HiDream Image

Author: hlky

src/diffusers/__init__.py

@@ -169,6 +169,7 @@
             "FluxControlNetModel",
             "FluxMultiControlNetModel",
             "FluxTransformer2DModel",
+            "HiDreamImageTransformer2DModel",
             "HunyuanDiT2DControlNetModel",
             "HunyuanDiT2DModel",
             "HunyuanDiT2DMultiControlNetModel",
@@ -366,6 +367,7 @@
             "FluxInpaintPipeline",
             "FluxPipeline",
             "FluxPriorReduxPipeline",
+            "HiDreamImagePipeline",
             "HunyuanDiTControlNetPipeline",
             "HunyuanDiTPAGPipeline",
             "HunyuanDiTPipeline",
@@ -745,6 +747,7 @@
             FluxControlNetModel,
             FluxMultiControlNetModel,
             FluxTransformer2DModel,
+            HiDreamImageTransformer2DModel,
             HunyuanDiT2DControlNetModel,
             HunyuanDiT2DModel,
             HunyuanDiT2DMultiControlNetModel,
@@ -921,6 +924,7 @@
             FluxInpaintPipeline,
             FluxPipeline,
             FluxPriorReduxPipeline,
+            HiDreamImagePipeline,
             HunyuanDiTControlNetPipeline,
             HunyuanDiTPAGPipeline,
             HunyuanDiTPipeline,

src/diffusers/models/__init__.py

@@ -75,6 +75,7 @@
     _import_structure["transformers.transformer_cogview4"] = ["CogView4Transformer2DModel"]
     _import_structure["transformers.transformer_easyanimate"] = ["EasyAnimateTransformer3DModel"]
     _import_structure["transformers.transformer_flux"] = ["FluxTransformer2DModel"]
+    _import_structure["transformers.transformer_hidream_image"] = ["HiDreamImageTransformer2DModel"]
     _import_structure["transformers.transformer_hunyuan_video"] = ["HunyuanVideoTransformer3DModel"]
     _import_structure["transformers.transformer_ltx"] = ["LTXVideoTransformer3DModel"]
     _import_structure["transformers.transformer_lumina2"] = ["Lumina2Transformer2DModel"]
@@ -149,6 +150,7 @@
             DualTransformer2DModel,
             EasyAnimateTransformer3DModel,
             FluxTransformer2DModel,
+            HiDreamImageTransformer2DModel,
             HunyuanDiT2DModel,
             HunyuanVideoTransformer3DModel,
             LatteTransformer3DModel,

src/diffusers/models/attention.py

@@ -1249,3 +1249,33 @@ def forward(self, hidden_states: torch.Tensor, *args, **kwargs) -> torch.Tensor:
         for module in self.net:
             hidden_states = module(hidden_states)
         return hidden_states
+
+
+class HiDreamImageFeedForwardSwiGLU(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        hidden_dim: int,
+        multiple_of: int = 256,
+        ffn_dim_multiplier: Optional[float] = None,
+    ):
+        super().__init__()
+        hidden_dim = int(2 * hidden_dim / 3)
+        # custom dim factor multiplier
+        if ffn_dim_multiplier is not None:
+            hidden_dim = int(ffn_dim_multiplier * hidden_dim)
+        hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
+
+        self.w1 = nn.Linear(dim, hidden_dim, bias=False)
+        self.w2 = nn.Linear(hidden_dim, dim, bias=False)
+        self.w3 = nn.Linear(dim, hidden_dim, bias=False)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            nn.init.xavier_uniform_(m.weight)
+            if m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+
+    def forward(self, x):
+        return self.w2(torch.nn.functional.silu(self.w1(x)) * self.w3(x))

src/diffusers/models/embeddings.py

@@ -2621,3 +2621,59 @@ def forward(self, image_embeds: List[torch.Tensor]):
             projected_image_embeds.append(image_embed)
 
         return projected_image_embeds
+
+
+class HiDreamImagePooledEmbed(nn.Module):
+    def __init__(self, text_emb_dim, hidden_size):
+        super().__init__()
+        self.pooled_embedder = TimestepEmbedding(in_channels=text_emb_dim, time_embed_dim=hidden_size)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            nn.init.normal_(m.weight, std=0.02)
+            if m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+
+    def forward(self, pooled_embed):
+        return self.pooled_embedder(pooled_embed)
+
+
+class HiDreamImageTimestepEmbed(nn.Module):
+    def __init__(self, hidden_size, frequency_embedding_size=256):
+        super().__init__()
+        self.time_proj = Timesteps(num_channels=frequency_embedding_size, flip_sin_to_cos=True, downscale_freq_shift=0)
+        self.timestep_embedder = TimestepEmbedding(in_channels=frequency_embedding_size, time_embed_dim=hidden_size)
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            nn.init.normal_(m.weight, std=0.02)
+            if m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+
+    def forward(self, timesteps, wdtype):
+        t_emb = self.time_proj(timesteps).to(dtype=wdtype)
+        t_emb = self.timestep_embedder(t_emb)
+        return t_emb
+
+
+class HiDreamImageOutEmbed(nn.Module):
+    def __init__(self, hidden_size, patch_size, out_channels):
+        super().__init__()
+        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+        self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
+        self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True))
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            nn.init.zeros_(m.weight)
+            if m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+
+    def forward(self, x, adaln_input):
+        shift, scale = self.adaLN_modulation(adaln_input).chunk(2, dim=1)
+        x = self.norm_final(x) * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+        x = self.linear(x)
+        return x

src/diffusers/models/transformers/__init__.py

@@ -21,6 +21,7 @@
     from .transformer_cogview4 import CogView4Transformer2DModel
     from .transformer_easyanimate import EasyAnimateTransformer3DModel
     from .transformer_flux import FluxTransformer2DModel
+    from .transformer_hidream_image import HiDreamImageTransformer2DModel
     from .transformer_hunyuan_video import HunyuanVideoTransformer3DModel
     from .transformer_ltx import LTXVideoTransformer3DModel
     from .transformer_lumina2 import Lumina2Transformer2DModel