Merge pull request #715 from modelscope/nexusgen-eligen

NexusGen and EliGen

diffsynth/configs/model_config.py

@@ -69,6 +69,8 @@ from ..models.flux_value_control import SingleValueEncoder
 from ..lora.flux_lora import FluxLoraPatcher
 from ..models.flux_lora_encoder import FluxLoRAEncoder
 
+from ..models.nexus_gen_projector import NexusGenAdapter, NexusGenImageEmbeddingMerger
+from ..models.nexus_gen import NexusGenAutoregressiveModel
 
 model_loader_configs = [
     # These configs are provided for detecting model type automatically.
@@ -155,6 +157,9 @@ model_loader_configs = [
     (None, "d30fb9e02b1dbf4e509142f05cf7dd50", ["flux_dit", "step1x_connector"], [FluxDiT, Qwen2Connector], "civitai"),
     (None, "30143afb2dea73d1ac580e0787628f8c", ["flux_lora_patcher"], [FluxLoraPatcher], "civitai"),
     (None, "77c2e4dd2440269eb33bfaa0d004f6ab", ["flux_lora_encoder"], [FluxLoRAEncoder], "civitai"),
+    (None, "3e6c61b0f9471135fc9c6d6a98e98b6d", ["flux_dit", "nexus_gen_generation_adapter"], [FluxDiT, NexusGenAdapter], "civitai"),
+    (None, "63c969fd37cce769a90aa781fbff5f81", ["flux_dit", "nexus_gen_editing_adapter"], [FluxDiT, NexusGenImageEmbeddingMerger], "civitai"),
+    (None, "2bd19e845116e4f875a0a048e27fc219", ["nexus_gen_llm"], [NexusGenAutoregressiveModel], "civitai"),
 ]
 huggingface_model_loader_configs = [
     # These configs are provided for detecting model type automatically.

diffsynth/models/flux_dit.py

@@ -2,7 +2,7 @@ import torch
 from .sd3_dit import TimestepEmbeddings, AdaLayerNorm, RMSNorm
 from einops import rearrange
 from .tiler import TileWorker
-from .utils import init_weights_on_device
+from .utils import init_weights_on_device, hash_state_dict_keys
 
 def interact_with_ipadapter(hidden_states, q, ip_k, ip_v, scale=1.0):
     batch_size, num_tokens = hidden_states.shape[0:2]
@@ -662,6 +662,9 @@ class FluxDiTStateDictConverter:
         return state_dict_
 
     def from_civitai(self, state_dict):
+        if hash_state_dict_keys(state_dict, with_shape=True) in ["3e6c61b0f9471135fc9c6d6a98e98b6d", "63c969fd37cce769a90aa781fbff5f81"]:
+            dit_state_dict = {key.replace("pipe.dit.", ""): value for key, value in state_dict.items() if key.startswith('pipe.dit.')}
+            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/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_new.py

@@ -22,6 +22,8 @@ from ..models.flux_value_control import MultiValueEncoder
 from ..models.flux_infiniteyou import InfiniteYouImageProjector
 from ..models.flux_lora_encoder import FluxLoRAEncoder, LoRALayerBlock
 from ..models.tiler import FastTileWorker
+from ..models.nexus_gen import NexusGenAutoregressiveModel
+from ..models.nexus_gen_projector import NexusGenAdapter, NexusGenImageEmbeddingMerger
 from ..utils import BasePipeline, ModelConfig, PipelineUnitRunner, PipelineUnit
 from ..lora.flux_lora import FluxLoRALoader, FluxLoraPatcher, FluxLoRAFuser
 
@@ -94,6 +96,9 @@ class FluxImagePipeline(BasePipeline):
         self.ipadapter_image_encoder = None
         self.qwenvl = None
         self.step1x_connector: Qwen2Connector = None
+        self.nexus_gen: NexusGenAutoregressiveModel = None
+        self.nexus_gen_generation_adapter: NexusGenAdapter = None
+        self.nexus_gen_editing_adapter: NexusGenImageEmbeddingMerger = None
         self.value_controller: MultiValueEncoder = None
         self.infinityou_processor: InfinitYou = None
         self.image_proj_model: InfiniteYouImageProjector = None
@@ -113,6 +118,7 @@ class FluxImagePipeline(BasePipeline):
             FluxImageUnit_ControlNet(),
             FluxImageUnit_IPAdapter(),
             FluxImageUnit_EntityControl(),
+            FluxImageUnit_NexusGen(),
             FluxImageUnit_TeaCache(),
             FluxImageUnit_Flex(),
             FluxImageUnit_Step1x(),
@@ -369,6 +375,7 @@ class FluxImagePipeline(BasePipeline):
         torch_dtype: torch.dtype = torch.bfloat16,
         device: Union[str, torch.device] = "cuda",
         model_configs: list[ModelConfig] = [],
+        nexus_gen_processor_config: ModelConfig = ModelConfig(model_id="DiffSynth-Studio/Nexus-GenV2", origin_file_pattern="processor/"),
     ):
         # Download and load models
         model_manager = ModelManager()
@@ -397,6 +404,12 @@ class FluxImagePipeline(BasePipeline):
             pipe.infinityou_processor = InfinitYou(device=device)
         pipe.lora_patcher = model_manager.fetch_model("flux_lora_patcher")
         pipe.lora_encoder = model_manager.fetch_model("flux_lora_encoder")
+        pipe.nexus_gen = model_manager.fetch_model("nexus_gen_llm")
+        pipe.nexus_gen_generation_adapter = model_manager.fetch_model("nexus_gen_generation_adapter")
+        pipe.nexus_gen_editing_adapter = model_manager.fetch_model("nexus_gen_editing_adapter")
+        if nexus_gen_processor_config is not None and pipe.nexus_gen is not None:
+            nexus_gen_processor_config.download_if_necessary()
+            pipe.nexus_gen.load_processor(nexus_gen_processor_config.path)
         
         # ControlNet
         controlnets = []
@@ -468,6 +481,8 @@ class FluxImagePipeline(BasePipeline):
         value_controller_inputs: Union[list[float], float] = None,
         # Step1x
         step1x_reference_image: Image.Image = None,
+        # NexusGen
+        nexus_gen_reference_image: Image.Image = None,
         # LoRA Encoder
         lora_encoder_inputs: Union[list[ModelConfig], ModelConfig, str] = None,
         lora_encoder_scale: float = 1.0,
@@ -504,6 +519,7 @@ class FluxImagePipeline(BasePipeline):
             "flex_inpaint_image": flex_inpaint_image, "flex_inpaint_mask": flex_inpaint_mask, "flex_control_image": flex_control_image, "flex_control_strength": flex_control_strength, "flex_control_stop": flex_control_stop,
             "value_controller_inputs": value_controller_inputs,
             "step1x_reference_image": step1x_reference_image,
+            "nexus_gen_reference_image": nexus_gen_reference_image,
             "lora_encoder_inputs": lora_encoder_inputs, "lora_encoder_scale": lora_encoder_scale,
             "tea_cache_l1_thresh": tea_cache_l1_thresh,
             "tiled": tiled, "tile_size": tile_size, "tile_stride": tile_stride,
@@ -755,15 +771,70 @@ class FluxImageUnit_EntityControl(PipelineUnit):
         if eligen_entity_prompts is None or eligen_entity_masks is None:
             return inputs_shared, inputs_posi, inputs_nega
         pipe.load_models_to_device(self.onload_model_names)
+        eligen_enable_on_negative = inputs_shared.get("eligen_enable_on_negative", False)
         eligen_kwargs_posi, eligen_kwargs_nega = self.prepare_eligen(pipe, inputs_nega,
             eligen_entity_prompts, eligen_entity_masks, inputs_shared["width"], inputs_shared["height"], 
-            inputs_shared["t5_sequence_length"], inputs_shared["eligen_enable_on_negative"], inputs_shared["cfg_scale"])
+            inputs_shared["t5_sequence_length"], eligen_enable_on_negative, inputs_shared["cfg_scale"])
         inputs_posi.update(eligen_kwargs_posi)
         if inputs_shared.get("cfg_scale", 1.0) != 1.0:
             inputs_nega.update(eligen_kwargs_nega)
         return inputs_shared, inputs_posi, inputs_nega
 
 
+class FluxImageUnit_NexusGen(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            take_over=True,
+            onload_model_names=("nexus_gen", "nexus_gen_generation_adapter", "nexus_gen_editing_adapter"),
+        )
+
+    def process(self, pipe: FluxImagePipeline, inputs_shared, inputs_posi, inputs_nega):
+        if pipe.nexus_gen is None:
+            return inputs_shared, inputs_posi, inputs_nega
+        pipe.load_models_to_device(self.onload_model_names)
+        if inputs_shared.get("nexus_gen_reference_image", None) is None:
+            assert pipe.nexus_gen_generation_adapter is not None, "NexusGen requires a generation adapter to be set."
+            embed = pipe.nexus_gen(inputs_posi["prompt"])[0].unsqueeze(0)
+            inputs_posi["prompt_emb"] = pipe.nexus_gen_generation_adapter(embed)
+            inputs_posi['text_ids'] = torch.zeros(embed.shape[0], embed.shape[1], 3).to(device=pipe.device, dtype=pipe.torch_dtype)
+        else:
+            assert pipe.nexus_gen_editing_adapter is not None, "NexusGen requires an editing adapter to be set."
+            embed, ref_embed, grids = pipe.nexus_gen(inputs_posi["prompt"], inputs_shared["nexus_gen_reference_image"])
+            embeds_grid = grids[1:2].to(device=pipe.device, dtype=torch.long)
+            ref_embeds_grid = grids[0:1].to(device=pipe.device, dtype=torch.long)
+
+            inputs_posi["prompt_emb"] = pipe.nexus_gen_editing_adapter(embed.unsqueeze(0), embeds_grid, ref_embed.unsqueeze(0), ref_embeds_grid)
+            inputs_posi["text_ids"] = self.get_editing_text_ids(
+                inputs_shared["latents"],
+                embeds_grid[0][1].item(), embeds_grid[0][2].item(),
+                ref_embeds_grid[0][1].item(), ref_embeds_grid[0][2].item(),
+                )
+        return inputs_shared, inputs_posi, inputs_nega
+
+
+    def get_editing_text_ids(self, latents, target_embed_height, target_embed_width, ref_embed_height, ref_embed_width):
+        # prepare text ids for target and reference embeddings
+        batch_size, height, width = latents.shape[0], target_embed_height, target_embed_width
+        embed_ids = torch.zeros(height // 2, width // 2, 3)
+        scale_factor_height, scale_factor_width = latents.shape[-2] / height, latents.shape[-1] / width
+        embed_ids[..., 1] = embed_ids[..., 1] + torch.arange(height // 2)[:, None] * scale_factor_height
+        embed_ids[..., 2] = embed_ids[..., 2] + torch.arange(width // 2)[None, :] * scale_factor_width
+        embed_ids = embed_ids[None, :].repeat(batch_size, 1, 1, 1).reshape(batch_size, height // 2 * width // 2, 3)
+        embed_text_ids = embed_ids.to(device=latents.device, dtype=latents.dtype)
+
+        batch_size, height, width = latents.shape[0], ref_embed_height, ref_embed_width
+        ref_embed_ids = torch.zeros(height // 2, width // 2, 3)
+        scale_factor_height, scale_factor_width = latents.shape[-2] / height, latents.shape[-1] / width
+        ref_embed_ids[..., 0] = ref_embed_ids[..., 0] + 1.0
+        ref_embed_ids[..., 1] = ref_embed_ids[..., 1] + torch.arange(height // 2)[:, None] * scale_factor_height
+        ref_embed_ids[..., 2] = ref_embed_ids[..., 2] + torch.arange(width // 2)[None, :] * scale_factor_width
+        ref_embed_ids = ref_embed_ids[None, :].repeat(batch_size, 1, 1, 1).reshape(batch_size, height // 2 * width // 2, 3)
+        ref_embed_text_ids = ref_embed_ids.to(device=latents.device, dtype=latents.dtype)
+
+        text_ids = torch.cat([embed_text_ids, ref_embed_text_ids], dim=1)
+        return text_ids
+
+
 class FluxImageUnit_Step1x(PipelineUnit):
     def __init__(self):
         super().__init__(take_over=True,onload_model_names=("qwenvl","vae_encoder"))

diffsynth/trainers/utils.py

@@ -120,8 +120,12 @@ class ImageDataset(torch.utils.data.Dataset):
         data = self.data[data_id % len(self.data)].copy()
         for key in self.data_file_keys:
             if key in data:
-                path = os.path.join(self.base_path, data[key])
-                data[key] = self.load_data(path)
+                if isinstance(data[key], list):
+                    path = [os.path.join(self.base_path, p) for p in data[key]]
+                    data[key] = [self.load_data(p) for p in path]
+                else:
+                    path = os.path.join(self.base_path, data[key])
+                    data[key] = self.load_data(path)
                 if data[key] is None:
                     warnings.warn(f"cannot load file {data[key]}.")
                     return None