log(td)
+4
-1
@@ -3,6 +3,7 @@ from .nodes.load_instance_scaleu_node import LoadInstanceScaleUNode
|
||||
from .nodes.load_instance_fusers_node import LoadInstanceFusersNode
|
||||
from .nodes.load_instance_positionnet_node import LoadInstancePositionNetNode
|
||||
from .nodes.instance_diffusion_tracking_prompt_node import InstanceDiffusionTrackingPromptNode
|
||||
from .nodes.download_and_load_models import DownloadInstanceDiffusionModels
|
||||
|
||||
|
||||
NODE_CLASS_MAPPINGS = {
|
||||
@@ -11,6 +12,7 @@ NODE_CLASS_MAPPINGS = {
|
||||
"LoadInstancePositionNetModel": LoadInstancePositionNetNode,
|
||||
"LoadInstanceFusersNode": LoadInstanceFusersNode,
|
||||
"InstanceDiffusionTrackingPrompt": InstanceDiffusionTrackingPromptNode,
|
||||
"DownloadInstanceDiffusionModels": DownloadInstanceDiffusionModels
|
||||
}
|
||||
|
||||
NODE_DISPLAY_NAME_MAPPINGS = {
|
||||
@@ -18,5 +20,6 @@ NODE_DISPLAY_NAME_MAPPINGS = {
|
||||
"LoadInstancePositionNetModel": "Load Instance PositionNet Model",
|
||||
"LoadInstanceScaleUModel": "Load Instance ScaleU Model",
|
||||
"LoadInstanceFusersNode": "Load Instance Fusers Model",
|
||||
"InstanceDiffusionTrackingPrompt": "Instance Diffusion Tracking Prompt"
|
||||
"InstanceDiffusionTrackingPrompt": "Instance Diffusion Tracking Prompt",
|
||||
"DownloadInstanceDiffusionModels": "(Down)Load Instance Diffusion Models"
|
||||
}
|
||||
|
||||
@@ -1,7 +1,13 @@
|
||||
import torch
|
||||
|
||||
from ..modules.attention import GatedSelfAttentionDense
|
||||
#from ..modules.attention import GatedSelfAttentionDense
|
||||
from comfy.gligen import GatedSelfAttentionDense as GSAD
|
||||
|
||||
class GatedSelfAttentionDense(GSAD):
|
||||
def forward(self, x, instance_options={}):
|
||||
objs = instance_options['objs']
|
||||
x = super().forward(x, objs)
|
||||
return x.to(torch.float16)
|
||||
|
||||
def prepare_fusers(fusers_ckpt) -> list[torch.nn.Module]:
|
||||
fusers_list = []
|
||||
|
||||
@@ -22,5 +22,5 @@ def get_positionnet_default_params():
|
||||
|
||||
def prepare_positionnet(checkpoint, params) -> torch.nn.Module:
|
||||
model = UniFusion(**params)
|
||||
model.load_state_dict(checkpoint)
|
||||
model.load_state_dict(checkpoint, strict=False)
|
||||
return model
|
||||
|
||||
@@ -1,184 +0,0 @@
|
||||
import torch
|
||||
import numpy as np
|
||||
from torch import nn
|
||||
from inspect import isfunction
|
||||
import torch.nn.functional as F
|
||||
|
||||
from comfy.gligen import GatedSelfAttentionDense as GSAD
|
||||
|
||||
|
||||
def exists(val):
|
||||
return val is not None
|
||||
|
||||
|
||||
def default(val, d):
|
||||
if exists(val):
|
||||
return val
|
||||
return d() if isfunction(d) else d
|
||||
|
||||
|
||||
# feedforward
|
||||
class GEGLU(nn.Module):
|
||||
def __init__(self, dim_in, dim_out):
|
||||
super().__init__()
|
||||
self.proj = nn.Linear(dim_in, dim_out * 2)
|
||||
|
||||
def forward(self, x):
|
||||
x, gate = self.proj(x).chunk(2, dim=-1)
|
||||
return x * F.gelu(gate)
|
||||
|
||||
|
||||
class FeedForward(nn.Module):
|
||||
def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0.):
|
||||
super().__init__()
|
||||
inner_dim = int(dim * mult)
|
||||
dim_out = default(dim_out, dim)
|
||||
project_in = nn.Sequential(
|
||||
nn.Linear(dim, inner_dim),
|
||||
nn.GELU()
|
||||
) if not glu else GEGLU(dim, inner_dim)
|
||||
|
||||
self.net = nn.Sequential(
|
||||
project_in,
|
||||
nn.Dropout(dropout),
|
||||
nn.Linear(inner_dim, dim_out)
|
||||
)
|
||||
|
||||
def forward(self, x):
|
||||
return self.net(x)
|
||||
|
||||
|
||||
class SelfAttentionFusers(nn.Module):
|
||||
def __init__(self, query_dim, heads=8, dim_head=64, dropout=0.):
|
||||
super().__init__()
|
||||
inner_dim = dim_head * heads
|
||||
self.scale = dim_head ** -0.5
|
||||
self.heads = heads
|
||||
|
||||
self.to_q = nn.Linear(query_dim, inner_dim, bias=False)
|
||||
self.to_k = nn.Linear(query_dim, inner_dim, bias=False)
|
||||
self.to_v = nn.Linear(query_dim, inner_dim, bias=False)
|
||||
|
||||
self.to_out = nn.Sequential(
|
||||
nn.Linear(inner_dim, query_dim), nn.Dropout(dropout))
|
||||
|
||||
def forward(self, x, instance_options={}):
|
||||
# TODO: use attention masking
|
||||
grounding_input = instance_options.get('grounding_input', None)
|
||||
drop_box_mask = instance_options.get('drop_box_mask', None)
|
||||
q = self.to_q(x) # B*N*(H*C)
|
||||
k = self.to_k(x) # B*N*(H*C)
|
||||
v = self.to_v(x) # B*N*(H*C)
|
||||
|
||||
B, N, HC = q.shape
|
||||
H = self.heads
|
||||
C = HC // H
|
||||
|
||||
q = q.view(B, N, H, C).permute(
|
||||
0, 2, 1, 3).reshape(B * H, N, C) # (B*H)*N*C
|
||||
k = k.view(B, N, H, C).permute(
|
||||
0, 2, 1, 3).reshape(B * H, N, C) # (B*H)*N*C
|
||||
v = v.view(B, N, H, C).permute(
|
||||
0, 2, 1, 3).reshape(B * H, N, C) # (B*H)*N*C
|
||||
|
||||
sim = torch.einsum('b i c, b j c -> b i j', q, k) * \
|
||||
self.scale # (B*H)*N*N
|
||||
|
||||
if grounding_input is not None:
|
||||
# att_masks: B*(n_objs*4)*sqrt(w_h)*sqrt(w_h)
|
||||
n_objs = grounding_input["att_masks"].shape[1]
|
||||
|
||||
att_masks_reshape = None
|
||||
# if N - n_objs * 4 == 64 * 64:
|
||||
# att_masks_reshape = grounding_input["att_masks"]
|
||||
# att_key = "att_masks_selfAtt64"
|
||||
|
||||
# if the mask is all zeros or box&mask are dropped, we do not need to mask the self-attention
|
||||
if att_masks_reshape is not None and torch.sum(att_masks_reshape) > 0.0 and not drop_box_mask:
|
||||
# the attention mask is shared across all blocks for more efficient forwarding
|
||||
# we only need to get the mask for the first block.
|
||||
if att_key in grounding_input:
|
||||
att_masks_ = grounding_input[att_key]
|
||||
else:
|
||||
w_h = att_masks_reshape.shape[2] * \
|
||||
att_masks_reshape.shape[3]
|
||||
att_masks_ = torch.ones(B, 1, N, N).type(
|
||||
att_masks_reshape.dtype).to(att_masks_reshape.device)
|
||||
# mask the self-attention between visual tokens
|
||||
# all object patches do not exchange information with each other
|
||||
# but object patches share information with all background regions
|
||||
att_masks_reshape_v = att_masks_reshape.view(
|
||||
B * n_objs, w_h, 1)
|
||||
|
||||
# get the self-attention between att_masks_reshape with batch matrix multiplication
|
||||
# 1. att_masks_reshape_v: (B*n_objs)*w_h*1
|
||||
# 2. att_masks_reshape_v.permute(0,1,3,2): (B*n_objs)*1*w_h
|
||||
# 3. torch.bmm(att_masks_reshape_v, att_masks_reshape_v.permute(0,2,1)): (B*n_objs)*w_h*w_h
|
||||
self_att_all = torch.bmm(
|
||||
att_masks_reshape_v, att_masks_reshape_v.permute(0, 2, 1))
|
||||
|
||||
# sum the values alone the n_objs dimension
|
||||
# 1. self_att_all: B*n_objs*w_h*w_h
|
||||
# 2. self_att_all.sum(dim=1): B*w_h*w_h
|
||||
self_att_ind_objs = self_att_all.view(
|
||||
B, n_objs, w_h, w_h).sum(dim=1)
|
||||
|
||||
# sum att_masks_reshape alone the n_objs dimension and measure the self-attention with batch matrix multiplication
|
||||
# 1. att_masks_reshape_all: B*n_objs*w_h*1
|
||||
# 2. att_masks_reshape_all.sum(dim=1): B*w_h*1
|
||||
# 3. set non-zero values to 1: B*w_h*1
|
||||
# 4. use batch matrix multiplication to measure the self-attention: B*w_h*w_h
|
||||
att_masks_reshape_all = att_masks_reshape_v.view(
|
||||
B, n_objs, w_h, 1).sum(dim=1)
|
||||
att_masks_reshape_all[att_masks_reshape_all >= 1.0] = 1.0
|
||||
self_att_all_objs = torch.bmm(
|
||||
att_masks_reshape_all, att_masks_reshape_all.permute(0, 2, 1))
|
||||
|
||||
# get the masks for avoiding information leakage between object patches
|
||||
visual_token_masks = self_att_all_objs + self_att_ind_objs
|
||||
|
||||
# avoid the communications between objects and background
|
||||
# objects-background can not communicate
|
||||
visual_token_masks[self_att_ind_objs < 1.0] = 0.0
|
||||
visual_token_masks[self_att_ind_objs >=
|
||||
1.0] = 1.0 # binay mask
|
||||
|
||||
att_masks_[:, :, :w_h, :w_h] = visual_token_masks.view(
|
||||
B, 1, w_h, w_h)
|
||||
|
||||
# mask the self-attention between grounded tokens and the visual tokens
|
||||
att_masks_reshape = att_masks_reshape.view(
|
||||
B, 1, n_objs, w_h)
|
||||
# the order of inputs are [box, point, scribble, mask]. only box and mask need to have masked self-attention
|
||||
att_masks_[:, :, w_h:, :w_h] = att_masks_reshape.repeat(
|
||||
1, 1, 4, 1)
|
||||
att_masks_[:, :, w_h + n_objs:w_h + n_objs * 3, :w_h] = 1
|
||||
att_masks_[:, :, :w_h, w_h:] = att_masks_reshape.permute(
|
||||
0, 1, 3, 2).repeat(1, 1, 1, 4)
|
||||
att_masks_[:, :, :w_h, w_h + n_objs:w_h + n_objs * 3] = 1
|
||||
|
||||
# add 1e-9 along the diagonal to avoid nan
|
||||
diagonal_epsilon = torch.eye(N).view(1, 1, N, N).type(
|
||||
att_masks_.dtype).to(att_masks_.device) * 1e-9
|
||||
att_masks_ = att_masks_ + diagonal_epsilon
|
||||
|
||||
# save the masks for later blocks
|
||||
grounding_input['next_att_mask'] = att_masks_
|
||||
# the larger the threshold is, the smaller area the grounded info can influence
|
||||
# if threshold = 1.01, all grounded info has no influence on the rest of the patches.
|
||||
sim = sim.view(B, H, N, N).masked_fill(
|
||||
att_masks_ <= 0.0, -np.inf).view(B * H, N, N) # -np.inf
|
||||
|
||||
attn = sim.softmax(dim=-1) # (B*H)*N*N
|
||||
|
||||
out = torch.einsum('b i j, b j c -> b i c', attn, v) # (B*H)*N*C
|
||||
out = out.view(B, H, N, C).permute(
|
||||
0, 2, 1, 3).reshape(B, N, (H * C)) # B*N*(H*C)
|
||||
return self.to_out(out)
|
||||
|
||||
|
||||
class GatedSelfAttentionDense(GSAD):
|
||||
def forward(self, x, instance_options={}):
|
||||
objs = instance_options['objs']
|
||||
x = super().forward(x, objs)
|
||||
return x.to(torch.float16)
|
||||
+48
-46
@@ -12,6 +12,8 @@ import torch.nn.functional as F
|
||||
from timm.models.layers import trunc_normal_, DropPath
|
||||
from timm.models.registry import register_model
|
||||
|
||||
import comfy.ops
|
||||
ops = comfy.ops.manual_cast
|
||||
|
||||
class Block(nn.Module):
|
||||
r""" ConvNeXt Block. There are two equivalent implementations:
|
||||
@@ -27,13 +29,13 @@ class Block(nn.Module):
|
||||
|
||||
def __init__(self, dim, drop_path=0., layer_scale_init_value=1e-6):
|
||||
super().__init__()
|
||||
self.dwconv = nn.Conv2d(dim, dim, kernel_size=7,
|
||||
self.dwconv = ops.Conv2d(dim, dim, kernel_size=7,
|
||||
padding=3, groups=dim) # depthwise conv
|
||||
self.norm = LayerNorm(dim, eps=1e-6)
|
||||
# pointwise/1x1 convs, implemented with linear layers
|
||||
self.pwconv1 = nn.Linear(dim, 4 * dim)
|
||||
self.pwconv1 = ops.Linear(dim, 4 * dim)
|
||||
self.act = nn.GELU()
|
||||
self.pwconv2 = nn.Linear(4 * dim, dim)
|
||||
self.pwconv2 = ops.Linear(4 * dim, dim)
|
||||
self.gamma = nn.Parameter(layer_scale_init_value * torch.ones((dim)),
|
||||
requires_grad=True) if layer_scale_init_value > 0 else None
|
||||
self.drop_path = DropPath(
|
||||
@@ -79,14 +81,14 @@ class ConvNeXt(nn.Module):
|
||||
# stem and 3 intermediate downsampling conv layers
|
||||
self.downsample_layers = nn.ModuleList()
|
||||
stem = nn.Sequential(
|
||||
nn.Conv2d(in_chans, dims[0], kernel_size=4, stride=4),
|
||||
ops.Conv2d(in_chans, dims[0], kernel_size=4, stride=4),
|
||||
LayerNorm(dims[0], eps=1e-6, data_format="channels_first")
|
||||
)
|
||||
self.downsample_layers.append(stem)
|
||||
for i in range(3):
|
||||
downsample_layer = nn.Sequential(
|
||||
LayerNorm(dims[i], eps=1e-6, data_format="channels_first"),
|
||||
nn.Conv2d(dims[i], dims[i + 1], kernel_size=2, stride=2),
|
||||
ops.Conv2d(dims[i], dims[i + 1], kernel_size=2, stride=2),
|
||||
)
|
||||
self.downsample_layers.append(downsample_layer)
|
||||
|
||||
@@ -104,7 +106,7 @@ class ConvNeXt(nn.Module):
|
||||
cur += depths[i]
|
||||
|
||||
def _init_weights(self, m):
|
||||
if isinstance(m, (nn.Conv2d, nn.Linear)):
|
||||
if isinstance(m, (ops.Conv2d, ops.Linear)):
|
||||
trunc_normal_(m.weight, std=.02)
|
||||
nn.init.constant_(m.bias, 0)
|
||||
|
||||
@@ -173,49 +175,49 @@ def convnext_tiny(pretrained=False, in_22k=False, **kwargs):
|
||||
return model
|
||||
|
||||
|
||||
@register_model
|
||||
def convnext_small(pretrained=False, in_22k=False, **kwargs):
|
||||
model = ConvNeXt(depths=[3, 3, 27, 3], dims=[96, 192, 384, 768], **kwargs)
|
||||
if pretrained:
|
||||
url = model_urls['convnext_small_22k'] if in_22k else model_urls['convnext_small_1k']
|
||||
checkpoint = torch.hub.load_state_dict_from_url(
|
||||
url=url, map_location="cpu")
|
||||
model.load_state_dict(checkpoint["model"])
|
||||
return model
|
||||
# @register_model
|
||||
# def convnext_small(pretrained=False, in_22k=False, **kwargs):
|
||||
# model = ConvNeXt(depths=[3, 3, 27, 3], dims=[96, 192, 384, 768], **kwargs)
|
||||
# if pretrained:
|
||||
# url = model_urls['convnext_small_22k'] if in_22k else model_urls['convnext_small_1k']
|
||||
# checkpoint = torch.hub.load_state_dict_from_url(
|
||||
# url=url, map_location="cpu")
|
||||
# model.load_state_dict(checkpoint["model"])
|
||||
# return model
|
||||
|
||||
|
||||
@register_model
|
||||
def convnext_base(pretrained=False, in_22k=False, **kwargs):
|
||||
model = ConvNeXt(depths=[3, 3, 27, 3], dims=[
|
||||
128, 256, 512, 1024], **kwargs)
|
||||
if pretrained:
|
||||
url = model_urls['convnext_base_22k'] if in_22k else model_urls['convnext_base_1k']
|
||||
checkpoint = torch.hub.load_state_dict_from_url(
|
||||
url=url, map_location="cpu")
|
||||
model.load_state_dict(checkpoint["model"])
|
||||
return model
|
||||
# @register_model
|
||||
# def convnext_base(pretrained=False, in_22k=False, **kwargs):
|
||||
# model = ConvNeXt(depths=[3, 3, 27, 3], dims=[
|
||||
# 128, 256, 512, 1024], **kwargs)
|
||||
# if pretrained:
|
||||
# url = model_urls['convnext_base_22k'] if in_22k else model_urls['convnext_base_1k']
|
||||
# checkpoint = torch.hub.load_state_dict_from_url(
|
||||
# url=url, map_location="cpu")
|
||||
# model.load_state_dict(checkpoint["model"])
|
||||
# return model
|
||||
|
||||
|
||||
@register_model
|
||||
def convnext_large(pretrained=False, in_22k=False, **kwargs):
|
||||
model = ConvNeXt(depths=[3, 3, 27, 3], dims=[
|
||||
192, 384, 768, 1536], **kwargs)
|
||||
if pretrained:
|
||||
url = model_urls['convnext_large_22k'] if in_22k else model_urls['convnext_large_1k']
|
||||
checkpoint = torch.hub.load_state_dict_from_url(
|
||||
url=url, map_location="cpu")
|
||||
model.load_state_dict(checkpoint["model"])
|
||||
return model
|
||||
# @register_model
|
||||
# def convnext_large(pretrained=False, in_22k=False, **kwargs):
|
||||
# model = ConvNeXt(depths=[3, 3, 27, 3], dims=[
|
||||
# 192, 384, 768, 1536], **kwargs)
|
||||
# if pretrained:
|
||||
# url = model_urls['convnext_large_22k'] if in_22k else model_urls['convnext_large_1k']
|
||||
# checkpoint = torch.hub.load_state_dict_from_url(
|
||||
# url=url, map_location="cpu")
|
||||
# model.load_state_dict(checkpoint["model"])
|
||||
# return model
|
||||
|
||||
|
||||
@register_model
|
||||
def convnext_xlarge(pretrained=False, in_22k=False, **kwargs):
|
||||
model = ConvNeXt(depths=[3, 3, 27, 3], dims=[
|
||||
256, 512, 1024, 2048], **kwargs)
|
||||
if pretrained:
|
||||
assert in_22k, "only ImageNet-22K pre-trained ConvNeXt-XL is available; please set in_22k=True"
|
||||
url = model_urls['convnext_xlarge_22k']
|
||||
checkpoint = torch.hub.load_state_dict_from_url(
|
||||
url=url, map_location="cpu")
|
||||
model.load_state_dict(checkpoint["model"])
|
||||
return model
|
||||
# @register_model
|
||||
# def convnext_xlarge(pretrained=False, in_22k=False, **kwargs):
|
||||
# model = ConvNeXt(depths=[3, 3, 27, 3], dims=[
|
||||
# 256, 512, 1024, 2048], **kwargs)
|
||||
# if pretrained:
|
||||
# assert in_22k, "only ImageNet-22K pre-trained ConvNeXt-XL is available; please set in_22k=True"
|
||||
# url = model_urls['convnext_xlarge_22k']
|
||||
# checkpoint = torch.hub.load_state_dict_from_url(
|
||||
# url=url, map_location="cpu")
|
||||
# model.load_state_dict(checkpoint["model"])
|
||||
# return model
|
||||
|
||||
@@ -3,12 +3,14 @@ import torch.nn as nn
|
||||
from .util import FourierEmbedder
|
||||
from .convnext import convnext_tiny
|
||||
|
||||
import comfy.ops
|
||||
ops = comfy.ops.manual_cast
|
||||
|
||||
class UniFusion(nn.Module):
|
||||
def __init__(self, in_dim, out_dim, mid_dim=3072, fourier_freqs=8,
|
||||
train_add_boxes=True, train_add_points=True, train_add_scribbles=True, train_add_masks=True,
|
||||
test_drop_boxes=False, test_drop_points=False, test_drop_scribbles=True, test_drop_masks=False,
|
||||
use_seperate_tokenizer=True):
|
||||
use_seperate_tokenizer=True, use_segs=True):
|
||||
super().__init__()
|
||||
self.in_dim = in_dim
|
||||
self.out_dim = out_dim
|
||||
@@ -26,14 +28,15 @@ class UniFusion(nn.Module):
|
||||
self.use_seperate_tokenizer = use_seperate_tokenizer
|
||||
|
||||
# Use instance masks as additional model inputs for mask conditioned image generation
|
||||
self.use_segs = True if self.add_masks else False
|
||||
#self.use_segs = True if self.add_masks else False
|
||||
self.use_segs = use_segs
|
||||
|
||||
if self.use_segs:
|
||||
in_dim = 30
|
||||
self.resize_input = 512
|
||||
self.down_factor = 64 # determined by the convnext backbone
|
||||
# from num_sem to 3 channels
|
||||
self.in_conv = nn.Conv2d(in_dim, 3, 3, 1, 1)
|
||||
self.in_conv = ops.Conv2d(in_dim, 3, 3, 1, 1)
|
||||
self.convnext_tiny_backbone = convnext_tiny(pretrained=True)
|
||||
self.num_tokens = (self.resize_input // self.down_factor) ** 2
|
||||
self.convnext_feature_dim = 3072
|
||||
@@ -79,19 +82,19 @@ class UniFusion(nn.Module):
|
||||
for idx, input_dim_ in enumerate(input_dim_list):
|
||||
mid_dim = self.mid_dim
|
||||
self.linears_list.append(nn.Sequential(
|
||||
nn.Linear(input_dim_, mid_dim),
|
||||
ops.Linear(input_dim_, mid_dim),
|
||||
nn.SiLU(),
|
||||
nn.Linear(mid_dim, mid_dim),
|
||||
ops.Linear(mid_dim, mid_dim),
|
||||
nn.SiLU(),
|
||||
nn.Linear(mid_dim, out_dim),
|
||||
ops.Linear(mid_dim, out_dim),
|
||||
))
|
||||
else:
|
||||
self.linears = nn.Sequential(
|
||||
nn.Linear(input_dim, self.mid_dim),
|
||||
ops.Linear(input_dim, self.mid_dim),
|
||||
nn.SiLU(),
|
||||
nn.Linear(self.mid_dim, self.mid_dim),
|
||||
ops.Linear(self.mid_dim, self.mid_dim),
|
||||
nn.SiLU(),
|
||||
nn.Linear(self.mid_dim, out_dim),
|
||||
ops.Linear(self.mid_dim, out_dim),
|
||||
)
|
||||
|
||||
self.null_positive_feature = torch.nn.Parameter(
|
||||
|
||||
@@ -0,0 +1,72 @@
|
||||
import os
|
||||
import folder_paths
|
||||
import comfy.utils
|
||||
from .. import constants as constants
|
||||
from ..model_helpers.prepare_positionnet import prepare_positionnet, get_positionnet_default_params
|
||||
from ..model_helpers.prepare_scaleu import prepare_scaleu_nets
|
||||
from ..model_helpers.prepare_fusers import prepare_fusers
|
||||
from huggingface_hub import snapshot_download
|
||||
|
||||
INSTANCE_FUSERS_DIR = "fuser_models"
|
||||
|
||||
INSTANCE_SCALEU_DIR = "scaleu_models"
|
||||
|
||||
class DownloadInstanceDiffusionModels:
|
||||
@classmethod
|
||||
def INPUT_TYPES(s):
|
||||
return {"required": {
|
||||
"use_segs": ("BOOLEAN", {"default": True}),
|
||||
}}
|
||||
|
||||
RETURN_TYPES = ("POSITIONNET", "FUSERS", "SCALEU", )
|
||||
FUNCTION = "load_model"
|
||||
|
||||
CATEGORY = "instance/loaders"
|
||||
|
||||
def load_model(self, use_segs: bool):
|
||||
repo_id = "logtd/instance_diffusion"
|
||||
instance_models_folder = os.path.join(folder_paths.models_dir, constants.INSTANCE_MODELS_DIR)
|
||||
|
||||
models_to_download = [
|
||||
("position_net", constants.INSTANCE_POSITIONNET_DIR, "position_net.ckpt"),
|
||||
("fusers", constants.INSTANCE_FUSERS_DIR, "fusers.ckpt"),
|
||||
("scaleu", constants.INSTANCE_SCALEU_DIR, "scaleu.ckpt")
|
||||
]
|
||||
|
||||
for model_name, model_folder, model_file in models_to_download:
|
||||
model_folder_path = os.path.join(instance_models_folder, model_folder)
|
||||
model_file_path = os.path.join(model_folder_path, model_file)
|
||||
|
||||
if not os.path.exists(model_file_path):
|
||||
print(f"Selected model: {model_file_path} not found, downloading...")
|
||||
allow_patterns = [f"*{model_name}*"]
|
||||
snapshot_download(repo_id=repo_id,
|
||||
allow_patterns=allow_patterns,
|
||||
local_dir=model_folder_path,
|
||||
local_dir_use_symlinks=False
|
||||
)
|
||||
|
||||
positionnet_file = os.path.join(instance_models_folder, constants.INSTANCE_POSITIONNET_DIR, "position_net.ckpt")
|
||||
fusers_file = os.path.join(instance_models_folder, constants.INSTANCE_FUSERS_DIR, "fusers.ckpt")
|
||||
scaleu_file = os.path.join(instance_models_folder, constants.INSTANCE_SCALEU_DIR, "scaleu.ckpt")
|
||||
|
||||
pos_checkpoint = comfy.utils.load_torch_file(positionnet_file, safe_load=True)
|
||||
params = get_positionnet_default_params()
|
||||
params["use_segs"] = use_segs
|
||||
model = prepare_positionnet(pos_checkpoint, params)
|
||||
positionnet = {
|
||||
'model': model,
|
||||
}
|
||||
|
||||
fusers_checkpoint = comfy.utils.load_torch_file(fusers_file, safe_load=True)
|
||||
fusers_list = prepare_fusers(fusers_checkpoint)
|
||||
fusers = {
|
||||
'model_list': fusers_list
|
||||
}
|
||||
scaleu_checkpoint = comfy.utils.load_torch_file(scaleu_file, safe_load=True)
|
||||
scaleu_list = prepare_scaleu_nets(scaleu_checkpoint)
|
||||
scaleu = {
|
||||
'model_list': scaleu_list
|
||||
}
|
||||
return (positionnet, fusers, scaleu)
|
||||
|
||||
@@ -8,17 +8,19 @@ class LoadInstancePositionNetNode:
|
||||
def INPUT_TYPES(s):
|
||||
return {"required": {
|
||||
"model_filename": (get_model_list(constants.INSTANCE_POSITIONNET_DIR),),
|
||||
"use_segs": ("BOOLEAN", {"default": True}),
|
||||
}}
|
||||
|
||||
RETURN_TYPES = ("POSITIONNET",)
|
||||
RETURN_TYPES = ("POSITIONNET", "FUSERS", "SCALEU",)
|
||||
FUNCTION = "load_model"
|
||||
|
||||
CATEGORY = "instance/loaders"
|
||||
|
||||
def load_model(self, model_filename: str):
|
||||
def load_model(self, model_filename: str, use_segs: bool):
|
||||
checkpoint = load_checkpoint(
|
||||
constants.INSTANCE_POSITIONNET_DIR, model_filename)
|
||||
params = get_positionnet_default_params()
|
||||
params["use_segs"] = use_segs
|
||||
model = prepare_positionnet(checkpoint, params)
|
||||
positionnet = {
|
||||
'model': model,
|
||||
|
||||
@@ -0,0 +1 @@
|
||||
huggingface_hub
|
||||
Reference in New Issue
Block a user