multi-node

.github/workflows/publish.yaml

@@ -20,7 +20,7 @@ jobs:
         with:
           python-version: '3.10'
       - name: Install wheel
-        run: pip install wheel && pip install -r requirements.txt
+        run: pip install wheel==0.44.0 && pip install -r requirements.txt
       - name: Build DiffSynth
         run: python setup.py sdist bdist_wheel
       - name: Publish package to PyPI

diffsynth/data/image_pulse.py

@@ -0,0 +1,125 @@
+import torch, os, json, torchvision
+from PIL import Image
+from torchvision.transforms import v2
+
+
+
+class SingleTaskDataset(torch.utils.data.Dataset):
+    def __init__(self, base_path, keys=(("image_1", "image_2", "editing_instruction"), ("image_2", "image_1", "reverse_editing_instruction")), height=1024, width=1024, random=True, steps_per_epoch=1000, metadata_path=None):
+        self.base_path = base_path
+        self.keys = keys
+        self.metadata = []
+        self.bad_data = []
+        self.height = height
+        self.width = width
+        self.random = random
+        self.steps_per_epoch = steps_per_epoch
+        self.image_process = v2.Compose([
+            v2.CenterCrop(size=(height, width)),
+            v2.ToImage(),
+            v2.ToDtype(torch.float32, scale=True),
+            v2.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+        ])
+        if metadata_path is None:
+            self.search_for_data("", report_data_log=True)
+            self.report_data_log()
+        else:
+            with open(metadata_path, "r", encoding="utf-8-sig") as f:
+                self.metadata = json.load(f)
+
+
+    def report_data_log(self):
+        print(f"{len(self.metadata)} valid data, {len(self.bad_data)} invalid data.")
+
+
+    def dump_metadata(self, path):
+        with open(path, "w", encoding="utf-8") as f:
+            json.dump(self.metadata, f, ensure_ascii=False, indent=4)
+        
+        
+    def parse_json_file(self, absolute_path, relative_path):
+        data_list = []
+        with open(absolute_path, "r") as f:
+            metadata = json.load(f)
+            for image_1, image_2, instruction in self.keys:
+                image_1 = os.path.join(relative_path, metadata[image_1])
+                image_2 = os.path.join(relative_path, metadata[image_2])
+                instruction = metadata[instruction]
+                data_list.append((image_1, image_2, instruction))
+        return data_list
+    
+        
+    def search_for_data(self, path, report_data_log=False):
+        now_path = os.path.join(self.base_path, path)
+        if os.path.isfile(now_path) and path.endswith(".json"):
+            try:
+                data_list = self.parse_json_file(now_path, os.path.dirname(path))
+                self.metadata.extend(data_list)
+            except:
+                self.bad_data.append(now_path)
+        elif os.path.isdir(now_path):
+            for sub_path in os.listdir(now_path):
+                self.search_for_data(os.path.join(path, sub_path))
+                if report_data_log and os.path.isdir(os.path.join(self.base_path, path, sub_path)):
+                    self.report_data_log()
+                
+                
+    def load_image(self, image_path):
+        image_path = os.path.join(self.base_path, image_path)
+        image = Image.open(image_path).convert("RGB")
+        width, height = image.size
+        scale = max(self.width / width, self.height / height)
+        image = torchvision.transforms.functional.resize(
+            image,
+            (round(height*scale), round(width*scale)),
+            interpolation=torchvision.transforms.InterpolationMode.BILINEAR
+        )
+        image = self.image_process(image)
+        return image
+                
+                
+    def load_data(self, data_id):
+        image_1, image_2, instruction = self.metadata[data_id]
+        image_1 = self.load_image(image_1)
+        image_2 = self.load_image(image_2)
+        return {"image_1": image_1, "image_2": image_2, "instruction": instruction}
+        
+        
+    def __getitem__(self, data_id):
+        if self.random:
+            while True:
+                try:
+                    data_id = (torch.randint(0, len(self.metadata), (1,))[0] + data_id) % len(self.metadata)
+                    data = self.load_data(data_id)
+                    return data
+                except:
+                    continue
+        else:
+            return self.load_data(data_id)
+
+
+    def __len__(self):
+        return self.steps_per_epoch if self.random else len(self.metadata)
+    
+    
+    
+class MultiTaskDataset(torch.utils.data.Dataset):
+    def __init__(self, dataset_list, dataset_weight, steps_per_epoch=1000):
+        self.dataset_list = dataset_list
+        self.dataset_weight = torch.tensor(dataset_weight, dtype=torch.float)
+        self.steps_per_epoch = steps_per_epoch
+
+        
+    def __getitem__(self, data_id):
+        while True:
+            try:
+                dataset_id = torch.multinomial(self.dataset_weight, 1).tolist()[0]
+                data_id = torch.randint(0, len(self.dataset_list[dataset_id]), (1,))[0]
+                data = self.dataset_list[dataset_id][data_id]
+                return data
+            except:
+                continue
+
+
+    def __len__(self):
+        return self.steps_per_epoch

diffsynth/models/flux_dit.py

@@ -20,10 +20,11 @@ class RoPEEmbedding(torch.nn.Module):
         self.axes_dim = axes_dim
 
 
-    def rope(self, pos: torch.Tensor, dim: int, theta: int) -> torch.Tensor:
+    def rope(self, pos: torch.Tensor, dim: int, theta: int, device="cpu") -> torch.Tensor:
         assert dim % 2 == 0, "The dimension must be even."
 
         scale = torch.arange(0, dim, 2, dtype=torch.float64, device=pos.device) / dim
+        scale = scale.to(device)
         omega = 1.0 / (theta**scale)
 
         batch_size, seq_length = pos.shape
@@ -36,9 +37,9 @@ class RoPEEmbedding(torch.nn.Module):
         return out.float()
 
 
-    def forward(self, ids):
+    def forward(self, ids, device="cpu"):
         n_axes = ids.shape[-1]
-        emb = torch.cat([self.rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(n_axes)], dim=-3)
+        emb = torch.cat([self.rope(ids[..., i], self.axes_dim[i], self.theta, device) for i in range(n_axes)], dim=-3)
         return emb.unsqueeze(1)
 
 

diffsynth/models/flux_reference_embedder.py

@@ -0,0 +1,31 @@
+from .sd3_dit import TimestepEmbeddings
+from .flux_dit import RoPEEmbedding
+import torch
+from einops import repeat
+
+
+class FluxReferenceEmbedder(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.pos_embedder = RoPEEmbedding(3072, 10000, [16, 56, 56])
+        self.idx_embedder = TimestepEmbeddings(256, 256)
+        self.proj = torch.nn.Linear(3072, 3072)
+        
+    def forward(self, image_ids, idx, dtype, device):
+        pos_emb = self.pos_embedder(image_ids, device=device)
+        idx_emb = self.idx_embedder(idx, dtype=dtype).to(device)
+        length = pos_emb.shape[2]
+        pos_emb = repeat(pos_emb, "B N L C H W -> 1 N (B L) C H W")
+        idx_emb = repeat(idx_emb, "B (C H W) -> 1 1 (B L) C H W", C=64, H=2, W=2, L=length)
+        image_rotary_emb = pos_emb + idx_emb
+        return image_rotary_emb
+    
+    def init(self):
+        self.idx_embedder.timestep_embedder[-1].load_state_dict({
+            "weight": torch.zeros((256, 256)),
+            "bias": torch.zeros((256,))
+        }),
+        self.proj.load_state_dict({
+            "weight": torch.eye(3072),
+            "bias": torch.zeros((3072,))
+        })

diffsynth/pipelines/flux_image.py

@@ -1,10 +1,12 @@
 from ..models import ModelManager, FluxDiT, SD3TextEncoder1, FluxTextEncoder2, FluxVAEDecoder, FluxVAEEncoder, FluxIpAdapter
 from ..controlnets import FluxMultiControlNetManager, ControlNetUnit, ControlNetConfigUnit, Annotator
+from ..models.flux_reference_embedder import FluxReferenceEmbedder
 from ..prompters import FluxPrompter
 from ..schedulers import FlowMatchScheduler
 from .base import BasePipeline
 from typing import List
 import torch
+from einops import rearrange
 from tqdm import tqdm
 import numpy as np
 from PIL import Image
@@ -32,6 +34,7 @@ class FluxImagePipeline(BasePipeline):
         self.ipadapter: FluxIpAdapter = None
         self.ipadapter_image_encoder: SiglipVisionModel = None
         self.infinityou_processor: InfinitYou = None
+        self.reference_embedder: FluxReferenceEmbedder = None
         self.model_names = ['text_encoder_1', 'text_encoder_2', 'dit', 'vae_decoder', 'vae_encoder', 'controlnet', 'ipadapter', 'ipadapter_image_encoder']
 
 
@@ -360,6 +363,20 @@ class FluxImagePipeline(BasePipeline):
             return self.infinityou_processor.prepare_infinite_you(self.image_proj_model, id_image, controlnet_image, infinityou_guidance, height, width)
         else:
             return {}, controlnet_image
+        
+        
+    def prepare_reference_images(self, reference_images, tiled=False, tile_size=64, tile_stride=32):
+        if reference_images is not None:
+            hidden_states_ref = []
+            for reference_image in reference_images:
+                self.load_models_to_device(['vae_encoder'])
+                reference_image = self.preprocess_image(reference_image).to(device=self.device, dtype=self.torch_dtype)
+                latents = self.encode_image(reference_image, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)
+                hidden_states_ref.append(latents)
+            hidden_states_ref = torch.concat(hidden_states_ref, dim=0)
+            return {"hidden_states_ref": hidden_states_ref}
+        else:
+            return {"hidden_states_ref": None}
 
 
     @torch.no_grad()
@@ -398,6 +415,8 @@ class FluxImagePipeline(BasePipeline):
         # InfiniteYou
         infinityou_id_image=None,
         infinityou_guidance=1.0,
+        # Reference images
+        reference_images=None,
         # TeaCache
         tea_cache_l1_thresh=None,
         # Tile
@@ -436,6 +455,9 @@ class FluxImagePipeline(BasePipeline):
 
         # ControlNets
         controlnet_kwargs_posi, controlnet_kwargs_nega, local_controlnet_kwargs = self.prepare_controlnet(controlnet_image, masks, controlnet_inpaint_mask, tiler_kwargs, enable_controlnet_on_negative)
+        
+        # Reference images
+        reference_kwargs = self.prepare_reference_images(reference_images, **tiler_kwargs)
 
         # TeaCache
         tea_cache_kwargs = {"tea_cache": TeaCache(num_inference_steps, rel_l1_thresh=tea_cache_l1_thresh) if tea_cache_l1_thresh is not None else None}
@@ -447,9 +469,9 @@ class FluxImagePipeline(BasePipeline):
 
             # Positive side
             inference_callback = lambda prompt_emb_posi, controlnet_kwargs: lets_dance_flux(
-                dit=self.dit, controlnet=self.controlnet,
+                dit=self.dit, controlnet=self.controlnet, reference_embedder=self.reference_embedder,
                 hidden_states=latents, timestep=timestep,
-                **prompt_emb_posi, **tiler_kwargs, **extra_input, **controlnet_kwargs, **ipadapter_kwargs_list_posi, **eligen_kwargs_posi, **tea_cache_kwargs, **infiniteyou_kwargs
+                **prompt_emb_posi, **tiler_kwargs, **extra_input, **controlnet_kwargs, **ipadapter_kwargs_list_posi, **eligen_kwargs_posi, **tea_cache_kwargs, **infiniteyou_kwargs, **reference_kwargs,
             )
             noise_pred_posi = self.control_noise_via_local_prompts(
                 prompt_emb_posi, prompt_emb_locals, masks, mask_scales, inference_callback,
@@ -464,9 +486,9 @@ class FluxImagePipeline(BasePipeline):
             if cfg_scale != 1.0:
                 # Negative side
                 noise_pred_nega = lets_dance_flux(
-                    dit=self.dit, controlnet=self.controlnet,
+                    dit=self.dit, controlnet=self.controlnet, reference_embedder=self.reference_embedder,
                     hidden_states=latents, timestep=timestep,
-                    **prompt_emb_nega, **tiler_kwargs, **extra_input, **controlnet_kwargs_nega, **ipadapter_kwargs_list_nega, **eligen_kwargs_nega, **infiniteyou_kwargs,
+                    **prompt_emb_nega, **tiler_kwargs, **extra_input, **controlnet_kwargs_nega, **ipadapter_kwargs_list_nega, **eligen_kwargs_nega, **infiniteyou_kwargs, **reference_kwargs,
                 )
                 noise_pred = noise_pred_nega + cfg_scale * (noise_pred_posi - noise_pred_nega)
             else:
@@ -586,6 +608,7 @@ class TeaCache:
 def lets_dance_flux(
     dit: FluxDiT,
     controlnet: FluxMultiControlNetManager = None,
+    reference_embedder: FluxReferenceEmbedder = None,
     hidden_states=None,
     timestep=None,
     prompt_emb=None,
@@ -594,6 +617,7 @@ def lets_dance_flux(
     text_ids=None,
     image_ids=None,
     controlnet_frames=None,
+    hidden_states_ref=None,
     tiled=False,
     tile_size=128,
     tile_stride=64,
@@ -603,6 +627,7 @@ def lets_dance_flux(
     id_emb=None,
     infinityou_guidance=None,
     tea_cache: TeaCache = None,
+    use_gradient_checkpointing=False,
     **kwargs
 ):
     if tiled:
@@ -669,28 +694,55 @@ def lets_dance_flux(
         prompt_emb, image_rotary_emb, attention_mask = dit.process_entity_masks(hidden_states, prompt_emb, entity_prompt_emb, entity_masks, text_ids, image_ids)
     else:
         prompt_emb = dit.context_embedder(prompt_emb)
-        image_rotary_emb = dit.pos_embedder(torch.cat((text_ids, image_ids), dim=1))
+        image_rotary_emb = dit.pos_embedder(torch.cat((text_ids, image_ids), dim=1), device=hidden_states.device)
         attention_mask = None
+        
+    # Reference images
+    if hidden_states_ref is not None:
+        # RoPE
+        image_ids_ref = dit.prepare_image_ids(hidden_states_ref)
+        idx = torch.arange(0, image_ids_ref.shape[0]).to(dtype=hidden_states.dtype, device=hidden_states.device) * 100
+        image_rotary_emb_ref = reference_embedder(image_ids_ref, idx, dtype=hidden_states.dtype, device=hidden_states.device)
+        image_rotary_emb = torch.cat((image_rotary_emb, image_rotary_emb_ref), dim=2)
+        # hidden_states
+        original_hidden_states_length = hidden_states.shape[1]
+        hidden_states_ref = dit.patchify(hidden_states_ref)
+        hidden_states_ref = dit.x_embedder(hidden_states_ref)
+        hidden_states_ref = rearrange(hidden_states_ref, "B L C -> 1 (B L) C")
+        hidden_states_ref = reference_embedder.proj(hidden_states_ref)
+        hidden_states = torch.cat((hidden_states, hidden_states_ref), dim=1)
 
     # TeaCache
     if tea_cache is not None:
         tea_cache_update = tea_cache.check(dit, hidden_states, conditioning)
     else:
         tea_cache_update = False
+        
+    def create_custom_forward(module):
+        def custom_forward(*inputs):
+            return module(*inputs)
+        return custom_forward
 
     if tea_cache_update:
         hidden_states = tea_cache.update(hidden_states)
     else:
         # Joint Blocks
         for block_id, block in enumerate(dit.blocks):
-            hidden_states, prompt_emb = block(
-                hidden_states,
-                prompt_emb,
-                conditioning,
-                image_rotary_emb,
-                attention_mask,
-                ipadapter_kwargs_list=ipadapter_kwargs_list.get(block_id, None)
-            )
+            if use_gradient_checkpointing:
+                hidden_states, prompt_emb = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states, prompt_emb, conditioning, image_rotary_emb, attention_mask, ipadapter_kwargs_list.get(block_id, None),
+                    use_reentrant=False,
+                )
+            else:
+                hidden_states, prompt_emb = block(
+                    hidden_states,
+                    prompt_emb,
+                    conditioning,
+                    image_rotary_emb,
+                    attention_mask,
+                    ipadapter_kwargs_list=ipadapter_kwargs_list.get(block_id, None)
+                )
             # ControlNet
             if controlnet is not None and controlnet_frames is not None:
                 hidden_states = hidden_states + controlnet_res_stack[block_id]
@@ -699,14 +751,21 @@ def lets_dance_flux(
         hidden_states = torch.cat([prompt_emb, hidden_states], dim=1)
         num_joint_blocks = len(dit.blocks)
         for block_id, block in enumerate(dit.single_blocks):
-            hidden_states, prompt_emb = block(
-                hidden_states,
-                prompt_emb,
-                conditioning,
-                image_rotary_emb,
-                attention_mask,
-                ipadapter_kwargs_list=ipadapter_kwargs_list.get(block_id + num_joint_blocks, None)
-            )
+            if use_gradient_checkpointing:
+                hidden_states, prompt_emb = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states, prompt_emb, conditioning, image_rotary_emb, attention_mask, ipadapter_kwargs_list.get(block_id + num_joint_blocks, None),
+                    use_reentrant=False,
+                )
+            else:
+                hidden_states, prompt_emb = block(
+                    hidden_states,
+                    prompt_emb,
+                    conditioning,
+                    image_rotary_emb,
+                    attention_mask,
+                    ipadapter_kwargs_list=ipadapter_kwargs_list.get(block_id + num_joint_blocks, None)
+                )
             # ControlNet
             if controlnet is not None and controlnet_frames is not None:
                 hidden_states[:, prompt_emb.shape[1]:] = hidden_states[:, prompt_emb.shape[1]:] + controlnet_single_res_stack[block_id]
@@ -715,6 +774,8 @@ def lets_dance_flux(
         if tea_cache is not None:
             tea_cache.store(hidden_states)
 
+    if hidden_states_ref is not None:
+        hidden_states = hidden_states[:, :original_hidden_states_length]
     hidden_states = dit.final_norm_out(hidden_states, conditioning)
     hidden_states = dit.final_proj_out(hidden_states)
     hidden_states = dit.unpatchify(hidden_states, height, width)

setup.py

@@ -14,7 +14,7 @@ else:
 
 setup(
     name="diffsynth",
-    version="1.1.4",
+    version="1.1.7",
     description="Enjoy the magic of Diffusion models!",
     author="Artiprocher",
     packages=find_packages(),

train_flux_reference.py

@@ -0,0 +1,241 @@
+from diffsynth import ModelManager, FluxImagePipeline
+from diffsynth.trainers.text_to_image import LightningModelForT2ILoRA, add_general_parsers, launch_training_task
+from diffsynth.models.lora import FluxLoRAConverter
+import torch, os, argparse
+import lightning as pl
+from diffsynth.data.image_pulse import SingleTaskDataset, MultiTaskDataset
+from diffsynth.pipelines.flux_image import lets_dance_flux
+from diffsynth.models.flux_reference_embedder import FluxReferenceEmbedder
+os.environ["TOKENIZERS_PARALLELISM"] = "True"
+
+
+class LightningModel(LightningModelForT2ILoRA):
+    def __init__(
+        self,
+        torch_dtype=torch.float16, pretrained_weights=[], preset_lora_path=None,
+        learning_rate=1e-4, use_gradient_checkpointing=True,
+        lora_rank=4, lora_alpha=4, lora_target_modules="to_q,to_k,to_v,to_out", init_lora_weights="kaiming", pretrained_lora_path=None,
+        state_dict_converter=None, quantize = None
+    ):
+        super().__init__(learning_rate=learning_rate, use_gradient_checkpointing=use_gradient_checkpointing, state_dict_converter=state_dict_converter)
+        # Load models
+        model_manager = ModelManager(torch_dtype=torch_dtype, device=self.device)
+        if quantize is None:
+            model_manager.load_models(pretrained_weights)
+        else:
+            model_manager.load_models(pretrained_weights[1:])
+            model_manager.load_model(pretrained_weights[0], torch_dtype=quantize)
+        if preset_lora_path is not None:
+            preset_lora_path = preset_lora_path.split(",")
+            for path in preset_lora_path:
+                model_manager.load_lora(path)
+            
+        self.pipe = FluxImagePipeline.from_model_manager(model_manager)
+        self.pipe.reference_embedder = FluxReferenceEmbedder()
+        self.pipe.reference_embedder.init()
+        
+        if quantize is not None:
+            self.pipe.dit.quantize()
+        
+        self.pipe.scheduler.set_timesteps(1000, training=True)
+
+        self.freeze_parameters()
+        self.pipe.reference_embedder.requires_grad_(True)
+        self.pipe.reference_embedder.train()
+        self.pipe.dit.requires_grad_(True)
+        self.pipe.dit.train()
+        # self.add_lora_to_model(
+        #     self.pipe.denoising_model(),
+        #     lora_rank=lora_rank,
+        #     lora_alpha=lora_alpha,
+        #     lora_target_modules=lora_target_modules,
+        #     init_lora_weights=init_lora_weights,
+        #     pretrained_lora_path=pretrained_lora_path,
+        #     state_dict_converter=FluxLoRAConverter.align_to_diffsynth_format
+        # )
+        
+        
+    def training_step(self, batch, batch_idx):
+        # Data
+        text, image = batch["instruction"], batch["image_2"]
+        image_ref = batch["image_1"]
+
+        # Prepare input parameters
+        self.pipe.device = self.device
+        prompt_emb = self.pipe.encode_prompt(text, positive=True)
+        if "latents" in batch:
+            latents = batch["latents"].to(dtype=self.pipe.torch_dtype, device=self.device)
+        else:
+            latents = self.pipe.vae_encoder(image.to(dtype=self.pipe.torch_dtype, device=self.device))
+        noise = torch.randn_like(latents)
+        timestep_id = torch.randint(0, self.pipe.scheduler.num_train_timesteps, (1,))
+        timestep = self.pipe.scheduler.timesteps[timestep_id].to(self.device)
+        extra_input = self.pipe.prepare_extra_input(latents)
+        noisy_latents = self.pipe.scheduler.add_noise(latents, noise, timestep)
+        training_target = self.pipe.scheduler.training_target(latents, noise, timestep)
+        
+        # Reference image
+        hidden_states_ref = self.pipe.vae_encoder(image_ref.to(dtype=self.pipe.torch_dtype, device=self.device))
+
+        # Compute loss
+        noise_pred = lets_dance_flux(
+            self.pipe.denoising_model(),
+            reference_embedder=self.pipe.reference_embedder,
+            hidden_states_ref=hidden_states_ref,
+            hidden_states=noisy_latents, timestep=timestep, **prompt_emb, **extra_input,
+            use_gradient_checkpointing=self.use_gradient_checkpointing
+        )
+        loss = torch.nn.functional.mse_loss(noise_pred.float(), training_target.float())
+        loss = loss * self.pipe.scheduler.training_weight(timestep)
+
+        # Record log
+        self.log("train_loss", loss, prog_bar=True)
+        return loss
+    
+    
+    def configure_optimizers(self):
+        trainable_modules = filter(lambda p: p.requires_grad, self.pipe.parameters())
+        optimizer = torch.optim.AdamW(trainable_modules, lr=self.learning_rate)
+        return optimizer
+    
+
+    def on_save_checkpoint(self, checkpoint):
+        checkpoint.clear()
+        trainable_param_names = list(filter(lambda named_param: named_param[1].requires_grad, self.pipe.named_parameters()))
+        trainable_param_names = set([named_param[0] for named_param in trainable_param_names])
+        state_dict = self.pipe.state_dict()
+        lora_state_dict = {}
+        for name, param in state_dict.items():
+            if name in trainable_param_names:
+                lora_state_dict[name] = param
+        if self.state_dict_converter is not None:
+            lora_state_dict = self.state_dict_converter(lora_state_dict, alpha=self.lora_alpha)
+        checkpoint.update(lora_state_dict)
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Simple example of a training script.")
+    parser.add_argument(
+        "--pretrained_text_encoder_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained text encoder model. For example, `models/FLUX/FLUX.1-dev/text_encoder/model.safetensors`.",
+    )
+    parser.add_argument(
+        "--pretrained_text_encoder_2_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained t5 text encoder model. For example, `models/FLUX/FLUX.1-dev/text_encoder_2`.",
+    )
+    parser.add_argument(
+        "--pretrained_dit_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained dit model. For example, `models/FLUX/FLUX.1-dev/flux1-dev.safetensors`.",
+    )
+    parser.add_argument(
+        "--pretrained_vae_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained vae model. For example, `models/FLUX/FLUX.1-dev/ae.safetensors`.",
+    )
+    parser.add_argument(
+        "--lora_target_modules",
+        type=str,
+        default="a_to_qkv,b_to_qkv,ff_a.0,ff_a.2,ff_b.0,ff_b.2,a_to_out,b_to_out,proj_out,norm.linear,norm1_a.linear,norm1_b.linear,to_qkv_mlp",
+        help="Layers with LoRA modules.",
+    )
+    parser.add_argument(
+        "--align_to_opensource_format",
+        default=False,
+        action="store_true",
+        help="Whether to export lora files aligned with other opensource format.",
+    )
+    parser.add_argument(
+        "--quantize",
+        type=str,
+        default=None,
+        choices=["float8_e4m3fn"],
+        help="Whether to use quantization when training the model, and in which format.",
+    )
+    parser.add_argument(
+        "--preset_lora_path",
+        type=str,
+        default=None,
+        help="Preset LoRA path.",
+    )
+    parser = add_general_parsers(parser)
+    args = parser.parse_args()
+    return args
+
+
+if __name__ == '__main__':
+    args = parse_args()
+    model = LightningModel(
+        torch_dtype={"32": torch.float32, "bf16": torch.bfloat16}.get(args.precision, torch.float16),
+        pretrained_weights=[args.pretrained_dit_path, args.pretrained_text_encoder_path, args.pretrained_text_encoder_2_path, args.pretrained_vae_path],
+        preset_lora_path=args.preset_lora_path,
+        learning_rate=args.learning_rate,
+        use_gradient_checkpointing=args.use_gradient_checkpointing,
+        lora_rank=args.lora_rank,
+        lora_alpha=args.lora_alpha,
+        lora_target_modules=args.lora_target_modules,
+        init_lora_weights=args.init_lora_weights,
+        pretrained_lora_path=args.pretrained_lora_path,
+        state_dict_converter=FluxLoRAConverter.align_to_opensource_format if args.align_to_opensource_format else None,
+        quantize={"float8_e4m3fn": torch.float8_e4m3fn}.get(args.quantize, None),
+    )
+    # dataset and data loader
+    dataset = MultiTaskDataset(
+        dataset_list=[
+            SingleTaskDataset(
+                "/shark/zhongjie/data/image_pulse_datasets/task1/data/dataset_change_add_remove",
+                keys=(("image_1", "image_2", "editing_instruction"), ("image_2", "image_1", "reverse_editing_instruction")),
+                metadata_path="/shark/zhongjie/data/image_pulse_datasets/task1/data/metadata/20250418_dataset_change_add_remove.json",
+                height=512, width=512,
+            ),
+            SingleTaskDataset(
+                "/shark/zhongjie/data/image_pulse_datasets/task1/data/dataset_zoomin_zoomout",
+                keys=(("image_1", "image_2", "editing_instruction"), ("image_2", "image_1", "reverse_editing_instruction")),
+                metadata_path="/shark/zhongjie/data/image_pulse_datasets/task1/data/metadata/20250418_dataset_zoomin_zoomout.json",
+                height=512, width=512,
+            ),
+            SingleTaskDataset(
+                "/shark/zhongjie/data/image_pulse_datasets/task1/data/dataset_style_transfer",
+                keys=(("image_1", "image_4", "editing_instruction"), ("image_4", "image_1", "reverse_editing_instruction")),
+                metadata_path="/shark/zhongjie/data/image_pulse_datasets/task1/data/metadata/20250418_dataset_style_transfer.json",
+                height=512, width=512,
+            ),
+            SingleTaskDataset(
+                "/shark/zhongjie/data/image_pulse_datasets/task1/data/dataset_faceid",
+                keys=(("image_1", "image_2", "editing_instruction"), ("image_2", "image_1", "reverse_editing_instruction")),
+                metadata_path="/shark/zhongjie/data/image_pulse_datasets/task1/data/metadata/20250418_dataset_faceid.json",
+                height=512, width=512,
+            ),
+        ],
+        dataset_weight=(4, 1, 4, 1),
+        steps_per_epoch=args.steps_per_epoch,
+    )
+    train_loader = torch.utils.data.DataLoader(
+        dataset,
+        shuffle=True,
+        batch_size=args.batch_size,
+        num_workers=args.dataloader_num_workers
+    )
+    # train
+    trainer = pl.Trainer(
+        max_epochs=args.max_epochs,
+        accelerator="gpu",
+        devices="auto",
+        precision=args.precision,
+        strategy=args.training_strategy,
+        default_root_dir=args.output_path,
+        accumulate_grad_batches=args.accumulate_grad_batches,
+        callbacks=[pl.pytorch.callbacks.ModelCheckpoint(save_top_k=-1)],
+        logger=None,
+    )
+    trainer.fit(model=model, train_dataloaders=train_loader)

train_flux_reference_multi_node.py

@@ -0,0 +1,248 @@
+from diffsynth import ModelManager, FluxImagePipeline
+from diffsynth.trainers.text_to_image import LightningModelForT2ILoRA, add_general_parsers, launch_training_task
+from diffsynth.models.lora import FluxLoRAConverter
+import torch, os, argparse
+import lightning as pl
+from diffsynth.data.image_pulse import SingleTaskDataset, MultiTaskDataset
+from diffsynth.pipelines.flux_image import lets_dance_flux
+from diffsynth.models.flux_reference_embedder import FluxReferenceEmbedder
+os.environ["TOKENIZERS_PARALLELISM"] = "True"
+
+
+class LightningModel(LightningModelForT2ILoRA):
+    def __init__(
+        self,
+        torch_dtype=torch.float16, pretrained_weights=[], preset_lora_path=None,
+        learning_rate=1e-4, use_gradient_checkpointing=True,
+        lora_rank=4, lora_alpha=4, lora_target_modules="to_q,to_k,to_v,to_out", init_lora_weights="kaiming", pretrained_lora_path=None,
+        state_dict_converter=None, quantize = None
+    ):
+        super().__init__(learning_rate=learning_rate, use_gradient_checkpointing=use_gradient_checkpointing, state_dict_converter=state_dict_converter)
+        # Load models
+        model_manager = ModelManager(torch_dtype=torch_dtype, device=self.device)
+        if quantize is None:
+            model_manager.load_models(pretrained_weights)
+        else:
+            model_manager.load_models(pretrained_weights[1:])
+            model_manager.load_model(pretrained_weights[0], torch_dtype=quantize)
+        if preset_lora_path is not None:
+            preset_lora_path = preset_lora_path.split(",")
+            for path in preset_lora_path:
+                model_manager.load_lora(path)
+            
+        self.pipe = FluxImagePipeline.from_model_manager(model_manager)
+        self.pipe.reference_embedder = FluxReferenceEmbedder()
+        self.pipe.reference_embedder.init()
+        
+        if quantize is not None:
+            self.pipe.dit.quantize()
+        
+        self.pipe.scheduler.set_timesteps(1000, training=True)
+
+        self.freeze_parameters()
+        self.pipe.reference_embedder.requires_grad_(True)
+        self.pipe.reference_embedder.train()
+        self.pipe.dit.requires_grad_(True)
+        self.pipe.dit.train()
+        # self.add_lora_to_model(
+        #     self.pipe.denoising_model(),
+        #     lora_rank=lora_rank,
+        #     lora_alpha=lora_alpha,
+        #     lora_target_modules=lora_target_modules,
+        #     init_lora_weights=init_lora_weights,
+        #     pretrained_lora_path=pretrained_lora_path,
+        #     state_dict_converter=FluxLoRAConverter.align_to_diffsynth_format
+        # )
+        
+        
+    def training_step(self, batch, batch_idx):
+        # Data
+        text, image = batch["instruction"], batch["image_2"]
+        image_ref = batch["image_1"]
+
+        # Prepare input parameters
+        self.pipe.device = self.device
+        prompt_emb = self.pipe.encode_prompt(text, positive=True)
+        if "latents" in batch:
+            latents = batch["latents"].to(dtype=self.pipe.torch_dtype, device=self.device)
+        else:
+            latents = self.pipe.vae_encoder(image.to(dtype=self.pipe.torch_dtype, device=self.device))
+        noise = torch.randn_like(latents)
+        timestep_id = torch.randint(0, self.pipe.scheduler.num_train_timesteps, (1,))
+        timestep = self.pipe.scheduler.timesteps[timestep_id].to(self.device)
+        extra_input = self.pipe.prepare_extra_input(latents)
+        noisy_latents = self.pipe.scheduler.add_noise(latents, noise, timestep)
+        training_target = self.pipe.scheduler.training_target(latents, noise, timestep)
+        
+        # Reference image
+        hidden_states_ref = self.pipe.vae_encoder(image_ref.to(dtype=self.pipe.torch_dtype, device=self.device))
+
+        # Compute loss
+        noise_pred = lets_dance_flux(
+            self.pipe.denoising_model(),
+            reference_embedder=self.pipe.reference_embedder,
+            hidden_states_ref=hidden_states_ref,
+            hidden_states=noisy_latents, timestep=timestep, **prompt_emb, **extra_input,
+            use_gradient_checkpointing=self.use_gradient_checkpointing
+        )
+        loss = torch.nn.functional.mse_loss(noise_pred.float(), training_target.float())
+        loss = loss * self.pipe.scheduler.training_weight(timestep)
+
+        # Record log
+        self.log("train_loss", loss, prog_bar=True)
+        return loss
+    
+    
+    def configure_optimizers(self):
+        trainable_modules = filter(lambda p: p.requires_grad, self.pipe.parameters())
+        optimizer = torch.optim.AdamW(trainable_modules, lr=self.learning_rate)
+        return optimizer
+    
+
+    def on_save_checkpoint(self, checkpoint):
+        checkpoint.clear()
+        trainable_param_names = list(filter(lambda named_param: named_param[1].requires_grad, self.pipe.named_parameters()))
+        trainable_param_names = set([named_param[0] for named_param in trainable_param_names])
+        state_dict = self.pipe.state_dict()
+        lora_state_dict = {}
+        for name, param in state_dict.items():
+            if name in trainable_param_names:
+                lora_state_dict[name] = param
+        if self.state_dict_converter is not None:
+            lora_state_dict = self.state_dict_converter(lora_state_dict, alpha=self.lora_alpha)
+        checkpoint.update(lora_state_dict)
+
+
+def parse_args():
+    parser = argparse.ArgumentParser(description="Simple example of a training script.")
+    parser.add_argument(
+        "--pretrained_text_encoder_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained text encoder model. For example, `models/FLUX/FLUX.1-dev/text_encoder/model.safetensors`.",
+    )
+    parser.add_argument(
+        "--pretrained_text_encoder_2_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained t5 text encoder model. For example, `models/FLUX/FLUX.1-dev/text_encoder_2`.",
+    )
+    parser.add_argument(
+        "--pretrained_dit_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained dit model. For example, `models/FLUX/FLUX.1-dev/flux1-dev.safetensors`.",
+    )
+    parser.add_argument(
+        "--pretrained_vae_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained vae model. For example, `models/FLUX/FLUX.1-dev/ae.safetensors`.",
+    )
+    parser.add_argument(
+        "--lora_target_modules",
+        type=str,
+        default="a_to_qkv,b_to_qkv,ff_a.0,ff_a.2,ff_b.0,ff_b.2,a_to_out,b_to_out,proj_out,norm.linear,norm1_a.linear,norm1_b.linear,to_qkv_mlp",
+        help="Layers with LoRA modules.",
+    )
+    parser.add_argument(
+        "--align_to_opensource_format",
+        default=False,
+        action="store_true",
+        help="Whether to export lora files aligned with other opensource format.",
+    )
+    parser.add_argument(
+        "--quantize",
+        type=str,
+        default=None,
+        choices=["float8_e4m3fn"],
+        help="Whether to use quantization when training the model, and in which format.",
+    )
+    parser.add_argument(
+        "--preset_lora_path",
+        type=str,
+        default=None,
+        help="Preset LoRA path.",
+    )
+    parser.add_argument(
+        "--num_nodes",
+        type=int,
+        default=1,
+        help="Num nodes.",
+    )
+    parser = add_general_parsers(parser)
+    args = parser.parse_args()
+    return args
+
+
+if __name__ == '__main__':
+    args = parse_args()
+    model = LightningModel(
+        torch_dtype={"32": torch.float32, "bf16": torch.bfloat16}.get(args.precision, torch.float16),
+        pretrained_weights=[args.pretrained_dit_path, args.pretrained_text_encoder_path, args.pretrained_text_encoder_2_path, args.pretrained_vae_path],
+        preset_lora_path=args.preset_lora_path,
+        learning_rate=args.learning_rate,
+        use_gradient_checkpointing=args.use_gradient_checkpointing,
+        lora_rank=args.lora_rank,
+        lora_alpha=args.lora_alpha,
+        lora_target_modules=args.lora_target_modules,
+        init_lora_weights=args.init_lora_weights,
+        pretrained_lora_path=args.pretrained_lora_path,
+        state_dict_converter=FluxLoRAConverter.align_to_opensource_format if args.align_to_opensource_format else None,
+        quantize={"float8_e4m3fn": torch.float8_e4m3fn}.get(args.quantize, None),
+    )
+    # dataset and data loader
+    dataset = MultiTaskDataset(
+        dataset_list=[
+            SingleTaskDataset(
+                "/shark/zhongjie/data/image_pulse_datasets/task1/data/dataset_change_add_remove",
+                keys=(("image_1", "image_2", "editing_instruction"), ("image_2", "image_1", "reverse_editing_instruction")),
+                metadata_path="/shark/zhongjie/data/image_pulse_datasets/task1/data/metadata/20250418_dataset_change_add_remove.json",
+                height=512, width=512,
+            ),
+            SingleTaskDataset(
+                "/shark/zhongjie/data/image_pulse_datasets/task1/data/dataset_zoomin_zoomout",
+                keys=(("image_1", "image_2", "editing_instruction"), ("image_2", "image_1", "reverse_editing_instruction")),
+                metadata_path="/shark/zhongjie/data/image_pulse_datasets/task1/data/metadata/20250418_dataset_zoomin_zoomout.json",
+                height=512, width=512,
+            ),
+            SingleTaskDataset(
+                "/shark/zhongjie/data/image_pulse_datasets/task1/data/dataset_style_transfer",
+                keys=(("image_1", "image_4", "editing_instruction"), ("image_4", "image_1", "reverse_editing_instruction")),
+                metadata_path="/shark/zhongjie/data/image_pulse_datasets/task1/data/metadata/20250418_dataset_style_transfer.json",
+                height=512, width=512,
+            ),
+            SingleTaskDataset(
+                "/shark/zhongjie/data/image_pulse_datasets/task1/data/dataset_faceid",
+                keys=(("image_1", "image_2", "editing_instruction"), ("image_2", "image_1", "reverse_editing_instruction")),
+                metadata_path="/shark/zhongjie/data/image_pulse_datasets/task1/data/metadata/20250418_dataset_faceid.json",
+                height=512, width=512,
+            ),
+        ],
+        dataset_weight=(4, 1, 4, 1),
+        steps_per_epoch=args.steps_per_epoch,
+    )
+    train_loader = torch.utils.data.DataLoader(
+        dataset,
+        shuffle=True,
+        batch_size=args.batch_size,
+        num_workers=args.dataloader_num_workers
+    )
+    # train
+    trainer = pl.Trainer(
+        max_epochs=args.max_epochs,
+        accelerator="gpu",
+        devices="auto",
+        num_nodes=args.num_nodes,
+        precision=args.precision,
+        strategy="ddp",
+        default_root_dir=args.output_path,
+        accumulate_grad_batches=args.accumulate_grad_batches,
+        callbacks=[pl.pytorch.callbacks.ModelCheckpoint(save_top_k=-1)],
+        logger=None,
+    )
+    trainer.fit(model=model, train_dataloaders=train_loader)