Add: FLUX

diffsynth/configs/model_configs.py

@@ -364,6 +364,55 @@ flux_series = [
         "model_name": "flux_lora_patcher",
         "model_class": "diffsynth.models.flux_lora_patcher.FluxLoraPatcher",
     },
+    {
+        # Example: ModelConfig(model_id="DiffSynth-Studio/Nexus-GenV2", origin_file_pattern="model*.safetensors")
+        "model_hash": "2bd19e845116e4f875a0a048e27fc219",
+        "model_name": "nexus_gen_llm",
+        "model_class": "diffsynth.models.nexus_gen.NexusGenAutoregressiveModel",
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.nexus_gen.NexusGenAutoregressiveModelStateDictConverter",
+    },
+    {
+        # Example: ModelConfig(model_id="DiffSynth-Studio/Nexus-GenV2", origin_file_pattern="edit_decoder.bin")
+        "model_hash": "63c969fd37cce769a90aa781fbff5f81",
+        "model_name": "nexus_gen_editing_adapter",
+        "model_class": "diffsynth.models.nexus_gen_projector.NexusGenImageEmbeddingMerger",
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.nexus_gen_projector.NexusGenMergerStateDictConverter",
+    },
+    {
+        # Example: ModelConfig(model_id="DiffSynth-Studio/Nexus-GenV2", origin_file_pattern="edit_decoder.bin")
+        "model_hash": "63c969fd37cce769a90aa781fbff5f81",
+        "model_name": "flux_dit",
+        "model_class": "diffsynth.models.flux_dit.FluxDiT",
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.flux_dit.FluxDiTStateDictConverter",
+    },
+    {
+        # Example: ModelConfig(model_id="DiffSynth-Studio/Nexus-GenV2", origin_file_pattern="generation_decoder.bin")
+        "model_hash": "3e6c61b0f9471135fc9c6d6a98e98b6d",
+        "model_name": "nexus_gen_generation_adapter",
+        "model_class": "diffsynth.models.nexus_gen_projector.NexusGenAdapter",
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.nexus_gen_projector.NexusGenAdapterStateDictConverter",
+    },
+    {
+        # Example: ModelConfig(model_id="DiffSynth-Studio/Nexus-GenV2", origin_file_pattern="generation_decoder.bin")
+        "model_hash": "3e6c61b0f9471135fc9c6d6a98e98b6d",
+        "model_name": "flux_dit",
+        "model_class": "diffsynth.models.flux_dit.FluxDiT",
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.flux_dit.FluxDiTStateDictConverter",
+    },
+    {
+        # Example: ModelConfig(model_id="InstantX/FLUX.1-dev-IP-Adapter", origin_file_pattern="ip-adapter.bin")
+        "model_hash": "4daaa66cc656a8fe369908693dad0a35",
+        "model_name": "flux_ipadapter",
+        "model_class": "diffsynth.models.flux_ipadapter.FluxIpAdapter",
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.flux_ipadapter.FluxIpAdapterStateDictConverter",
+    },
+    {
+        # Example: ModelConfig(model_id="google/siglip-so400m-patch14-384", origin_file_pattern="model.safetensors")
+        "model_hash": "04d8c1e20a1f1b25f7434f111992a33f",
+        "model_name": "siglip_vision_model",
+        "model_class": "diffsynth.models.flux_ipadapter.SiglipVisionModelSO400M",
+        "state_dict_converter": "diffsynth.utils.state_dict_converters.flux_ipadapter.SiglipStateDictConverter",
+    },
 ]
 
 MODEL_CONFIGS = qwen_image_series + wan_series + flux_series

diffsynth/models/flux_ipadapter.py

@@ -5,34 +5,21 @@ import torch
 
 class SiglipVisionModelSO400M(SiglipVisionModel):
     def __init__(self):
-        config = SiglipVisionConfig(**{
-            "architectures": [
-                "SiglipModel"
-            ],
-            "initializer_factor": 1.0,
-            "model_type": "siglip",
-            "text_config": {
-                "hidden_size": 1152,
-                "intermediate_size": 4304,
-                "model_type": "siglip_text_model",
-                "num_attention_heads": 16,
-                "num_hidden_layers": 27
-            },
-            "torch_dtype": "float32",
-            "transformers_version": "4.37.0.dev0",
-            "vision_config": {
-                "hidden_size": 1152,
-                "image_size": 384,
-                "intermediate_size": 4304,
-                "model_type": "siglip_vision_model",
-                "num_attention_heads": 16,
-                "num_hidden_layers": 27,
-                "patch_size": 14
-            }
-        })
+        config = SiglipVisionConfig(
+            hidden_size=1152,
+            image_size=384,
+            intermediate_size=4304,
+            model_type="siglip_vision_model",
+            num_attention_heads=16,
+            num_hidden_layers=27,
+            patch_size=14,
+            architectures=["SiglipModel"],
+            initializer_factor=1.0,
+            torch_dtype="float32",
+            transformers_version="4.37.0.dev0"
+        )
         super().__init__(config)
 
-
 class MLPProjModel(torch.nn.Module):
     def __init__(self, cross_attention_dim=768, id_embeddings_dim=512, num_tokens=4):
         super().__init__()

diffsynth/models/flux_lora_patcher.py

@@ -1,4 +1,251 @@
-import torch
+import torch, math
+from ..core.loader import load_state_dict
+from typing import Union
+
+class GeneralLoRALoader:
+    def __init__(self, device="cpu", torch_dtype=torch.float32):
+        self.device = device
+        self.torch_dtype = torch_dtype
+    
+    
+    def get_name_dict(self, lora_state_dict):
+        lora_name_dict = {}
+        for key in lora_state_dict:
+            if ".lora_B." not in key:
+                continue
+            keys = key.split(".")
+            if len(keys) > keys.index("lora_B") + 2:
+                keys.pop(keys.index("lora_B") + 1)
+            keys.pop(keys.index("lora_B"))
+            if keys[0] == "diffusion_model":
+                keys.pop(0)
+            keys.pop(-1)
+            target_name = ".".join(keys)
+            lora_name_dict[target_name] = (key, key.replace(".lora_B.", ".lora_A."))
+        return lora_name_dict
+
+
+    def load(self, model: torch.nn.Module, state_dict_lora, alpha=1.0):
+        updated_num = 0
+        lora_name_dict = self.get_name_dict(state_dict_lora)
+        for name, module in model.named_modules():
+            if name in lora_name_dict:
+                weight_up = state_dict_lora[lora_name_dict[name][0]].to(device=self.device, dtype=self.torch_dtype)
+                weight_down = state_dict_lora[lora_name_dict[name][1]].to(device=self.device, dtype=self.torch_dtype)
+                if len(weight_up.shape) == 4:
+                    weight_up = weight_up.squeeze(3).squeeze(2)
+                    weight_down = weight_down.squeeze(3).squeeze(2)
+                    weight_lora = alpha * torch.mm(weight_up, weight_down).unsqueeze(2).unsqueeze(3)
+                else:
+                    weight_lora = alpha * torch.mm(weight_up, weight_down)
+                state_dict = module.state_dict()
+                state_dict["weight"] = state_dict["weight"].to(device=self.device, dtype=self.torch_dtype) + weight_lora
+                module.load_state_dict(state_dict)
+                updated_num += 1
+        print(f"{updated_num} tensors are updated by LoRA.")
+
+class FluxLoRALoader(GeneralLoRALoader):
+    def __init__(self, device="cpu", torch_dtype=torch.float32):
+        super().__init__(device=device, torch_dtype=torch_dtype)
+    
+        self.diffusers_rename_dict = {
+            "transformer.single_transformer_blocks.blockid.attn.to_k.lora_A.weight":"single_blocks.blockid.a_to_k.lora_A.default.weight",
+            "transformer.single_transformer_blocks.blockid.attn.to_k.lora_B.weight":"single_blocks.blockid.a_to_k.lora_B.default.weight",
+            "transformer.single_transformer_blocks.blockid.attn.to_q.lora_A.weight":"single_blocks.blockid.a_to_q.lora_A.default.weight",
+            "transformer.single_transformer_blocks.blockid.attn.to_q.lora_B.weight":"single_blocks.blockid.a_to_q.lora_B.default.weight",
+            "transformer.single_transformer_blocks.blockid.attn.to_v.lora_A.weight":"single_blocks.blockid.a_to_v.lora_A.default.weight",
+            "transformer.single_transformer_blocks.blockid.attn.to_v.lora_B.weight":"single_blocks.blockid.a_to_v.lora_B.default.weight",
+            "transformer.single_transformer_blocks.blockid.norm.linear.lora_A.weight":"single_blocks.blockid.norm.linear.lora_A.default.weight",
+            "transformer.single_transformer_blocks.blockid.norm.linear.lora_B.weight":"single_blocks.blockid.norm.linear.lora_B.default.weight",
+            "transformer.single_transformer_blocks.blockid.proj_mlp.lora_A.weight":"single_blocks.blockid.proj_in_besides_attn.lora_A.default.weight",
+            "transformer.single_transformer_blocks.blockid.proj_mlp.lora_B.weight":"single_blocks.blockid.proj_in_besides_attn.lora_B.default.weight",
+            "transformer.single_transformer_blocks.blockid.proj_out.lora_A.weight":"single_blocks.blockid.proj_out.lora_A.default.weight",
+            "transformer.single_transformer_blocks.blockid.proj_out.lora_B.weight":"single_blocks.blockid.proj_out.lora_B.default.weight",
+            "transformer.transformer_blocks.blockid.attn.add_k_proj.lora_A.weight":"blocks.blockid.attn.b_to_k.lora_A.default.weight",
+            "transformer.transformer_blocks.blockid.attn.add_k_proj.lora_B.weight":"blocks.blockid.attn.b_to_k.lora_B.default.weight",
+            "transformer.transformer_blocks.blockid.attn.add_q_proj.lora_A.weight":"blocks.blockid.attn.b_to_q.lora_A.default.weight",
+            "transformer.transformer_blocks.blockid.attn.add_q_proj.lora_B.weight":"blocks.blockid.attn.b_to_q.lora_B.default.weight",
+            "transformer.transformer_blocks.blockid.attn.add_v_proj.lora_A.weight":"blocks.blockid.attn.b_to_v.lora_A.default.weight",
+            "transformer.transformer_blocks.blockid.attn.add_v_proj.lora_B.weight":"blocks.blockid.attn.b_to_v.lora_B.default.weight",
+            "transformer.transformer_blocks.blockid.attn.to_add_out.lora_A.weight":"blocks.blockid.attn.b_to_out.lora_A.default.weight",
+            "transformer.transformer_blocks.blockid.attn.to_add_out.lora_B.weight":"blocks.blockid.attn.b_to_out.lora_B.default.weight",
+            "transformer.transformer_blocks.blockid.attn.to_k.lora_A.weight":"blocks.blockid.attn.a_to_k.lora_A.default.weight",
+            "transformer.transformer_blocks.blockid.attn.to_k.lora_B.weight":"blocks.blockid.attn.a_to_k.lora_B.default.weight",
+            "transformer.transformer_blocks.blockid.attn.to_out.0.lora_A.weight":"blocks.blockid.attn.a_to_out.lora_A.default.weight",
+            "transformer.transformer_blocks.blockid.attn.to_out.0.lora_B.weight":"blocks.blockid.attn.a_to_out.lora_B.default.weight",
+            "transformer.transformer_blocks.blockid.attn.to_q.lora_A.weight":"blocks.blockid.attn.a_to_q.lora_A.default.weight",
+            "transformer.transformer_blocks.blockid.attn.to_q.lora_B.weight":"blocks.blockid.attn.a_to_q.lora_B.default.weight",
+            "transformer.transformer_blocks.blockid.attn.to_v.lora_A.weight":"blocks.blockid.attn.a_to_v.lora_A.default.weight",
+            "transformer.transformer_blocks.blockid.attn.to_v.lora_B.weight":"blocks.blockid.attn.a_to_v.lora_B.default.weight",
+            "transformer.transformer_blocks.blockid.ff.net.0.proj.lora_A.weight":"blocks.blockid.ff_a.0.lora_A.default.weight",
+            "transformer.transformer_blocks.blockid.ff.net.0.proj.lora_B.weight":"blocks.blockid.ff_a.0.lora_B.default.weight",
+            "transformer.transformer_blocks.blockid.ff.net.2.lora_A.weight":"blocks.blockid.ff_a.2.lora_A.default.weight",
+            "transformer.transformer_blocks.blockid.ff.net.2.lora_B.weight":"blocks.blockid.ff_a.2.lora_B.default.weight",
+            "transformer.transformer_blocks.blockid.ff_context.net.0.proj.lora_A.weight":"blocks.blockid.ff_b.0.lora_A.default.weight",
+            "transformer.transformer_blocks.blockid.ff_context.net.0.proj.lora_B.weight":"blocks.blockid.ff_b.0.lora_B.default.weight",
+            "transformer.transformer_blocks.blockid.ff_context.net.2.lora_A.weight":"blocks.blockid.ff_b.2.lora_A.default.weight",
+            "transformer.transformer_blocks.blockid.ff_context.net.2.lora_B.weight":"blocks.blockid.ff_b.2.lora_B.default.weight",
+            "transformer.transformer_blocks.blockid.norm1.linear.lora_A.weight":"blocks.blockid.norm1_a.linear.lora_A.default.weight",
+            "transformer.transformer_blocks.blockid.norm1.linear.lora_B.weight":"blocks.blockid.norm1_a.linear.lora_B.default.weight",
+            "transformer.transformer_blocks.blockid.norm1_context.linear.lora_A.weight":"blocks.blockid.norm1_b.linear.lora_A.default.weight",
+            "transformer.transformer_blocks.blockid.norm1_context.linear.lora_B.weight":"blocks.blockid.norm1_b.linear.lora_B.default.weight",
+        }
+
+        self.civitai_rename_dict = {
+            "lora_unet_double_blocks_blockid_img_mod_lin.lora_down.weight": "blocks.blockid.norm1_a.linear.lora_A.default.weight",
+            "lora_unet_double_blocks_blockid_img_mod_lin.lora_up.weight": "blocks.blockid.norm1_a.linear.lora_B.default.weight",
+            "lora_unet_double_blocks_blockid_txt_mod_lin.lora_down.weight": "blocks.blockid.norm1_b.linear.lora_A.default.weight",
+            "lora_unet_double_blocks_blockid_txt_mod_lin.lora_up.weight": "blocks.blockid.norm1_b.linear.lora_B.default.weight",
+            "lora_unet_double_blocks_blockid_img_attn_qkv.lora_down.weight": "blocks.blockid.attn.a_to_qkv.lora_A.default.weight",
+            "lora_unet_double_blocks_blockid_img_attn_qkv.lora_up.weight": "blocks.blockid.attn.a_to_qkv.lora_B.default.weight",
+            "lora_unet_double_blocks_blockid_txt_attn_qkv.lora_down.weight": "blocks.blockid.attn.b_to_qkv.lora_A.default.weight",
+            "lora_unet_double_blocks_blockid_txt_attn_qkv.lora_up.weight": "blocks.blockid.attn.b_to_qkv.lora_B.default.weight",
+            "lora_unet_double_blocks_blockid_img_attn_proj.lora_down.weight": "blocks.blockid.attn.a_to_out.lora_A.default.weight",
+            "lora_unet_double_blocks_blockid_img_attn_proj.lora_up.weight": "blocks.blockid.attn.a_to_out.lora_B.default.weight",
+            "lora_unet_double_blocks_blockid_txt_attn_proj.lora_down.weight": "blocks.blockid.attn.b_to_out.lora_A.default.weight",
+            "lora_unet_double_blocks_blockid_txt_attn_proj.lora_up.weight": "blocks.blockid.attn.b_to_out.lora_B.default.weight",
+            "lora_unet_double_blocks_blockid_img_mlp_0.lora_down.weight": "blocks.blockid.ff_a.0.lora_A.default.weight",
+            "lora_unet_double_blocks_blockid_img_mlp_0.lora_up.weight": "blocks.blockid.ff_a.0.lora_B.default.weight",
+            "lora_unet_double_blocks_blockid_img_mlp_2.lora_down.weight": "blocks.blockid.ff_a.2.lora_A.default.weight",
+            "lora_unet_double_blocks_blockid_img_mlp_2.lora_up.weight": "blocks.blockid.ff_a.2.lora_B.default.weight",
+            "lora_unet_double_blocks_blockid_txt_mlp_0.lora_down.weight": "blocks.blockid.ff_b.0.lora_A.default.weight",
+            "lora_unet_double_blocks_blockid_txt_mlp_0.lora_up.weight": "blocks.blockid.ff_b.0.lora_B.default.weight",
+            "lora_unet_double_blocks_blockid_txt_mlp_2.lora_down.weight": "blocks.blockid.ff_b.2.lora_A.default.weight",
+            "lora_unet_double_blocks_blockid_txt_mlp_2.lora_up.weight": "blocks.blockid.ff_b.2.lora_B.default.weight",
+            "lora_unet_single_blocks_blockid_modulation_lin.lora_down.weight": "single_blocks.blockid.norm.linear.lora_A.default.weight",
+            "lora_unet_single_blocks_blockid_modulation_lin.lora_up.weight": "single_blocks.blockid.norm.linear.lora_B.default.weight",
+            "lora_unet_single_blocks_blockid_linear1.lora_down.weight": "single_blocks.blockid.to_qkv_mlp.lora_A.default.weight",
+            "lora_unet_single_blocks_blockid_linear1.lora_up.weight": "single_blocks.blockid.to_qkv_mlp.lora_B.default.weight",
+            "lora_unet_single_blocks_blockid_linear2.lora_down.weight": "single_blocks.blockid.proj_out.lora_A.default.weight",
+            "lora_unet_single_blocks_blockid_linear2.lora_up.weight": "single_blocks.blockid.proj_out.lora_B.default.weight",
+        }
+
+    def load(self, model: torch.nn.Module, state_dict_lora, alpha=1.0):
+        super().load(model, state_dict_lora, alpha)
+
+    
+    def convert_state_dict(self,state_dict):
+
+        def guess_block_id(name,model_resource):
+            if model_resource == 'civitai':
+                names = name.split("_")
+                for i in names:
+                    if i.isdigit():
+                        return i, name.replace(f"_{i}_", "_blockid_")
+            if model_resource == 'diffusers':
+                names = name.split(".")
+                for i in names:
+                    if i.isdigit():
+                        return i, name.replace(f"transformer_blocks.{i}.", "transformer_blocks.blockid.")
+            return None, None
+
+        def guess_resource(state_dict):
+            for k in state_dict:
+                if "lora_unet_" in k:
+                    return 'civitai'
+                elif k.startswith("transformer."):
+                    return 'diffusers'
+                else:
+                    None
+        
+        model_resource = guess_resource(state_dict)
+        if model_resource is None:
+            return state_dict
+
+        rename_dict = self.diffusers_rename_dict if model_resource == 'diffusers' else self.civitai_rename_dict
+        def guess_alpha(state_dict):
+                for name, param in state_dict.items():
+                    if ".alpha" in name:
+                        for suffix in [".lora_down.weight", ".lora_A.weight"]:
+                            name_ = name.replace(".alpha", suffix)
+                            if name_ in state_dict:
+                                lora_alpha = param.item() / state_dict[name_].shape[0]
+                                lora_alpha = math.sqrt(lora_alpha)
+                                return lora_alpha
+
+                return 1
+        
+        alpha = guess_alpha(state_dict)
+        
+        state_dict_ = {}
+        for name, param in state_dict.items():
+            block_id, source_name = guess_block_id(name,model_resource)
+            if alpha != 1:
+                param *= alpha
+            if source_name in rename_dict:
+                target_name = rename_dict[source_name]
+                target_name = target_name.replace(".blockid.", f".{block_id}.")
+                state_dict_[target_name] = param
+            else:
+                state_dict_[name] = param
+        
+        if model_resource == 'diffusers':
+            for name in list(state_dict_.keys()):
+                if "single_blocks." in name and ".a_to_q." in name:
+                    mlp = state_dict_.get(name.replace(".a_to_q.", ".proj_in_besides_attn."), None)
+                    if mlp is None:
+                        dim = 4
+                        if 'lora_A' in name:
+                            dim = 1
+                        mlp = torch.zeros(dim * state_dict_[name].shape[0],
+                                        *state_dict_[name].shape[1:],
+                                        dtype=state_dict_[name].dtype)
+                    else:
+                        state_dict_.pop(name.replace(".a_to_q.", ".proj_in_besides_attn."))
+                    if 'lora_A' in name:
+                        param = torch.concat([
+                            state_dict_.pop(name),
+                            state_dict_.pop(name.replace(".a_to_q.", ".a_to_k.")),
+                            state_dict_.pop(name.replace(".a_to_q.", ".a_to_v.")),
+                            mlp,
+                        ], dim=0)
+                    elif 'lora_B' in name:
+                        d, r = state_dict_[name].shape
+                        param = torch.zeros((3*d+mlp.shape[0], 3*r+mlp.shape[1]), dtype=state_dict_[name].dtype, device=state_dict_[name].device)
+                        param[:d, :r] = state_dict_.pop(name)
+                        param[d:2*d, r:2*r] = state_dict_.pop(name.replace(".a_to_q.", ".a_to_k."))
+                        param[2*d:3*d, 2*r:3*r] = state_dict_.pop(name.replace(".a_to_q.", ".a_to_v."))
+                        param[3*d:, 3*r:] = mlp
+                    else:
+                        param = torch.concat([
+                            state_dict_.pop(name),
+                            state_dict_.pop(name.replace(".a_to_q.", ".a_to_k.")),
+                            state_dict_.pop(name.replace(".a_to_q.", ".a_to_v.")),
+                            mlp,
+                        ], dim=0)
+                    name_ = name.replace(".a_to_q.", ".to_qkv_mlp.")
+                    state_dict_[name_] = param
+            for name in list(state_dict_.keys()):
+                for component in ["a", "b"]:
+                    if f".{component}_to_q." in name:
+                        name_ = name.replace(f".{component}_to_q.", f".{component}_to_qkv.")
+                        concat_dim = 0
+                        if 'lora_A' in name:
+                            param = torch.concat([
+                                state_dict_[name.replace(f".{component}_to_q.", f".{component}_to_q.")],
+                                state_dict_[name.replace(f".{component}_to_q.", f".{component}_to_k.")],
+                                state_dict_[name.replace(f".{component}_to_q.", f".{component}_to_v.")],
+                            ], dim=0)
+                        elif 'lora_B' in name:
+                            origin = state_dict_[name.replace(f".{component}_to_q.", f".{component}_to_q.")]
+                            d, r = origin.shape
+                            # print(d, r)
+                            param = torch.zeros((3*d, 3*r), dtype=origin.dtype, device=origin.device)
+                            param[:d, :r] = state_dict_[name.replace(f".{component}_to_q.", f".{component}_to_q.")]
+                            param[d:2*d, r:2*r] = state_dict_[name.replace(f".{component}_to_q.", f".{component}_to_k.")]
+                            param[2*d:3*d, 2*r:3*r] = state_dict_[name.replace(f".{component}_to_q.", f".{component}_to_v.")]
+                        else:
+                            param = torch.concat([
+                                state_dict_[name.replace(f".{component}_to_q.", f".{component}_to_q.")],
+                                state_dict_[name.replace(f".{component}_to_q.", f".{component}_to_k.")],
+                                state_dict_[name.replace(f".{component}_to_q.", f".{component}_to_v.")],
+                            ], dim=0)
+                        state_dict_[name_] = param
+                        state_dict_.pop(name.replace(f".{component}_to_q.", f".{component}_to_q."))
+                        state_dict_.pop(name.replace(f".{component}_to_q.", f".{component}_to_k."))
+                        state_dict_.pop(name.replace(f".{component}_to_q.", f".{component}_to_v."))  
+        return state_dict_
+
 
 class LoraMerger(torch.nn.Module):
     def __init__(self, dim):
@@ -57,4 +304,3 @@ class FluxLoraPatcher(torch.nn.Module):
         
     def forward(self, base_output, lora_outputs, name):
         return self.model_dict[name.replace(".", "___")](base_output, lora_outputs)
-    

diffsynth/models/nexus_gen.py

@@ -0,0 +1,161 @@
+import torch
+from PIL import Image
+
+
+class NexusGenAutoregressiveModel(torch.nn.Module):
+    def __init__(self, max_length=1024, max_pixels=262640):
+        super(NexusGenAutoregressiveModel, self).__init__()
+        from .nexus_gen_ar_model import Qwen2_5_VLForConditionalGeneration
+        from transformers import Qwen2_5_VLConfig
+        self.max_length = max_length
+        self.max_pixels = max_pixels
+        model_config = Qwen2_5_VLConfig(**{
+            "_name_or_path": "DiffSynth-Studio/Nexus-GenV2",
+            "architectures": [
+                "Qwen2_5_VLForConditionalGeneration"
+            ],
+            "attention_dropout": 0.0,
+            "auto_map": {
+                "AutoConfig": "configuration_qwen2_5_vl.Qwen2_5_VLConfig",
+                "AutoModel": "modeling_qwen2_5_vl.Qwen2_5_VLModel",
+                "AutoModelForCausalLM": "modeling_qwen2_5_vl.Qwen2_5_VLForConditionalGeneration"
+            },
+            "bos_token_id": 151643,
+            "eos_token_id": 151645,
+            "hidden_act": "silu",
+            "hidden_size": 3584,
+            "image_token_id": 151655,
+            "initializer_range": 0.02,
+            "intermediate_size": 18944,
+            "max_position_embeddings": 128000,
+            "max_window_layers": 28,
+            "model_type": "qwen2_5_vl",
+            "num_attention_heads": 28,
+            "num_hidden_layers": 28,
+            "num_key_value_heads": 4,
+            "pad_token_id": 151643,
+            "rms_norm_eps": 1e-06,
+            "rope_scaling": {
+                "mrope_section": [
+                16,
+                24,
+                24
+                ],
+                "rope_type": "default",
+                "type": "default"
+            },
+            "rope_theta": 1000000.0,
+            "sliding_window": 32768,
+            "tie_word_embeddings": False,
+            "torch_dtype": "bfloat16",
+            "transformers_version": "4.49.0",
+            "use_cache": False,
+            "use_sliding_window": False,
+            "video_token_id": 151656,
+            "vision_config": {
+                "hidden_size": 1280,
+                "in_chans": 3,
+                "model_type": "qwen2_5_vl",
+                "spatial_patch_size": 14,
+                "tokens_per_second": 2,
+                "torch_dtype": "bfloat16"
+            },
+            "vision_end_token_id": 151653,
+            "vision_start_token_id": 151652,
+            "vision_token_id": 151654,
+            "vocab_size": 152064
+        })
+        self.model = Qwen2_5_VLForConditionalGeneration(model_config)
+        self.processor = None
+        
+        
+    def load_processor(self, path):
+        from .nexus_gen_ar_model import Qwen2_5_VLProcessor
+        self.processor = Qwen2_5_VLProcessor.from_pretrained(path)
+
+
+    @staticmethod
+    def state_dict_converter():
+        return NexusGenAutoregressiveModelStateDictConverter()
+
+    def bound_image(self, image, max_pixels=262640):
+        from qwen_vl_utils import smart_resize
+        resized_height, resized_width = smart_resize(
+            image.height,
+            image.width,
+            max_pixels=max_pixels,
+        )
+        return image.resize((resized_width, resized_height))
+
+    def get_editing_msg(self, instruction):
+        if '<image>' not in instruction:
+            instruction = '<image> ' + instruction
+        messages = [{"role":"user", "content":instruction}, {"role":"assistant", "content":"Here is the image: <image>"}]
+        return messages
+
+    def get_generation_msg(self, instruction):
+        instruction = "Generate an image according to the following description: {}".format(instruction)
+        messages = [{"role":"user", "content":instruction}, {"role":"assistant", "content":"Here is an image based on the description: <image>"}]
+        return messages
+
+    def forward(self, instruction, ref_image=None, num_img_tokens=81):
+        """
+        Generate target embeddings for the given instruction and reference image.
+        """
+        if ref_image is not None:
+            messages = self.get_editing_msg(instruction)
+            images = [self.bound_image(ref_image)] + [Image.new(mode='RGB', size=(252, 252), color=(255, 255, 255))]
+            output_image_embeddings = self.get_target_embeddings(images, messages, self.processor, self.model, num_img_tokens)
+        else:
+            messages = self.get_generation_msg(instruction)
+            images = [Image.new(mode='RGB', size=(252, 252), color=(255, 255, 255))]
+            output_image_embeddings = self.get_target_embeddings(images, messages, self.processor, self.model, num_img_tokens)
+
+        return output_image_embeddings
+
+    def get_target_embeddings(self, images, messages, processor, model, num_img_tokens=81):
+        text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=False)
+        text = text.replace('<image>', '<|vision_start|><|image_pad|><|vision_end|>')
+        inputs = processor(
+            text=[text],
+            images=images,
+            padding=True,
+            return_tensors="pt",
+        )
+        inputs = inputs.to(model.device)
+
+        input_embeds = model.model.embed_tokens(inputs['input_ids'])
+        image_embeds = model.visual(inputs['pixel_values'], grid_thw=inputs['image_grid_thw'])
+        ground_truth_image_embeds = image_embeds[-num_img_tokens:]
+        input_image_embeds = image_embeds[:-num_img_tokens]
+
+        image_mask = inputs['input_ids'] == model.config.image_token_id
+        indices = image_mask.cumsum(dim=1)
+        input_image_mask = torch.logical_and(indices <= (image_embeds.shape[0] - ground_truth_image_embeds.shape[0]), image_mask)
+        gt_image_mask = torch.logical_and(image_mask, ~input_image_mask)
+        input_image_mask = input_image_mask.unsqueeze(-1).expand_as(input_embeds)
+        input_embeds = input_embeds.masked_scatter(input_image_mask, input_image_embeds)
+
+        image_prefill_embeds = model.image_prefill_embeds(
+            torch.arange(81, device=model.device).long()
+        )
+        input_embeds = input_embeds.masked_scatter(gt_image_mask.unsqueeze(-1).expand_as(input_embeds), image_prefill_embeds)
+
+        position_ids, _ = model.get_rope_index(
+            inputs['input_ids'],
+            inputs['image_grid_thw'],
+            attention_mask=inputs['attention_mask'])
+        position_ids = position_ids.contiguous()
+        outputs = model(inputs_embeds=input_embeds, position_ids=position_ids, attention_mask=inputs['attention_mask'], return_dict=True)
+        output_image_embeddings = outputs.image_embeddings[:, :-1, :]
+        output_image_embeddings = output_image_embeddings[gt_image_mask[:, 1:]]
+        return output_image_embeddings, input_image_embeds, inputs['image_grid_thw']
+
+
+class NexusGenAutoregressiveModelStateDictConverter:
+    def __init__(self):
+        pass
+
+    def from_civitai(self, state_dict):
+        state_dict = {"model." + key: value for key, value in state_dict.items()}
+        return state_dict

diffsynth/models/nexus_gen_projector.py

@@ -0,0 +1,417 @@
+import math
+import torch
+import torch.nn as nn
+from typing import Optional, Tuple
+
+
+
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section, unsqueeze_dim=1):
+    mrope_section = mrope_section * 2
+    cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
+        unsqueeze_dim
+    )
+    sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
+        unsqueeze_dim
+    )
+
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+class Qwen2_5_VLRotaryEmbedding(nn.Module):
+    def __init__(self, config, device=None):
+        super().__init__()
+        # BC: "rope_type" was originally "type"
+        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+        from transformers.modeling_rope_utils import _compute_default_rope_parameters
+        self.rope_init_fn = _compute_default_rope_parameters
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+        self.original_inv_freq = self.inv_freq
+
+
+    def _dynamic_frequency_update(self, position_ids, device):
+        """
+        dynamic RoPE layers should recompute `inv_freq` in the following situations:
+        1 - growing beyond the cached sequence length (allow scaling)
+        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
+        """
+        seq_len = torch.max(position_ids) + 1
+        if seq_len > self.max_seq_len_cached:  # growth
+            inv_freq, self.attention_scaling = self.rope_init_fn(
+                self.config, device, seq_len=seq_len, **self.rope_kwargs
+            )
+            self.register_buffer("inv_freq", inv_freq, persistent=False)  # TODO joao: may break with compilation
+            self.max_seq_len_cached = seq_len
+
+        if seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len:  # reset
+            self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+            self.max_seq_len_cached = self.original_max_seq_len
+
+
+    @torch.no_grad()
+    def forward(self, x, position_ids):
+        if "dynamic" in self.rope_type:
+            self._dynamic_frequency_update(position_ids, device=x.device)
+
+        # Core RoPE block. In contrast to other models, Qwen2_5_VL has different position ids for the grids
+        # So we expand the inv_freq to shape (3, ...)
+        inv_freq_expanded = self.inv_freq[None, None, :, None].float().expand(3, position_ids.shape[1], -1, 1)
+        position_ids_expanded = position_ids[:, :, None, :].float()  # shape (3, bs, 1, positions)
+        # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
+        device_type = x.device.type
+        device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+        with torch.autocast(device_type=device_type, enabled=False):
+            freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(2, 3)
+            emb = torch.cat((freqs, freqs), dim=-1)
+            cos = emb.cos()
+            sin = emb.sin()
+
+        # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
+        cos = cos * self.attention_scaling
+        sin = sin * self.attention_scaling
+
+        return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+    """
+    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+    """
+    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+    if n_rep == 1:
+        return hidden_states
+    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+class Qwen2_5_VLAttention(nn.Module):
+    def __init__(self, config, layer_idx: Optional[int] = None):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.num_key_value_heads = config.num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.is_causal = True
+        self.attention_dropout = config.attention_dropout
+        self.rope_scaling = config.rope_scaling
+
+        if (self.head_dim * self.num_heads) != self.hidden_size:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+                f" and `num_heads`: {self.num_heads})."
+            )
+        self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True)
+        self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
+        self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
+        self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        bsz, q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2)
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_multimodal_rotary_pos_emb(
+            query_states, key_states, cos, sin, self.rope_scaling["mrope_section"]
+        )
+
+        # repeat k/v heads if n_kv_heads < n_heads
+        key_states = repeat_kv(key_states, self.num_key_value_groups)
+        value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+        # Fix precision issues in Qwen2-VL float16 inference
+        # Replace inf values with zeros in attention weights to prevent NaN propagation
+        if query_states.dtype == torch.float16:
+            attn_weights = torch.where(torch.isinf(attn_weights), torch.zeros_like(attn_weights), attn_weights)
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+            raise ValueError(
+                f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+                f" {attn_output.size()}"
+            )
+
+        attn_output = attn_output.transpose(1, 2).contiguous()
+        attn_output = attn_output.reshape(bsz, q_len, -1)
+
+        attn_output = self.o_proj(attn_output)
+
+        return attn_output
+
+
+class Qwen2MLP(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        from transformers.activations import ACT2FN
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.intermediate_size = config.intermediate_size
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+        self.act_fn = ACT2FN[config.hidden_act]
+
+    def forward(self, x):
+        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+class Qwen2RMSNorm(nn.Module):
+    def __init__(self, hidden_size, eps=1e-6):
+        """
+        Qwen2RMSNorm is equivalent to T5LayerNorm
+        """
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(hidden_size))
+        self.variance_epsilon = eps
+
+    def forward(self, hidden_states):
+        input_dtype = hidden_states.dtype
+        hidden_states = hidden_states.to(torch.float32)
+        variance = hidden_states.pow(2).mean(-1, keepdim=True)
+        hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+        return self.weight * hidden_states.to(input_dtype)
+
+    def extra_repr(self):
+        return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+class Qwen2_5_VLDecoderLayer(nn.Module):
+    def __init__(self, config, layer_idx):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+
+        self.self_attn = Qwen2_5_VLAttention(config, layer_idx)
+
+        self.mlp = Qwen2MLP(config)
+        self.input_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # necessary, but kept here for BC
+    ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+
+        residual = hidden_states
+
+        hidden_states = self.input_layernorm(hidden_states)
+
+        # Self Attention
+        hidden_states = self.self_attn(
+            hidden_states=hidden_states,
+            position_embeddings=position_embeddings,
+        )
+        hidden_states = residual + hidden_states
+
+        # Fully Connected
+        residual = hidden_states
+        hidden_states = self.post_attention_layernorm(hidden_states)
+        hidden_states = self.mlp(hidden_states)
+        hidden_states = residual + hidden_states
+
+        return hidden_states
+
+
+class NexusGenImageEmbeddingMerger(nn.Module):
+    def __init__(self, num_layers=1, out_channel=4096, expand_ratio=4, device='cpu'):
+        super().__init__()
+        from transformers import Qwen2_5_VLConfig
+        from transformers.activations import ACT2FN
+        config = Qwen2_5_VLConfig(**{
+            "_name_or_path": "DiffSynth-Studio/Nexus-GenV2",
+            "architectures": [
+                "Qwen2_5_VLForConditionalGeneration"
+            ],
+            "attention_dropout": 0.0,
+            "auto_map": {
+                "AutoConfig": "configuration_qwen2_5_vl.Qwen2_5_VLConfig",
+                "AutoModel": "modeling_qwen2_5_vl.Qwen2_5_VLModel",
+                "AutoModelForCausalLM": "modeling_qwen2_5_vl.Qwen2_5_VLForConditionalGeneration"
+            },
+            "bos_token_id": 151643,
+            "eos_token_id": 151645,
+            "hidden_act": "silu",
+            "hidden_size": 3584,
+            "image_token_id": 151655,
+            "initializer_range": 0.02,
+            "intermediate_size": 18944,
+            "max_position_embeddings": 128000,
+            "max_window_layers": 28,
+            "model_type": "qwen2_5_vl",
+            "num_attention_heads": 28,
+            "num_hidden_layers": 28,
+            "num_key_value_heads": 4,
+            "pad_token_id": 151643,
+            "rms_norm_eps": 1e-06,
+            "rope_scaling": {
+                "mrope_section": [
+                16,
+                24,
+                24
+                ],
+                "rope_type": "default",
+                "type": "default"
+            },
+            "rope_theta": 1000000.0,
+            "sliding_window": 32768,
+            "tie_word_embeddings": False,
+            "torch_dtype": "bfloat16",
+            "transformers_version": "4.49.0",
+            "use_cache": False,
+            "use_sliding_window": False,
+            "video_token_id": 151656,
+            "vision_config": {
+                "hidden_size": 1280,
+                "in_chans": 3,
+                "model_type": "qwen2_5_vl",
+                "spatial_patch_size": 14,
+                "tokens_per_second": 2,
+                "torch_dtype": "bfloat16"
+            },
+            "vision_end_token_id": 151653,
+            "vision_start_token_id": 151652,
+            "vision_token_id": 151654,
+            "vocab_size": 152064
+        })
+        self.config = config
+        self.num_layers = num_layers
+        self.layers = nn.ModuleList([Qwen2_5_VLDecoderLayer(config, layer_idx) for layer_idx in range(num_layers)])
+        self.projector = nn.Sequential(Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps),
+                                       nn.Linear(config.hidden_size, out_channel * expand_ratio),
+                                       Qwen2RMSNorm(out_channel * expand_ratio, eps=config.rms_norm_eps),
+                                       ACT2FN[config.hidden_act], nn.Linear(out_channel * expand_ratio, out_channel),
+                                       Qwen2RMSNorm(out_channel, eps=config.rms_norm_eps))
+        self.base_grid = torch.tensor([[1, 72, 72]], device=device)
+        self.rotary_emb = Qwen2_5_VLRotaryEmbedding(config=config, device=device)
+
+    def get_position_ids(self, image_grid_thw):
+        """
+        Generates position ids for the input embeddings grid.
+        modified from the qwen2_vl mrope.
+        """
+        batch_size = image_grid_thw.shape[0]
+        spatial_merge_size = self.config.vision_config.spatial_merge_size
+        t, h, w = (
+            image_grid_thw[0][0],
+            image_grid_thw[0][1],
+            image_grid_thw[0][2],
+        )
+        llm_grid_t, llm_grid_h, llm_grid_w = (
+            t.item(),
+            h.item() // spatial_merge_size,
+            w.item() // spatial_merge_size,
+        )
+        scale_h = self.base_grid[0][1].item() / h.item()
+        scale_w = self.base_grid[0][2].item() / w.item()
+
+        range_tensor = torch.arange(llm_grid_t).view(-1, 1)
+        expanded_range = range_tensor.expand(-1, llm_grid_h * llm_grid_w)
+        time_tensor = expanded_range * self.config.vision_config.tokens_per_second
+        t_index = time_tensor.long().flatten().to(image_grid_thw.device)
+        h_index = torch.arange(llm_grid_h).view(1, -1, 1).expand(llm_grid_t, -1, llm_grid_w).flatten().to(image_grid_thw.device) * scale_h
+        w_index = torch.arange(llm_grid_w).view(1, 1, -1).expand(llm_grid_t, llm_grid_h, -1).flatten().to(image_grid_thw.device) * scale_w
+        # 3, B, L
+        position_ids = torch.stack([t_index, h_index, w_index]).unsqueeze(0).repeat(batch_size, 1, 1).permute(1, 0, 2)
+        return position_ids
+
+    def forward(self, embeds, embeds_grid, ref_embeds=None, ref_embeds_grid=None):
+        position_ids = self.get_position_ids(embeds_grid)
+        hidden_states = embeds
+        if ref_embeds is not None:
+            position_ids_ref_embeds = self.get_position_ids(ref_embeds_grid)
+            position_ids = torch.cat((position_ids, position_ids_ref_embeds), dim=-1)
+            hidden_states = torch.cat((embeds, ref_embeds), dim=1)
+
+        position_embeddings = self.rotary_emb(hidden_states, position_ids)
+        for layer in self.layers:
+            hidden_states = layer(hidden_states, position_embeddings)
+
+        hidden_states = self.projector(hidden_states)
+        return hidden_states
+
+    @staticmethod
+    def state_dict_converter():
+        return NexusGenMergerStateDictConverter()
+
+
+class NexusGenMergerStateDictConverter:
+    def __init__(self):
+        pass
+
+    def from_diffusers(self, state_dict):
+        return state_dict
+    
+    def from_civitai(self, state_dict):
+        merger_state_dict = {key.replace("embedding_merger.", ""): value for key, value in state_dict.items() if key.startswith('embedding_merger.')}
+        return merger_state_dict
+
+
+class NexusGenAdapter(nn.Module):
+    """
+    Adapter for Nexus-Gen generation decoder.
+    """
+    def __init__(self, input_dim=3584, output_dim=4096):
+        super(NexusGenAdapter, self).__init__()
+        self.adapter = nn.Sequential(nn.Linear(input_dim, output_dim),
+                                     nn.LayerNorm(output_dim), nn.ReLU(),
+                                     nn.Linear(output_dim, output_dim),
+                                     nn.LayerNorm(output_dim))
+
+    def forward(self, x):
+        return self.adapter(x)
+
+    @staticmethod
+    def state_dict_converter():
+        return NexusGenAdapterStateDictConverter()
+
+
+class NexusGenAdapterStateDictConverter:
+    def __init__(self):
+        pass
+
+    def from_diffusers(self, state_dict):
+        return state_dict
+    
+    def from_civitai(self, state_dict):
+        adapter_state_dict = {key: value for key, value in state_dict.items() if key.startswith('adapter.')}
+        return adapter_state_dict

diffsynth/pipelines/flux_image.py

@@ -16,7 +16,7 @@ from ..models.flux_text_encoder_clip import FluxTextEncoderClip
 from ..models.flux_text_encoder_t5 import FluxTextEncoderT5
 from ..models.flux_vae import FluxVAEEncoder, FluxVAEDecoder
 from ..models.flux_value_control import MultiValueEncoder
-
+from ..core.vram.layers import AutoWrappedLinear
 
 class MultiControlNet(torch.nn.Module):
     def __init__(self, models: list[torch.nn.Module]):
@@ -104,7 +104,12 @@ class FluxImagePipeline(BasePipeline):
         self.lora_loader = FluxLoRALoader
 
     def enable_lora_magic(self):
-        pass
+        if self.lora_patcher is not None:
+            for name, module in self.dit.named_modules():
+                if isinstance(module, AutoWrappedLinear):
+                    merger_name = name.replace(".", "___")
+                    if merger_name in self.lora_patcher.model_dict:
+                        module.lora_merger = self.lora_patcher.model_dict[merger_name]
 
     @staticmethod
     def from_pretrained(

diffsynth/utils/state_dict_converters/flux_dit.py

@@ -1,4 +1,41 @@
+import torch
+import hashlib
+
+def convert_state_dict_keys_to_single_str(state_dict, with_shape=True):
+    keys = []
+    all_keys = sorted(list(state_dict))
+    
+    for key in all_keys:
+        value = state_dict[key]
+        if isinstance(key, str):
+            if isinstance(value, torch.Tensor):
+                if with_shape:
+                    shape = "_".join(map(str, list(value.shape)))
+                    keys.append(key + ":" + shape)
+                keys.append(key)
+            elif isinstance(value, dict):
+                keys.append(key + "|" + convert_state_dict_keys_to_single_str(value, with_shape=with_shape))
+    keys.sort()
+    keys_str = ",".join(keys)
+    return keys_str
+
+def hash_state_dict_keys(state_dict, with_shape=True):
+    keys_str = convert_state_dict_keys_to_single_str(state_dict, with_shape=with_shape)
+    keys_str = keys_str.encode(encoding="UTF-8")
+    return hashlib.md5(keys_str).hexdigest()
+
 def FluxDiTStateDictConverter(state_dict):
+    model_hash = hash_state_dict_keys(state_dict, with_shape=True)
+    
+    if model_hash in ["3e6c61b0f9471135fc9c6d6a98e98b6d", "63c969fd37cce769a90aa781fbff5f81"]:
+        dit_state_dict = {}
+        for key in state_dict:
+            if key.startswith('pipe.dit.'):
+                value = state_dict[key]
+                new_key = key.replace("pipe.dit.", "")
+                dit_state_dict[new_key] = value
+        return dit_state_dict
+
     rename_dict = {
         "time_in.in_layer.bias": "time_embedder.timestep_embedder.0.bias",
         "time_in.in_layer.weight": "time_embedder.timestep_embedder.0.weight",

diffsynth/utils/state_dict_converters/flux_ipadapter.py

@@ -0,0 +1,34 @@
+import torch
+
+def FluxIpAdapterStateDictConverter(state_dict):
+    state_dict_ = {}
+    
+    if "ip_adapter" in state_dict and isinstance(state_dict["ip_adapter"], dict):
+        for name, param in state_dict["ip_adapter"].items():
+            name_ = 'ipadapter_modules.' + name
+            state_dict_[name_] = param
+        
+        if "image_proj" in state_dict:
+            for name, param in state_dict["image_proj"].items():
+                name_ = "image_proj." + name
+                state_dict_[name_] = param
+        return state_dict_
+
+    for key, value in state_dict.items():
+        if key.startswith("image_proj."):
+            state_dict_[key] = value
+        elif key.startswith("ip_adapter."):
+            new_key = key.replace("ip_adapter.", "ipadapter_modules.")
+            state_dict_[new_key] = value
+        else:
+            pass
+            
+    return state_dict_
+
+
+def SiglipStateDictConverter(state_dict):
+    new_state_dict = {}
+    for key in state_dict:
+        if key.startswith("vision_model."):
+            new_state_dict[key] = state_dict[key] 
+    return new_state_dict

diffsynth/utils/state_dict_converters/nexus_gen.py

@@ -0,0 +1,8 @@
+import torch
+
+def NexusGenAutoregressiveModelStateDictConverter(state_dict):
+    new_state_dict = {}
+    for key in state_dict:
+        value = state_dict[key]
+        new_state_dict["model." + key] = value
+    return new_state_dict

diffsynth/utils/state_dict_converters/nexus_gen_projector.py

@@ -0,0 +1,17 @@
+import torch
+
+def NexusGenMergerStateDictConverter(state_dict):
+    merger_state_dict = {}
+    for key in state_dict:
+        if key.startswith('embedding_merger.'):
+            value = state_dict[key]
+            new_key = key.replace("embedding_merger.", "")
+            merger_state_dict[new_key] = value
+    return merger_state_dict
+
+def NexusGenAdapterStateDictConverter(state_dict):
+    adapter_state_dict = {}
+    for key in state_dict:
+        if key.startswith('adapter.'):
+            adapter_state_dict[key] = state_dict[key]
+    return adapter_state_dict