](uncond)
|
|
v_cond = ipadapter.ip_layers.to_kvs[self.v_key](cond)
|
|
v_uncond = ipadapter.ip_layers.to_kvs[self.v_key](uncond)
|
|
else:
|
|
k_cond = ipadapter.ip_layers.to_kvs[self.k_key](cond).repeat(batch_prompt, 1, 1)
|
|
k_uncond = ipadapter.ip_layers.to_kvs[self.k_key](uncond).repeat(batch_prompt, 1, 1)
|
|
v_cond = ipadapter.ip_layers.to_kvs[self.v_key](cond).repeat(batch_prompt, 1, 1)
|
|
v_uncond = ipadapter.ip_layers.to_kvs[self.v_key](uncond).repeat(batch_prompt, 1, 1)
|
|
|
|
if weight_type.startswith("linear"):
|
|
ip_k = torch.cat([(k_cond, k_uncond)[i] for i in cond_or_uncond], dim=0) * weight
|
|
ip_v = torch.cat([(v_cond, v_uncond)[i] for i in cond_or_uncond], dim=0) * weight
|
|
else:
|
|
ip_k = torch.cat([(k_cond, k_uncond)[i] for i in cond_or_uncond], dim=0)
|
|
ip_v = torch.cat([(v_cond, v_uncond)[i] for i in cond_or_uncond], dim=0)
|
|
|
|
if weight_type.startswith("channel"):
|
|
# code by Lvmin Zhang at Stanford University as also seen on Fooocus IPAdapter implementation
|
|
# please read licensing notes https://github.com/lllyasviel/Fooocus/blob/main/fooocus_extras/ip_adapter.py#L225
|
|
ip_v_mean = torch.mean(ip_v, dim=1, keepdim=True)
|
|
ip_v_offset = ip_v - ip_v_mean
|
|
_, _, C = ip_k.shape
|
|
channel_penalty = float(C) / 1280.0
|
|
W = weight * channel_penalty
|
|
ip_k = ip_k * W
|
|
ip_v = ip_v_offset + ip_v_mean * W
|
|
|
|
out_ip = optimized_attention(q, ip_k, ip_v, extra_options["n_heads"])
|
|
if weight_type.startswith("original"):
|
|
out_ip = out_ip * weight
|
|
|
|
if mask is not None:
|
|
# TODO: needs checking
|
|
mask_h = max(1, round(lh / math.sqrt(lh * lw / qs)))
|
|
mask_w = qs // mask_h
|
|
|
|
# check if using AnimateDiff and sliding context window
|
|
if (mask.shape[0] > 1 and ad_params is not None and ad_params["sub_idxs"] is not None):
|
|
# if mask length matches or exceeds full_length, just get sub_idx masks, resize, and continue
|
|
if mask.shape[0] >= ad_params["full_length"]:
|
|
mask_downsample = torch.Tensor(mask[ad_params["sub_idxs"]])
|
|
mask_downsample = F.interpolate(mask_downsample.unsqueeze(1), size=(mask_h, mask_w), mode="bicubic").squeeze(1)
|
|
# otherwise, need to do more to get proper sub_idxs masks
|
|
else:
|
|
# resize to needed attention size (to save on memory)
|
|
mask_downsample = F.interpolate(mask.unsqueeze(1), size=(mask_h, mask_w), mode="bicubic").squeeze(1)
|
|
# check if mask length matches full_length - if not, make it match
|
|
if mask_downsample.shape[0] < ad_params["full_length"]:
|
|
mask_downsample = torch.cat((mask_downsample, mask_downsample[-1:].repeat((ad_params["full_length"]-mask_downsample.shape[0], 1, 1))), dim=0)
|
|
# if we have too many remove the excess (should not happen, but just in case)
|
|
if mask_downsample.shape[0] > ad_params["full_length"]:
|
|
mask_downsample = mask_downsample[:ad_params["full_length"]]
|
|
# now, select sub_idxs masks
|
|
mask_downsample = mask_downsample[ad_params["sub_idxs"]]
|
|
# otherwise, perform usual mask interpolation
|
|
else:
|
|
mask_downsample = F.interpolate(mask.unsqueeze(1), size=(mask_h, mask_w), mode="bicubic").squeeze(1)
|
|
|
|
# if we don't have enough masks repeat the last one until we reach the right size
|
|
if mask_downsample.shape[0] < batch_prompt:
|
|
mask_downsample = torch.cat((mask_downsample, mask_downsample[-1:, :, :].repeat((batch_prompt-mask_downsample.shape[0], 1, 1))), dim=0)
|
|
# if we have too many remove the exceeding
|
|
elif mask_downsample.shape[0] > batch_prompt:
|
|
mask_downsample = mask_downsample[:batch_prompt, :, :]
|
|
|
|
# repeat the masks
|
|
mask_downsample = mask_downsample.repeat(len(cond_or_uncond), 1, 1)
|
|
mask_downsample = mask_downsample.view(mask_downsample.shape[0], -1, 1).repeat(1, 1, out.shape[2])
|
|
|
|
out_ip = out_ip * mask_downsample
|
|
|
|
out = out + out_ip
|
|
|
|
return out.to(dtype=org_dtype)
|
|
|
|
|
|
|
|
class IPAdapterApplyImport:
|
|
@classmethod
|
|
def INPUT_TYPES(s):
|
|
return {
|
|
"required": {
|
|
"ipadapter": ("IPADAPTER", ),
|
|
"clip_vision": ("CLIP_VISION",),
|
|
"image": ("IMAGE",),
|
|
"model": ("MODEL", ),
|
|
"weight": ("FLOAT", { "default": 1.0, "min": -1, "max": 3, "step": 0.05 }),
|
|
"noise": ("FLOAT", { "default": 0.0, "min": 0.0, "max": 1.0, "step": 0.01 }),
|
|
"weight_type": (["original", "linear", "channel penalty"], ),
|
|
"start_at": ("FLOAT", { "default": 0.0, "min": 0.0, "max": 1.0, "step": 0.001 }),
|
|
"end_at": ("FLOAT", { "default": 1.0, "min": 0.0, "max": 1.0, "step": 0.001 }),
|
|
"unfold_batch": ("BOOLEAN", { "default": False }),
|
|
},
|
|
"optional": {
|
|
"attn_mask": ("MASK",),
|
|
}
|
|
}
|
|
|
|
RETURN_TYPES = ("MODEL",)
|
|
FUNCTION = "apply_ipadapter"
|
|
CATEGORY = "ipadapter"
|
|
|
|
def apply_ipadapter(self, ipadapter, model, weight, clip_vision=None, image=None, weight_type="original", noise=None, embeds=None, attn_mask=None, start_at=0.0, end_at=1.0, unfold_batch=False):
|
|
self.dtype = model.model.diffusion_model.dtype
|
|
self.device = comfy.model_management.get_torch_device()
|
|
self.weight = weight
|
|
self.is_full = "proj.0.weight" in ipadapter["image_proj"]
|
|
self.is_plus = self.is_full or "latents" in ipadapter["image_proj"]
|
|
|
|
output_cross_attention_dim = ipadapter["ip_adapter"]["1.to_k_ip.weight"].shape[1]
|
|
self.is_sdxl = output_cross_attention_dim == 2048
|
|
cross_attention_dim = 1280 if self.is_plus and self.is_sdxl else output_cross_attention_dim
|
|
clip_extra_context_tokens = 16 if self.is_plus else 4
|
|
|
|
if embeds is not None:
|
|
embeds = torch.unbind(embeds)
|
|
clip_embed = embeds[0].cpu()
|
|
clip_embed_zeroed = embeds[1].cpu()
|
|
else:
|
|
if image.shape[1] != image.shape[2]:
|
|
print("\033[33mINFO: the IPAdapter reference image is not a square, CLIPImageProcessor will resize and crop it at the center. If the main focus of the picture is not in the middle the result might not be what you are expecting.\033[0m")
|
|
|
|
clip_embed = clip_vision.encode_image(image)
|
|
neg_image = image_add_noise(image, noise) if noise > 0 else None
|
|
|
|
if self.is_plus:
|
|
clip_embed = clip_embed.penultimate_hidden_states
|
|
if noise > 0:
|
|
clip_embed_zeroed = clip_vision.encode_image(neg_image).penultimate_hidden_states
|
|
else:
|
|
clip_embed_zeroed = zeroed_hidden_states(clip_vision, image.shape[0])
|
|
else:
|
|
clip_embed = clip_embed.image_embeds
|
|
if noise > 0:
|
|
clip_embed_zeroed = clip_vision.encode_image(neg_image).image_embeds
|
|
else:
|
|
clip_embed_zeroed = torch.zeros_like(clip_embed)
|
|
|
|
clip_embeddings_dim = clip_embed.shape[-1]
|
|
|
|
self.ipadapter = IPAdapterImport(
|
|
ipadapter,
|
|
cross_attention_dim=cross_attention_dim,
|
|
output_cross_attention_dim=output_cross_attention_dim,
|
|
clip_embeddings_dim=clip_embeddings_dim,
|
|
clip_extra_context_tokens=clip_extra_context_tokens,
|
|
is_sdxl=self.is_sdxl,
|
|
is_plus=self.is_plus,
|
|
is_full=self.is_full,
|
|
)
|
|
|
|
self.ipadapter.to(self.device, dtype=self.dtype)
|
|
|
|
image_prompt_embeds, uncond_image_prompt_embeds = self.ipadapter.get_image_embeds(clip_embed.to(self.device, self.dtype), clip_embed_zeroed.to(self.device, self.dtype))
|
|
image_prompt_embeds = image_prompt_embeds.to(self.device, dtype=self.dtype)
|
|
uncond_image_prompt_embeds = uncond_image_prompt_embeds.to(self.device, dtype=self.dtype)
|
|
|
|
work_model = model.clone()
|
|
|
|
if attn_mask is not None:
|
|
attn_mask = attn_mask.to(self.device)
|
|
|
|
sigma_start = model.model.model_sampling.percent_to_sigma(start_at)
|
|
sigma_end = model.model.model_sampling.percent_to_sigma(end_at)
|
|
|
|
patch_kwargs = {
|
|
"number": 0,
|
|
"weight": self.weight,
|
|
"ipadapter": self.ipadapter,
|
|
"device": self.device,
|
|
"dtype": self.dtype,
|
|
"cond": image_prompt_embeds,
|
|
"uncond": uncond_image_prompt_embeds,
|
|
"weight_type": weight_type,
|
|
"mask": attn_mask,
|
|
"sigma_start": sigma_start,
|
|
"sigma_end": sigma_end,
|
|
"unfold_batch": unfold_batch,
|
|
}
|
|
|
|
if not self.is_sdxl:
|
|
for id in [1,2,4,5,7,8]: # id of input_blocks that have cross attention
|
|
set_model_patch_replace(work_model, patch_kwargs, ("input", id))
|
|
patch_kwargs["number"] += 1
|
|
for id in [3,4,5,6,7,8,9,10,11]: # id of output_blocks that have cross attention
|
|
set_model_patch_replace(work_model, patch_kwargs, ("output", id))
|
|
patch_kwargs["number"] += 1
|
|
set_model_patch_replace(work_model, patch_kwargs, ("middle", 0))
|
|
else:
|
|
for id in [4,5,7,8]: # id of input_blocks that have cross attention
|
|
block_indices = range(2) if id in [4, 5] else range(10) # transformer_depth
|
|
for index in block_indices:
|
|
set_model_patch_replace(work_model, patch_kwargs, ("input", id, index))
|
|
patch_kwargs["number"] += 1
|
|
for id in range(6): # id of output_blocks that have cross attention
|
|
block_indices = range(2) if id in [3, 4, 5] else range(10) # transformer_depth
|
|
for index in block_indices:
|
|
set_model_patch_replace(work_model, patch_kwargs, ("output", id, index))
|
|
patch_kwargs["number"] += 1
|
|
for index in range(10):
|
|
set_model_patch_replace(work_model, patch_kwargs, ("middle", 0, index))
|
|
patch_kwargs["number"] += 1
|
|
|
|
return (work_model, )
|
|
|
|
def prep_image(image, interpolation="LANCZOS", crop_position="center", sharpening=0.0):
|
|
_, oh, ow, _ = image.shape
|
|
output = image.permute([0,3,1,2])
|
|
|
|
if "pad" in crop_position:
|
|
target_length = max(oh, ow)
|
|
pad_l = (target_length - ow) // 2
|
|
pad_r = (target_length - ow) - pad_l
|
|
pad_t = (target_length - oh) // 2
|
|
pad_b = (target_length - oh) - pad_t
|
|
output = F.pad(output, (pad_l, pad_r, pad_t, pad_b), value=0, mode="constant")
|
|
else:
|
|
crop_size = min(oh, ow)
|
|
x = (ow-crop_size) // 2
|
|
y = (oh-crop_size) // 2
|
|
if "top" in crop_position:
|
|
y = 0
|
|
elif "bottom" in crop_position:
|
|
y = oh-crop_size
|
|
elif "left" in crop_position:
|
|
x = 0
|
|
elif "right" in crop_position:
|
|
x = ow-crop_size
|
|
|
|
x2 = x+crop_size
|
|
y2 = y+crop_size
|
|
|
|
# crop
|
|
output = output[:, :, y:y2, x:x2]
|
|
|
|
# resize (apparently PIL resize is better than tourchvision interpolate)
|
|
imgs = []
|
|
for i in range(output.shape[0]):
|
|
img = TT.ToPILImage()(output[i])
|
|
img = img.resize((224,224), resample=Image.Resampling[interpolation])
|
|
imgs.append(TT.ToTensor()(img))
|
|
output = torch.stack(imgs, dim=0)
|
|
|
|
if sharpening > 0:
|
|
output = contrast_adaptive_sharpening(output, sharpening)
|
|
|
|
output = output.permute([0,2,3,1])
|
|
|
|
return (output,)
|
|
|
|
class ResamplerImport(nn.Module):
|
|
def __init__(
|
|
self,
|
|
dim=1024,
|
|
depth=8,
|
|
dim_head=64,
|
|
heads=16,
|
|
num_queries=8,
|
|
embedding_dim=768,
|
|
output_dim=1024,
|
|
ff_mult=4,
|
|
):
|
|
super().__init__()
|
|
|
|
self.latents = nn.Parameter(torch.randn(1, num_queries, dim) / dim**0.5)
|
|
|
|
self.proj_in = nn.Linear(embedding_dim, dim)
|
|
|
|
self.proj_out = nn.Linear(dim, output_dim)
|
|
self.norm_out = nn.LayerNorm(output_dim)
|
|
|
|
self.layers = nn.ModuleList([])
|
|
for _ in range(depth):
|
|
self.layers.append(
|
|
nn.ModuleList(
|
|
[
|
|
PerceiverAttention(dim=dim, dim_head=dim_head, heads=heads),
|
|
FeedForward(dim=dim, mult=ff_mult),
|
|
]
|
|
)
|
|
)
|
|
|
|
def forward(self, x):
|
|
|
|
latents = self.latents.repeat(x.size(0), 1, 1)
|
|
|
|
x = self.proj_in(x)
|
|
|
|
for attn, ff in self.layers:
|
|
latents = attn(x, latents) + latents
|
|
latents = ff(latents) + latents
|
|
|
|
latents = self.proj_out(latents)
|
|
return self.norm_out(latents)
|
|
|
|
|
|
class IPAdapterEncoderImport:
|
|
@classmethod
|
|
def INPUT_TYPES(s):
|
|
return {"required": {
|
|
"clip_vision": ("CLIP_VISION",),
|
|
"image_1": ("IMAGE",),
|
|
"ipadapter_plus": ("BOOLEAN", { "default": False }),
|
|
"noise": ("FLOAT", { "default": 0.0, "min": 0.0, "max": 1.0, "step": 0.01 }),
|
|
"weight_1": ("FLOAT", { "default": 1.0, "min": 0, "max": 1.0, "step": 0.01 }),
|
|
},
|
|
"optional": {
|
|
"image_2": ("IMAGE",),
|
|
"image_3": ("IMAGE",),
|
|
"image_4": ("IMAGE",),
|
|
"weight_2": ("FLOAT", { "default": 1.0, "min": 0, "max": 1.0, "step": 0.01 }),
|
|
"weight_3": ("FLOAT", { "default": 1.0, "min": 0, "max": 1.0, "step": 0.01 }),
|
|
"weight_4": ("FLOAT", { "default": 1.0, "min": 0, "max": 1.0, "step": 0.01 }),
|
|
}
|
|
}
|
|
|
|
RETURN_TYPES = ("EMBEDS",)
|
|
FUNCTION = "preprocess"
|
|
CATEGORY = "ipadapter"
|
|
|
|
def preprocess(self, clip_vision, image_1, ipadapter_plus, noise, weight_1, image_2=None, image_3=None, image_4=None, weight_2=1.0, weight_3=1.0, weight_4=1.0):
|
|
weight_1 *= (0.1 + (weight_1 - 0.1))
|
|
weight_1 = 1.19e-05 if weight_1 <= 1.19e-05 else weight_1
|
|
weight_2 *= (0.1 + (weight_2 - 0.1))
|
|
weight_2 = 1.19e-05 if weight_2 <= 1.19e-05 else weight_2
|
|
weight_3 *= (0.1 + (weight_3 - 0.1))
|
|
weight_3 = 1.19e-05 if weight_3 <= 1.19e-05 else weight_3
|
|
weight_4 *= (0.1 + (weight_4 - 0.1))
|
|
weight_5 = 1.19e-05 if weight_4 <= 1.19e-05 else weight_4
|
|
|
|
image = image_1
|
|
weight = [weight_1]*image_1.shape[0]
|
|
|
|
if image_2 is not None:
|
|
if image_1.shape[1:] != image_2.shape[1:]:
|
|
image_2 = comfy.utils.common_upscale(image_2.movedim(-1,1), image.shape[2], image.shape[1], "bilinear", "center").movedim(1,-1)
|
|
image = torch.cat((image, image_2), dim=0)
|
|
weight += [weight_2]*image_2.shape[0]
|
|
if image_3 is not None:
|
|
if image.shape[1:] != image_3.shape[1:]:
|
|
image_3 = comfy.utils.common_upscale(image_3.movedim(-1,1), image.shape[2], image.shape[1], "bilinear", "center").movedim(1,-1)
|
|
image = torch.cat((image, image_3), dim=0)
|
|
weight += [weight_3]*image_3.shape[0]
|
|
if image_4 is not None:
|
|
if image.shape[1:] != image_4.shape[1:]:
|
|
image_4 = comfy.utils.common_upscale(image_4.movedim(-1,1), image.shape[2], image.shape[1], "bilinear", "center").movedim(1,-1)
|
|
image = torch.cat((image, image_4), dim=0)
|
|
weight += [weight_4]*image_4.shape[0]
|
|
|
|
clip_embed = clip_vision.encode_image(image)
|
|
neg_image = image_add_noise(image, noise) if noise > 0 else None
|
|
|
|
if ipadapter_plus:
|
|
clip_embed = clip_embed.penultimate_hidden_states
|
|
if noise > 0:
|
|
clip_embed_zeroed = clip_vision.encode_image(neg_image).penultimate_hidden_states
|
|
else:
|
|
clip_embed_zeroed = zeroed_hidden_states(clip_vision, image.shape[0])
|
|
else:
|
|
clip_embed = clip_embed.image_embeds
|
|
if noise > 0:
|
|
clip_embed_zeroed = clip_vision.encode_image(neg_image).image_embeds
|
|
else:
|
|
clip_embed_zeroed = torch.zeros_like(clip_embed)
|
|
|
|
if any(e != 1.0 for e in weight):
|
|
weight = torch.tensor(weight).unsqueeze(-1) if not ipadapter_plus else torch.tensor(weight).unsqueeze(-1).unsqueeze(-1)
|
|
clip_embed = clip_embed * weight
|
|
|
|
output = torch.stack((clip_embed, clip_embed_zeroed))
|
|
|
|
return( output, )
|
|
|
|
|
|
|
|
|
|
class IPAdapterBatchEmbedsImport:
|
|
@classmethod
|
|
def INPUT_TYPES(s):
|
|
return {"required": {
|
|
"embed1": ("EMBEDS",),
|
|
"embed2": ("EMBEDS",),
|
|
}}
|
|
|
|
RETURN_TYPES = ("EMBEDS",)
|
|
FUNCTION = "batch"
|
|
CATEGORY = "ipadapter"
|
|
|
|
def batch(self, embed1, embed2):
|
|
output = torch.cat((embed1, embed2), dim=1)
|
|
return (output, ) |