submit app.py

LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright [yyyy] [name of copyright owner]
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

app.py

@@ -0,0 +1,239 @@
+import json
+import numpy as np
+import math
+import csv
+import random
+import argparse
+import torch
+import os
+import torch.distributed as dist
+import gradio as gr
+from PIL import Image
+from torch.nn.parallel import DistributedDataParallel as DDP
+import spaces
+from accelerate.utils import set_seed
+
+from diffusion_pipeline.sd35_pipeline import StableDiffusion3Pipeline, FlowMatchEulerInverseScheduler
+from diffusion_pipeline.sdxl_pipeline import StableDiffusionXLPipeline
+from diffusers import BitsAndBytesConfig, SD3Transformer2DModel
+from diffusers import FlowMatchEulerDiscreteScheduler, DDIMInverseScheduler, DDIMScheduler
+
+device = torch.device('cuda')
+
+
+@spaces.GPU
+def generate_image(
+    model_name,
+    seed,
+    num_steps,
+    guidance_scale,
+    inv_cfg,
+    w2s_guidance,
+    end_timesteps,
+    prompt,
+    method,
+    size,
+):
+    try:
+        # 根据传入的参数生成图像
+        torch.cuda.empty_cache()
+        dtype = torch.float16
+        set_seed(seed)
+        if model_name == 'sd35':
+            nf4_config = BitsAndBytesConfig(
+                load_in_4bit=True,
+                bnb_4bit_quant_type="nf4",
+                bnb_4bit_compute_dtype=torch.bfloat16
+            )
+            model_nf4 = SD3Transformer2DModel.from_pretrained(
+                "stabilityai/stable-diffusion-3.5-large",
+                subfolder="transformer",
+                quantization_config=nf4_config,
+                torch_dtype=torch.bfloat16
+            )
+
+            pipe = StableDiffusion3Pipeline.from_pretrained(
+                "stabilityai/stable-diffusion-3.5-large", 
+                transformer=model_nf4,
+                torch_dtype=torch.bfloat16,
+            )
+
+            pipe.scheduler = FlowMatchEulerDiscreteScheduler.from_config(pipe.scheduler.config)
+            inverse_scheduler = FlowMatchEulerInverseScheduler.from_pretrained("stabilityai/stable-diffusion-3.5-large",
+                                                                  subfolder='scheduler')
+            pipe.inv_scheduler = inverse_scheduler
+
+        elif model_name == "sdxl":
+            pipe = StableDiffusionXLPipeline.from_pretrained(
+                "stabilityai/stable-diffusion-xl-base-1.0", 
+                torch_dtype=torch.float16, 
+                variant="fp16",
+                use_safetensors=True
+            ).to("cuda")
+
+            pipe.scheduler = DDIMScheduler.from_config(pipe.scheduler.config)
+            inverse_scheduler = DDIMInverseScheduler.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0",
+                                                                  subfolder='scheduler')
+            pipe.inv_scheduler = inverse_scheduler
+
+        pipe.to(device)
+        pipe.enable_model_cpu_offload()
+
+        # TODO: load noise model
+        if method == 'core' or method == 'z-core':
+            from diffusion_pipeline.refine_model import PromptSD35Net, PromptSDXLNet
+            from diffusion_pipeline.lora import replace_linear_with_lora, lora_true
+
+            if model_name == 'sd35':
+                refine_model = PromptSD35Net()
+                replace_linear_with_lora(refine_model, rank=64, alpha=1.0, number_of_lora=28)
+                lora_true(refine_model, lora_idx=0)
+
+                os.makedirs('./weights', exist_ok=True)
+                if not os.path.exists('./weights/sd35_noise_model.pth'):
+                    os.system('wget https://huggingface.co/sst12345/CoRe2/resolve/main/weights/sd35_noise_model.pth')
+                    os.system('mv sd35_noise_model.pth ./weights/')
+                checkpoint = torch.load('./weights/sd35_noise_model.pth', map_location='cpu')
+                refine_model.load_state_dict(checkpoint)
+            elif model_name == 'sdxl':
+                refine_model = PromptSDXLNet()
+                replace_linear_with_lora(refine_model, rank=48, alpha=1.0, number_of_lora=50)
+                lora_true(refine_model, lora_idx=0)
+                os.makedirs('./weights', exist_ok=True)
+                if not os.path.exists('./weights/sdxl_noise_model.pth'):
+                    os.system('wget https://huggingface.co/sst12345/CoRe2/resolve/main/weights/sdxl_noise_model.pth')
+                    os.system('mv sdxl_noise_model.pth ./weights/')
+                checkpoint = torch.load('./weights/sdxl_noise_model.pth', map_location='cpu')
+                refine_model.load_state_dict(checkpoint)
+
+            print("Load Lora Success")
+            refine_model = refine_model.to(device)
+            refine_model = refine_model.to(torch.bfloat16)
+
+        # 根据模型类型设置形状
+        if model_name == 'sdxl':
+            shape = (1, 4, size // 8, size // 8)
+        else:
+            shape = (1, 16, size // 8, size // 8)
+        
+        start_latents = torch.randn(shape, dtype=dtype).to(device)
+        
+        # 根据方法选择生成图像
+        if model_name == 'sdxl':
+            if method == 'core':
+                output = pipe.core(
+                    prompt=prompt,
+                    guidance_scale=guidance_scale,
+                    num_inference_steps=num_steps,
+                    latents=start_latents,
+                    return_dict=False,
+                    refine_model=refine_model,
+                    lora_true=lora_true,
+                    end_timesteps=end_timesteps,
+                    w2s_guidance=w2s_guidance)[0][0]
+            elif method == 'zigzag':
+                output = pipe.zigzag(
+                    prompt=prompt,
+                    guidance_scale=guidance_scale,
+                    latents=start_latents,
+                    return_dict=False,
+                    num_inference_steps=num_steps,
+                    inv_cfg=inv_cfg)[0][0]
+            elif method == 'z-core':
+                output = pipe.z_core(
+                    prompt=prompt,
+                    guidance_scale=guidance_scale,
+                    num_inference_steps=num_steps,
+                    latents=start_latents,
+                    return_dict=False,
+                    refine_model=refine_model,
+                    lora_true=lora_true,
+                    end_timesteps=end_timesteps,
+                    w2s_guidance=w2s_guidance,
+                    inv_cfg=inv_cfg)[0][0]
+            elif method == 'standard':
+                output = pipe(
+                    prompt=prompt,
+                    guidance_scale=guidance_scale,
+                    latents=start_latents,
+                    return_dict=False,
+                    num_inference_steps=num_steps)[0][0]
+            else:
+                raise ValueError("Invalid method")
+        else:
+            if method == 'core':
+                output = pipe.core(
+                    prompt=prompt,
+                    guidance_scale=guidance_scale,
+                    num_inference_steps=num_steps,
+                    latents=start_latents,
+                    max_sequence_length=512,
+                    return_dict=False,
+                    refine_model=refine_model,
+                    lora_true=lora_true,
+                    end_timesteps=end_timesteps,
+                    w2s_guidance=w2s_guidance)[0][0]
+            elif method == 'zigzag':
+                output = pipe.zigzag(
+                    prompt=prompt,
+                    max_sequence_length=512,
+                    guidance_scale=guidance_scale,
+                    latents=start_latents,
+                    return_dict=False,
+                    num_inference_steps=num_steps,
+                    inv_cfg=inv_cfg)[0][0]
+            elif method == 'z-core':
+                output = pipe.z_core(
+                    prompt=prompt,
+                    guidance_scale=guidance_scale,
+                    num_inference_steps=num_steps,
+                    latents=start_latents,
+                    return_dict=False,
+                    max_sequence_length=512,
+                    refine_model=refine_model,
+                    lora_true=lora_true,
+                    end_timesteps=end_timesteps,
+                    w2s_guidance=w2s_guidance)[0][0]
+            elif method == 'standard':
+                output = pipe(
+                    prompt=prompt,
+                    guidance_scale=guidance_scale,
+                    latents=start_latents,
+                    return_dict=False,
+                    max_sequence_length=512,
+                    num_inference_steps=num_steps)[0][0]
+            else:
+                raise ValueError("Invalid method")
+
+        # 将生成的图像保存为临时文件并返回
+        output_path = f'{model_name}_{method}.png'
+        output.save(output_path)
+        return output_path
+
+    except Exception as e:
+        print(f"An error occurred: {e}")
+        return None
+
+
+
+if __name__ == '__main__':
+      # 创建Gradio接口
+      iface = gr.Interface(
+      fn=generate_image,
+      inputs=[
+            gr.Dropdown(choices=['sdxl', 'sd35'], value='sdxl', label="Model"),  # 设置默认模型为 'sdxl'
+            gr.Slider(minimum=1, maximum=1000000, value=1, label="seed"),  # 设置默认种子为 1
+            gr.Slider(minimum=1, maximum=100, value=50, label="Inference Steps"),  # 设置默认推理步数为 50
+            gr.Slider(minimum=1, maximum=10, value=5.5, label="CFG"),  # 设置默认CFG为 5.5
+            gr.Slider(minimum=-10, maximum=10, value=-1, label="Inverse CFG"),  # 设置默认逆CFG为 -1
+            gr.Slider(minimum=1, maximum=3.5, value=2.5, label="W2S Guidance"),  # 设置默认W2S指导为 2.5
+            gr.Slider(minimum=1, maximum=100, value=50, label="End Timesteps"),  # 设置默认结束时间步为 50
+            gr.Textbox(label="Prompt"),  # 文本框没有默认值
+            gr.Dropdown(choices=['standard', 'core', 'zigzag', 'z-core'], value='core', label="Method"),  # 设置默认方法为 'core'
+            gr.Slider(minimum=1024, maximum=2048, value=1024, label="Size")  # 设置默认大小为 1024
+      ],
+      outputs=gr.Image(type="filepath"),  # 修改了type参数
+      title="Image Generation with CoRe^2"
+      )
+      iface.launch()
+    

diffusion_pipeline/gemma.py

@@ -0,0 +1,53 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import AutoTokenizer, Gemma2ForTokenClassification, BitsAndBytesConfig
+
+import os
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+torch.set_float32_matmul_precision("high")
+
+def repeat_function(xs, max_length = 128):
+    new_xs = []
+    for x in xs:
+        if x.shape[1] >= max_length-1:
+            new_xs.append(x[:,:max_length-1,:])
+        else:
+            new_xs.append(x)
+    xs = new_xs
+    mean_xs = [x.mean(1,keepdim=True).expand(-1,max_length - x.shape[1],-1) for x in xs]
+    xs = [torch.cat([x,mean_x],1) for mean_x, x in zip(mean_xs, xs)]
+    return xs
+
+class Gemma2Model(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.tokenizer = AutoTokenizer.from_pretrained("google/gemma-2-2b", )
+        self.tokenizer_max_length = 128
+        # quantization_config = BitsAndBytesConfig(load_in_8bit=True)
+
+        self.model = Gemma2ForTokenClassification.from_pretrained(
+            "google/gemma-2-2b",
+            # device_map="auto",
+            # quantization_config=quantization_config,
+        ).float()
+        self.model.score = nn.Identity()
+    
+    @torch.no_grad()
+    def forward(self, input_prompt):
+        input_prompt = list(input_prompt)
+        outputs = []
+        for _input_prompt in input_prompt:
+            input_ids = self.tokenizer(_input_prompt, add_special_tokens=False, max_length=77, return_tensors="pt").to("cuda")
+            _outputs = self.model(**input_ids)["logits"]
+            outputs.append(_outputs)
+        outputs = repeat_function(outputs)
+        outputs = torch.cat(outputs,0)
+        return outputs
+
+if __name__ == "__main__":
+    model = Gemma2Model().cuda()
+    input_text = ["Write me a poem about Machine Learning.", "Write me a poem about Deep Learning."]
+    print(model(input_text))
+    print(model(input_text)[0].shape)
+    print(model(input_text).shape)

diffusion_pipeline/lora.py

@@ -0,0 +1,62 @@
+import torch
+import torch.nn as nn
+
+class LoRALayer(torch.nn.Module):
+    def __init__(self, in_dim, out_dim, rank, alpha):
+        super().__init__()
+        std_dev = 1 / torch.sqrt(torch.tensor(rank).float())
+        self.A = torch.nn.Parameter(torch.randn(in_dim, rank) * std_dev)
+        self.B = torch.nn.Parameter(torch.zeros(rank, out_dim))
+        self.alpha = alpha
+
+    def forward(self, x):
+        x = self.alpha * (x @ self.A @ self.B)
+        return x
+
+class LinearWithLoRA(torch.nn.Module):
+    def __init__(self, linear, rank, alpha, 
+                 weak_lora_alpha=0.1, number_of_lora=1):
+        super().__init__()
+        self.linear = linear
+        self.lora = nn.ModuleList([LoRALayer(
+            linear.in_features, linear.out_features, rank, alpha
+        ) for _ in range(number_of_lora)])
+        self.use_lora = True
+        self.lora_idx = 0
+        
+    def forward(self, x):
+        if self.use_lora:
+            return self.linear(x) + self.lora[self.lora_idx](x)
+        else:
+            return self.linear(x)
+
+def replace_linear_with_lora(module, rank=64, alpha=1., tag=0, weak_lora_alpha=0.1, number_of_lora=1):
+    for name, child in module.named_children():
+        if isinstance(child, nn.Linear):
+            setattr(module, name, LinearWithLoRA(child, rank, alpha, weak_lora_alpha=weak_lora_alpha, number_of_lora=number_of_lora))
+        else:
+            replace_linear_with_lora(child, rank, alpha, tag, weak_lora_alpha=weak_lora_alpha, number_of_lora=number_of_lora)
+
+
+def lora_false(model, lora_idx=0):
+    for name, module in model.named_modules():
+        if isinstance(module, LinearWithLoRA):
+            module.use_lora = False
+            module.lora_idx = lora_idx
+
+def lora_true(model, lora_idx=0):
+    for name, module in model.named_modules():
+        if isinstance(module, LinearWithLoRA):
+            module.use_lora = True
+            module.lora_idx = lora_idx
+            for i, lora in enumerate(module.lora):
+                if i != lora_idx:
+                    lora.A.requires_grad = False
+                    lora.B.requires_grad = False
+                    if lora.A.grad is not None:
+                        del lora.A.grad
+                    if lora.B.grad is not None:
+                        del lora.B.grad
+                else:
+                    lora.A.requires_grad = True
+                    lora.B.requires_grad = True

diffusion_pipeline/refine_model.py

@@ -0,0 +1,312 @@
+
+import torch
+import torch.nn as nn
+import os
+import json
+import torch.nn.functional as F
+import random
+from torch.utils.data import Dataset
+from transformers import AutoTokenizer
+from glob import glob
+import math
+from PIL import Image
+device = torch.device('cuda')
+import numpy as np
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+
+from diffusers.utils import logging
+from diffusers.models.embeddings import PatchEmbed
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+from diffusers.models.attention import BasicTransformerBlock
+from diffusers.models.normalization import AdaLayerNormContinuous
+from torchvision import transforms
+    
+def add_hook_to_module(model, module_name):
+    outputs = []
+    def hook(module, input, output):
+        outputs.append(output)
+    module = dict(model.named_modules()).get(module_name)
+    if module is None:
+        raise ValueError(f"can't find module {module_name}")
+    hook_handle = module.register_forward_hook(hook)
+    return hook_handle, outputs
+
+class PromptSD35Net(nn.Module):
+
+    def __init__(self,         
+        sample_size: int = 128,
+        patch_size: int = 2,
+        in_channels: int = 16,
+        num_layers: int = 8,
+        attention_head_dim: int = 64,
+        num_attention_heads: int = 24,
+        out_channels: int = 16,
+        pos_embed_max_size: int = 192
+        ):
+        super().__init__()
+        self.sample_size = sample_size
+        self.patch_size = patch_size
+        self.in_channels = in_channels
+        self.num_layers = num_layers
+        self.attention_head_dim = attention_head_dim
+        self.num_attention_heads = num_attention_heads
+        self.out_channels = out_channels
+        self.pos_embed_max_size = pos_embed_max_size
+        self.inner_dim = self.num_attention_heads * self.attention_head_dim
+        
+        self.pos_embed = PatchEmbed(
+            height=self.sample_size,
+            width=self.sample_size,
+            patch_size=self.patch_size,
+            in_channels=self.in_channels,
+            embed_dim=self.inner_dim,
+            pos_embed_max_size=pos_embed_max_size
+        )
+        
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    dim=self.inner_dim,
+                    num_attention_heads=self.num_attention_heads,
+                    attention_head_dim=self.attention_head_dim,
+                    ff_inner_dim=2*self.inner_dim   # mult should be 4 by default
+                )
+                for i in range(self.num_layers)
+            ]
+        )
+        self.norm_out = AdaLayerNormContinuous(self.inner_dim, self.inner_dim, elementwise_affine=False, eps=1e-6)
+        self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True)
+        
+        self.noise_shape = (1, 16, 128, 128) # (667, 4096)
+        self.pre8_linear = nn.Sequential(nn.Linear(4096, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536))
+        self.pre16_linear = nn.Sequential(nn.Linear(4096, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536))
+        self.pre24_linear = nn.Sequential(nn.Linear(4096, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536))
+
+        self.pre8_linear2 = nn.Sequential(nn.Linear(4096, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536))
+        self.pre16_linear2 = nn.Sequential(nn.Linear(4096, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536))
+        self.pre24_linear2 = nn.Sequential(nn.Linear(4096, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536))
+        
+        self.last_linear = nn.Sequential(nn.Linear(4096, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536))
+        # self.last_linear2 = nn.Sequential(nn.Linear(667, 32))
+        self.skip_connection2 = nn.Linear(4096, 1, bias=False)
+        self.skip_connection = nn.Linear(667, 32, bias=False)
+        self.trans_linear = nn.Linear(666+1+4096, 1536, bias=False)
+        nn.init.constant_(self.skip_connection.weight.data, 0)
+        nn.init.constant_(self.trans_linear.weight.data, 0)
+        nn.init.constant_(self.trans_linear.weight.data, 0)
+        nn.init.constant_(self.pre8_linear[-1].weight.data, 0)
+        nn.init.constant_(self.pre16_linear[-1].weight.data, 0)
+        nn.init.constant_(self.pre24_linear[-1].weight.data, 0)
+        nn.init.constant_(self.pre8_linear2[-1].weight.data, 0)
+        nn.init.constant_(self.pre16_linear2[-1].weight.data, 0)
+        nn.init.constant_(self.pre24_linear2[-1].weight.data, 0)
+
+    def forward(self, noise: torch.Tensor, _s, _v, _d, _pool_embedding) -> torch.Tensor:
+        
+        assert noise is not None
+        _ori_v = _v.clone()
+        _v = torch.stack([torch.diag(_v[jj]) for jj in range(_v.shape[0])], dim=0)
+        positive_embedding = _s.permute(0, 2, 1) @ _v @ _d # [2, 64, 666] [2, 64] [2, 64, 4096]
+        pool_embedding = _pool_embedding[:, None, :]
+        embedding = torch.cat([positive_embedding, pool_embedding], dim=1)
+        bs = noise.shape[0]
+        height, width = noise.shape[-2:]
+        embed_8 = embedding
+        embed_16 = embedding
+        embed_24 = embedding
+        scale_8 = self.pre8_linear2(embed_8).mean(1)
+        scale_16 = self.pre16_linear2(embed_16).mean(1)
+        scale_24 = self.pre24_linear2(embed_24).mean(1)
+        embed_8 = self.pre8_linear(embed_8).mean(1)
+        embed_16 = self.pre16_linear(embed_16).mean(1)
+        embed_24 = self.pre24_linear(embed_24).mean(1)
+        embed_last = self.last_linear(embedding).mean(1)
+        embed_trans = self.trans_linear(torch.cat([_s, _ori_v[...,None], _d], dim=2)).mean(1)
+        skip_embedding = self.skip_connection(self.skip_connection2(embedding).permute(0,2,1)).permute(0,2,1)
+        scale_skip, embed_skip = skip_embedding.chunk(2,dim=1)
+        
+        ori_noise = noise * (scale_skip[...,None]) + embed_skip[...,None]
+        noise = self.pos_embed(noise)        
+        noise = noise * (1 + scale_8[:, None, :] + embed_trans[:, None, :]) + embed_8[:, None, :]
+        scale_list = [scale_16, scale_24]
+        embed_list = [embed_16, embed_24]
+        for _ii, block in enumerate(self.transformer_blocks):
+            noise = block(noise)  
+            if len(scale_list)!=0 and len(embed_list)!=0:
+                noise = noise * (1 + scale_list[int(_ii//4)][:, None, :] + embed_trans[:, None, :]) + embed_list[int(_ii//4)][:, None, :]
+        
+        hidden_states = noise
+        hidden_states = self.norm_out(hidden_states, embed_last)
+        hidden_states = self.proj_out(hidden_states)
+
+        # unpatchify
+        patch_size = self.patch_size
+        height = height // patch_size
+        width = width // patch_size
+    
+        hidden_states = hidden_states.reshape(
+            shape=(hidden_states.shape[0], height, width, patch_size, patch_size, self.out_channels)
+        )
+        hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
+        output = hidden_states.reshape(
+            shape=(hidden_states.shape[0], self.out_channels, height * patch_size, width * patch_size)
+        )
+        return output + ori_noise
+
+    def weak_load_state_dict(self, state_dict: os.Mapping[str, torch.any], strict: bool = True, assign: bool = False):
+        return load_filtered_state_dict(self, state_dict)
+
+class PromptSDXLNet(nn.Module):
+
+    def __init__(self,         
+        sample_size: int = 128,
+        patch_size: int = 2,
+        in_channels: int = 4,
+        num_layers: int = 4,
+        attention_head_dim: int = 64,
+        num_attention_heads: int = 24,
+        out_channels: int = 4,
+        pos_embed_max_size: int = 192
+        ):
+        super().__init__()
+        self.sample_size = sample_size
+        self.patch_size = patch_size
+        self.in_channels = in_channels
+        self.num_layers = num_layers
+        self.attention_head_dim = attention_head_dim
+        self.num_attention_heads = num_attention_heads
+        self.out_channels = out_channels
+        self.pos_embed_max_size = pos_embed_max_size
+        self.inner_dim = self.num_attention_heads * self.attention_head_dim
+        
+        self.pos_embed = PatchEmbed(
+            height=self.sample_size,
+            width=self.sample_size,
+            patch_size=self.patch_size,
+            in_channels=self.in_channels,
+            embed_dim=self.inner_dim,
+            pos_embed_max_size=pos_embed_max_size
+        )
+        
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    dim=self.inner_dim,
+                    num_attention_heads=self.num_attention_heads,
+                    attention_head_dim=self.attention_head_dim,
+                    ff_inner_dim=2*self.inner_dim   # mult should be 4 by default
+                )
+                for i in range(self.num_layers)
+            ]
+        )
+        self.norm_out = AdaLayerNormContinuous(self.inner_dim, self.inner_dim, elementwise_affine=False, eps=1e-6)
+        self.proj_out = nn.Linear(self.inner_dim, patch_size * patch_size * self.out_channels, bias=True)
+        
+        self.noise_shape = (1, 4, 128, 128) 
+        self.pre8_linear = nn.Sequential(nn.Linear(2048, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536))
+        self.pre16_linear = nn.Sequential(nn.Linear(2048, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536))
+        self.pre24_linear = nn.Sequential(nn.Linear(2048, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536))
+
+        self.pre8_linear2 = nn.Sequential(nn.Linear(2048, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536))
+        self.pre16_linear2 = nn.Sequential(nn.Linear(2048, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536))
+        self.pre24_linear2 = nn.Sequential(nn.Linear(2048, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536))
+        
+        self.last_linear = nn.Sequential(nn.Linear(2048, 128), nn.SiLU(), nn.LayerNorm(128), nn.Linear(128, 1536))
+        # self.last_linear2 = nn.Sequential(nn.Linear(667, 32))
+        self.skip_connection2 = nn.Linear(2048, 1, bias=False)
+        self.skip_connection = nn.Linear(154+1, 8, bias=False)
+        self.trans_linear = nn.Linear(154+1+2048, 1536, bias=False)
+        self.pool_prompt_linear = nn.Linear(2560, 2048, bias=False)
+        nn.init.constant_(self.skip_connection.weight.data, 0)
+        nn.init.constant_(self.trans_linear.weight.data, 0)
+        nn.init.constant_(self.trans_linear.weight.data, 0)
+        nn.init.constant_(self.pre8_linear[-1].weight.data, 0)
+        nn.init.constant_(self.pre16_linear[-1].weight.data, 0)
+        nn.init.constant_(self.pre24_linear[-1].weight.data, 0)
+        nn.init.constant_(self.pre8_linear2[-1].weight.data, 0)
+        nn.init.constant_(self.pre16_linear2[-1].weight.data, 0)
+        nn.init.constant_(self.pre24_linear2[-1].weight.data, 0)
+
+    def forward(self, noise: torch.Tensor, _s, _v, _d, _pool_embedding) -> torch.Tensor:
+        
+        assert noise is not None
+        _ori_v = _v.clone()
+        _v = torch.stack([torch.diag(_v[jj]) for jj in range(_v.shape[0])], dim=0)
+        positive_embedding = _s.permute(0, 2, 1) @ _v @ _d # [2, 64, 154] [2, 64] [2, 64, 2048]
+        pool_embedding = self.pool_prompt_linear(_pool_embedding[:, None, :])
+        embedding = torch.cat([positive_embedding, pool_embedding], dim=1)
+        bs = noise.shape[0]
+        height, width = noise.shape[-2:]
+        embed_8 = embedding
+        embed_16 = embedding
+        embed_24 = embedding
+        scale_8 = self.pre8_linear2(embed_8).mean(1)
+        scale_16 = self.pre16_linear2(embed_16).mean(1)
+        scale_24 = self.pre24_linear2(embed_24).mean(1)
+        embed_8 = self.pre8_linear(embed_8).mean(1)
+        embed_16 = self.pre16_linear(embed_16).mean(1)
+        embed_24 = self.pre24_linear(embed_24).mean(1)
+        embed_last = self.last_linear(embedding).mean(1)
+        embed_trans = self.trans_linear(torch.cat([_s, _ori_v[...,None], _d], dim=2)).mean(1)
+        skip_embedding = self.skip_connection(self.skip_connection2(embedding).permute(0,2,1)).permute(0,2,1)
+        scale_skip, embed_skip = skip_embedding.chunk(2,dim=1)
+        
+        ori_noise = noise * (scale_skip[...,None]) + embed_skip[...,None]
+        noise = self.pos_embed(noise)        
+        noise = noise * (1 + scale_8[:, None, :] + embed_trans[:, None, :]) + embed_8[:, None, :]
+        scale_list = [scale_16, scale_24]
+        embed_list = [embed_16, embed_24]
+        for _ii, block in enumerate(self.transformer_blocks):
+            noise = block(noise)  
+            if len(scale_list)!=0 and len(embed_list)!=0:
+                noise = noise * (1 + scale_list[int(_ii//4)][:, None, :] + embed_trans[:, None, :]) + embed_list[int(_ii//4)][:, None, :]
+        
+        hidden_states = noise
+        hidden_states = self.norm_out(hidden_states, embed_last)
+        hidden_states = self.proj_out(hidden_states)
+
+        # unpatchify
+        patch_size = self.patch_size
+        height = height // patch_size
+        width = width // patch_size
+    
+        hidden_states = hidden_states.reshape(
+            shape=(hidden_states.shape[0], height, width, patch_size, patch_size, self.out_channels)
+        )
+        hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
+        output = hidden_states.reshape(
+            shape=(hidden_states.shape[0], self.out_channels, height * patch_size, width * patch_size)
+        )
+        return output + ori_noise
+
+    def weak_load_state_dict(self, state_dict: os.Mapping[str, torch.any], strict: bool = True, assign: bool = False):
+        return load_filtered_state_dict(self, state_dict)
+
+
+
+def load_filtered_state_dict(model, state_dict):
+    model_state_dict = model.state_dict()
+    filtered_state_dict = {}
+    for k, v in state_dict.items():
+        if k in model_state_dict:
+            if model_state_dict[k].size() == v.size():
+                filtered_state_dict[k] = v
+            else:
+                print(f"Skipping {k}: shape mismatch ({model_state_dict[k].size()} vs {v.size()})")
+        else:
+            print(f"Skipping {k}: not found in model's state_dict.")
+    model.load_state_dict(filtered_state_dict, strict=False)
+    return model
+
+def custom_collate_fn_2_0(batch):
+    noise_pred_texts, prompts, noise_preds, max_scores = zip(*batch)
+    
+    noise_pred_texts = torch.stack(noise_pred_texts)
+    noise_preds = torch.stack(noise_preds)
+    max_scores = torch.stack(max_scores)
+    
+    return noise_pred_texts, prompts, noise_preds, max_scores
+

requirements.txt

@@ -0,0 +1,14 @@
+diffusers
+transformers
+einops 
+wandb
+accelerate
+pandas
+imageio
+gradio
+imageio-ffmpeg
+omegaconf
+spaces
+torch==2.4 --index-url https://download.pytorch.org/whl/cu124
+torchaudio==2.4 --index-url https://download.pytorch.org/whl/cu124
+torchvision==0.19.0 --index-url https://download.pytorch.org/whl/cu124