more

Author: yiyixuxu

src/diffusers/loaders/unet.py

@@ -22,7 +22,7 @@
 from huggingface_hub.utils import validate_hf_hub_args
 from torch import nn
 
-from ..models.embeddings import ImageProjection, MLPProjection, IPAdapterPlusImageProjection
+from ..models.embeddings import ImageProjection, IPAdapterFullImageProjection, IPAdapterPlusImageProjection
 from ..models.modeling_utils import _LOW_CPU_MEM_USAGE_DEFAULT, load_model_dict_into_meta
 from ..utils import (
     USE_PEFT_BACKEND,
@@ -689,7 +689,7 @@ def _convert_ip_adapter_image_proj_to_diffusers(self, state_dict):
             clip_embeddings_dim = state_dict["proj.0.weight"].shape[0]
             cross_attention_dim = state_dict["proj.3.weight"].shape[0]
 
-            image_projection = MLPProjection(
+            image_projection = IPAdapterFullImageProjection(
                 cross_attention_dim=cross_attention_dim, image_embed_dim=clip_embeddings_dim
             )
 
@@ -707,7 +707,7 @@ def _convert_ip_adapter_image_proj_to_diffusers(self, state_dict):
             hidden_dims = state_dict["latents"].shape[2]
             heads = state_dict["layers.0.0.to_q.weight"].shape[0] // 64
 
-            image_projection = Resampler(
+            image_projection = IPAdapterPlusImageProjection(
                 embed_dims=embed_dims,
                 output_dims=output_dims,
                 hidden_dims=hidden_dims,

src/diffusers/models/embeddings.py

@@ -461,7 +461,7 @@ def forward(self, image_embeds: torch.FloatTensor):
         return image_embeds
 
 
-class MLPProjection(nn.Module):
+class IPAdapterFullImageProjection(nn.Module):
     def __init__(self, image_embed_dim=1024, cross_attention_dim=1024):
         super().__init__()
         from .attention import FeedForward
@@ -640,6 +640,7 @@ def get_fourier_embeds_from_boundingbox(embed_dim, box):
 
     return emb
 
+
 class GLIGENTextBoundingboxProjection(nn.Module):
     def __init__(self, positive_len, out_dim, feature_type="text-only", fourier_freqs=8):
         super().__init__()

src/diffusers/models/unet_2d_condition.py

@@ -32,10 +32,10 @@
 )
 from .embeddings import (
     GaussianFourierProjection,
+    GLIGENTextBoundingboxProjection,
     ImageHintTimeEmbedding,
     ImageProjection,
     ImageTimeEmbedding,
-    GLIGENTextBoundingboxProjection,
     TextImageProjection,
     TextImageTimeEmbedding,
     TextTimeEmbedding,