diff --git a/__init__.py b/__init__.py index ec86e01..f61d71f 100644 --- a/__init__.py +++ b/__init__.py @@ -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" } diff --git a/model_helpers/prepare_fusers.py b/model_helpers/prepare_fusers.py index 2d2858a..8951bec 100644 --- a/model_helpers/prepare_fusers.py +++ b/model_helpers/prepare_fusers.py @@ -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 = [] diff --git a/model_helpers/prepare_positionnet.py b/model_helpers/prepare_positionnet.py index 5c05830..4eff4e6 100644 --- a/model_helpers/prepare_positionnet.py +++ b/model_helpers/prepare_positionnet.py @@ -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 diff --git a/modules/attention.py b/modules/attention.py deleted file mode 100644 index 54ed650..0000000 --- a/modules/attention.py +++ /dev/null @@ -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) diff --git a/modules/convnext.py b/modules/convnext.py index 54e058e..8771527 100644 --- a/modules/convnext.py +++ b/modules/convnext.py @@ -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 diff --git a/modules/text_grounding_net.py b/modules/text_grounding_net.py index ba77d7f..3302226 100644 --- a/modules/text_grounding_net.py +++ b/modules/text_grounding_net.py @@ -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( diff --git a/nodes/download_and_load_models.py b/nodes/download_and_load_models.py new file mode 100644 index 0000000..052e129 --- /dev/null +++ b/nodes/download_and_load_models.py @@ -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) + diff --git a/nodes/load_instance_positionnet_node.py b/nodes/load_instance_positionnet_node.py index 5ee8b9d..7548607 100644 --- a/nodes/load_instance_positionnet_node.py +++ b/nodes/load_instance_positionnet_node.py @@ -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, diff --git a/requirements.txt b/requirements.txt index e69de29..b2642b3 100644 --- a/requirements.txt +++ b/requirements.txt @@ -0,0 +1 @@ +huggingface_hub \ No newline at end of file