generator device + seed_everything + custom_pipeline

app.py

@@ -21,6 +21,7 @@ import base64
 import os
 import time
 import re
+import random
 
 from transformers import (
     AutoTokenizer,
@@ -37,6 +38,15 @@ class StopAtSpecificTokenCriteria(StoppingCriteria):
     @add_start_docstrings(STOPPING_CRITERIA_INPUTS_DOCSTRING)
     def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
         return input_ids[0][-1].detach().cpu().numpy() in self.token_id_list
+    
+def seed_everything(seed):
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if torch.cuda.is_available():
+        torch.cuda.manual_seed(seed)
+        torch.cuda.manual_seed_all(seed)
+    torch.backends.cudnn.deterministic = True
 
 def ensure_space_after_period(input_string):
     # 去除多余的空格
@@ -232,13 +242,8 @@ def construction_all():
     global model
     global quantizer
     global tokenizer
-    global pipeline
-    global transp_vae
     from modeling_crello import CrelloModel, CrelloModelConfig
     from quantizer import get_quantizer
-    from custom_model_mmdit import CustomFluxTransformer2DModel
-    from custom_model_transp_vae import AutoencoderKLTransformerTraining as CustomVAE
-    from custom_pipeline import CustomFluxPipelineCfg
     
     params_dict = {  
         "input_model": "/openseg_blob/v-sirui/temporary/2024-02-21/Layout_train/COLEv2/Design_LLM/checkpoint/Meta-Llama-3-8B", 
@@ -314,13 +319,28 @@ def construction_all():
     for token in added_special_tokens_list:
         quantizer.additional_special_tokens.add(token)
         
+    global pipeline
+    global transp_vae
+    seed_everything(42)
+    from custom_model_mmdit import CustomFluxTransformer2DModel
+    from custom_pipeline import CustomFluxPipelineCfg
+    
     transformer = CustomFluxTransformer2DModel.from_pretrained(
         "WYBar/ART_test_weights",
         subfolder="fused_transformer", 
         torch_dtype=torch.bfloat16,
-        # cache_dir="/openseg_blob/v-yanbin/GradioDemo/cache_dir"
+        cache_dir="/openseg_blob/v-yanbin/GradioDemo/cache_dir"
     )
-    
+
+    pipeline = CustomFluxPipelineCfg.from_pretrained(
+        "black-forest-labs/FLUX.1-dev",
+        transformer=transformer,
+        torch_dtype=torch.bfloat16,
+        cache_dir="/openseg_blob/v-yanbin/GradioDemo/cache_dir",
+    ).to("cuda")
+    # pipeline.enable_model_cpu_offload(gpu_id=0) # save vram
+
+    from custom_model_transp_vae import AutoencoderKLTransformerTraining as CustomVAE
     transp_vae = CustomVAE.from_pretrained(
         "WYBar/ART_test_weights",
         subfolder="custom_vae", 
@@ -328,16 +348,7 @@ def construction_all():
         use_safetensors=True,
         # cache_dir="/openseg_blob/v-yanbin/GradioDemo/cache_dir"
     )
-    
-    token = os.environ.get("HF_TOKEN")
-    pipeline = CustomFluxPipelineCfg.from_pretrained(
-        "black-forest-labs/FLUX.1-dev",
-        transformer=transformer,
-        torch_dtype=torch.bfloat16,
-        token=token,
-        # cache_dir="/openseg_blob/v-yanbin/GradioDemo/cache_dir"
-    ).to("cuda")
-    pipeline.enable_model_cpu_offload(gpu_id=0) # Save GPU memory
+    transp_vae.eval()
         
     print(f"before .to(device):{model.device} {model.lm.device} {pipeline.device}")
     model = model.to("cuda")
@@ -421,6 +432,7 @@ def test_one_sample(validation_box, validation_prompt, true_gs, inference_steps,
         num_layers=len(validation_box),
         guidance_scale=4.0,
         num_inference_steps=inference_steps,
+        sdxl_vae=transp_vae,
         transparent_decoder=transp_vae,
         true_gs=true_gs
     )
@@ -440,25 +452,29 @@ def test_one_sample(validation_box, validation_prompt, true_gs, inference_steps,
 
 def svg_test_one_sample(validation_prompt, validation_box_str, seed, true_gs, inference_steps, pipeline, transp_vae):
     print(f"svg_test_one_sample {model.device} {model.lm.device} {pipeline.device}")
-    generator = torch.Generator().manual_seed(seed)
+    # generator = torch.Generator().manual_seed(seed)
+    generator = torch.Generator(device=torch.device("cuda", index=0)).manual_seed(seed)
     try:
-        validation_box = ast.literal_eval(validation_box_str)
+        if isinstance(validation_box_str, (list, tuple)):
+            validation_box = validation_box_str
+        else:
+            validation_box = ast.literal_eval(validation_box_str)
     except Exception as e:
         return [f"Error parsing validation_box: {e}"]
+
     if not isinstance(validation_box, list) or not all(isinstance(t, tuple) and len(t) == 4 for t in validation_box):
         return ["validation_box must be a list of tuples, each of length 4."]
-
     validation_box = adjust_validation_box(validation_box)
     
     print("result_images = test_one_sample")
     result_images = test_one_sample(validation_box, validation_prompt, true_gs, inference_steps, pipeline, generator, transp_vae)
     print("after result_images = test_one_sample")
     svg_img = pngs_to_svg(result_images[1:])
-    
+
     svg_file_path = './image.svg'
     os.makedirs(os.path.dirname(svg_file_path), exist_ok=True)
     with open(svg_file_path, 'w', encoding='utf-8') as f:
-        f.write(svg_img)    
+        f.write(svg_img)           
         
     if not isinstance(result_images, list):
         raise TypeError("result_images 必须是一个列表")
@@ -475,7 +491,7 @@ def svg_test_one_sample(validation_prompt, validation_box_str, seed, true_gs, in
 def process_svg(text_input, tuple_input, seed, true_gs, inference_steps):
     print(f"precess_svg {model.device} {model.lm.device} {pipeline.device}")
     result_images = []
-    result_images, svg_file_path = svg_test_one_sample(text_input, tuple_input, seed, true_gs, inference_steps, pipeline=pipeline, transp_vae=transp_vae)
+    result_images, svg_file_path = gradio_test_one_sample(text_input, tuple_input, seed, true_gs, inference_steps, pipeline=pipeline, transp_vae=transp_vae)
     # result_images, svg_file_path = gradio_test_one_sample_partial(text_input, tuple_input, seed, true_gs, inference_steps)
 
     url, unique_filename = upload_to_github(file_path=svg_file_path)
@@ -543,45 +559,6 @@ def main():
     construction_all()
     print(f"after construction_all:{model.device} {model.lm.device} {pipeline.device}")
     
-    # def process_preddate(intention, generate_method='v1'):
-    #     list_box = [(0, 0, 512, 512), (0, 0, 512, 512), (136, 184, 512, 512), (144, 0, 512, 512), (0, 0, 328, 136), (160, 112, 512, 360), (168, 112, 512, 360), (40, 232, 112, 296), (32, 88, 248, 176), (48, 424, 144, 448), (48, 464, 144, 488), (240, 464, 352, 488), (384, 464, 488, 488), (48, 480, 144, 504), (240, 480, 360, 504), (456, 0, 512, 56), (0, 0, 56, 40), (440, 0, 512, 40), (0, 24, 48, 88), (48, 168, 168, 240)]
-    #     wholecaption = "Design an engaging and vibrant recruitment advertisement for our company. The image should feature three animated characters in a modern cityscape, depicting a dynamic and collaborative work environment. Incorporate a light bulb graphic with a question mark, symbolizing innovation, creativity, and problem-solving. Use bold text to announce \"WE ARE RECRUITING\" and provide the company's social media handle \"@reallygreatsite\" and a contact phone number \"+123-456-7890\" for interested individuals. The overall design should be playful and youthful, attracting potential recruits who are innovative and eager to contribute to a lively team."
-    #     json_file = "/home/wyb/openseg_blob/v-yanbin/GradioDemo/LLM-For-Layout-Planning/inference_test.json"
-    #     return wholecaption, str(list_box), json_file
-
-    # pipeline, transp_vae = construction()
-
-    # gradio_test_one_sample_partial = partial(
-    #     svg_test_one_sample,
-    #     pipeline=pipeline,
-    #     transp_vae=transp_vae,
-    # )
-    
-    # def process_svg(text_input, tuple_input, seed, true_gs, inference_steps):
-    #     print("precess_svg")
-    #     result_images = []
-    #     result_images, svg_file_path = svg_test_one_sample(text_input, tuple_input, seed, true_gs, inference_steps, pipeline=pipeline, transp_vae=transp_vae)
-    #     # result_images, svg_file_path = gradio_test_one_sample_partial(text_input, tuple_input, seed, true_gs, inference_steps)
-
-    #     url, unique_filename = upload_to_github(file_path=svg_file_path)
-    #     unique_filename = f'{unique_filename}'
-        
-    #     if url != None:
-    #         print(f"File uploaded to: {url}")
-    #         svg_editor = f"""
-    #             <iframe src="https://svgedit.netlify.app/editor/index.html?\
-    #             storagePrompt=false&url={url}" \
-    #             width="100%", height="800px"></iframe>
-    #         """
-    #     else:
-    #         print('upload_to_github FAILED!')
-    #         svg_editor = f"""
-    #             <iframe src="https://svgedit.netlify.app/editor/index.html" \
-    #             width="100%", height="800px"></iframe>
-    #         """
-        
-    #     return result_images, svg_file_path, svg_editor
-    
     def one_click_generate(intention_input, temperature, top_p, seed, true_gs, inference_steps):
         # 首先调用process_preddate
         list_box_output, intention_input, list_box_output = process_preddate(intention_input, temperature, top_p)

custom_pipeline.py

@@ -1,16 +1,18 @@
+import torch
+import torch.nn as nn
 import numpy as np
+import math
 from typing import Any, Callable, Dict, List, Optional, Union
 
-import torch
-import torch.nn as nn
+from diffusers.pipelines.flux.pipeline_flux import calculate_shift, retrieve_timesteps, FluxPipeline
+from diffusers.pipelines.flux.pipeline_output import FluxPipelineOutput
 
-from diffusers.utils.torch_utils import randn_tensor
 from diffusers.utils import is_torch_xla_available, logging
-from diffusers.pipelines.flux.pipeline_output import FluxPipelineOutput
-from diffusers.pipelines.flux.pipeline_flux import calculate_shift, retrieve_timesteps, FluxPipeline
+from diffusers.utils.torch_utils import randn_tensor
 
 if is_torch_xla_available():
     import torch_xla.core.xla_model as xm # type: ignore
+
     XLA_AVAILABLE = True
 else:
     XLA_AVAILABLE = False
@@ -55,7 +57,6 @@ def _get_clip_prompt_embeds(
 
     return prompt_embeds
 
-
 def _get_t5_prompt_embeds(
     tokenizer,
     text_encoder,
@@ -111,6 +112,7 @@ def encode_prompt(
     prompt = [prompt] if isinstance(prompt, str) else prompt
     prompt_2 = prompt_2 or prompt
     prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2
+    batch_size = len(prompt)
 
     # We only use the pooled prompt output from the CLIPTextModel
     pooled_prompt_embeds = _get_clip_prompt_embeds(
@@ -469,73 +471,328 @@ class CustomFluxPipeline(FluxPipeline):
         return FluxPipelineOutput(images=image), result_list, vis_list
 
 
-class CustomFluxPipelineCfg(FluxPipeline):
+class CustomFluxPipelineCfg(CustomFluxPipeline):
 
-    @staticmethod
-    def _prepare_latent_image_ids(height, width, list_layer_box, device, dtype):
+    @torch.no_grad()
+    def __call__(
+        self,
+        prompt: Union[str, List[str]] = None,
+        prompt_2: Optional[Union[str, List[str]]] = None,
+        validation_box: List[tuple] = None,
+        height: Optional[int] = None,
+        width: Optional[int] = None,
+        num_inference_steps: int = 28,
+        timesteps: List[int] = None,
+        guidance_scale: float = 3.5,
+        true_gs: float = 3.5,
+        num_images_per_prompt: Optional[int] = 1,
+        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
+        latents: Optional[torch.FloatTensor] = None,
+        prompt_embeds: Optional[torch.FloatTensor] = None,
+        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
+        output_type: Optional[str] = "pil",
+        return_dict: bool = True,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+        callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
+        callback_on_step_end_tensor_inputs: List[str] = ["latents"],
+        max_sequence_length: int = 512,
+        num_layers: int = 5,
+        sdxl_vae: nn.Module = None,
+        transparent_decoder: nn.Module = None,
+    ):
+        r"""
+        Function invoked when calling the pipeline for generation.
 
-        latent_image_ids_list = []
-        for layer_idx in range(len(list_layer_box)):
-            if list_layer_box[layer_idx] == None:
-                continue
-            else:
-                latent_image_ids = torch.zeros(height // 2, width // 2, 3) # [h/2, w/2, 3]
-                latent_image_ids[..., 0] = layer_idx # use the first dimension for layer representation
-                latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height // 2)[:, None]
-                latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width // 2)[None, :]
+        Args:
+            prompt (`str` or `List[str]`, *optional*):
+                The prompt or prompts to guide the image generation. If not defined, one has to pass `prompt_embeds`.
+                instead.
+            prompt_2 (`str` or `List[str]`, *optional*):
+                The prompt or prompts to be sent to `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
+                will be used instead
+            height (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The height in pixels of the generated image. This is set to 1024 by default for the best results.
+            width (`int`, *optional*, defaults to self.unet.config.sample_size * self.vae_scale_factor):
+                The width in pixels of the generated image. This is set to 1024 by default for the best results.
+            num_inference_steps (`int`, *optional*, defaults to 50):
+                The number of denoising steps. More denoising steps usually lead to a higher quality image at the
+                expense of slower inference.
+            timesteps (`List[int]`, *optional*):
+                Custom timesteps to use for the denoising process with schedulers which support a `timesteps` argument
+                in their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is
+                passed will be used. Must be in descending order.
+            guidance_scale (`float`, *optional*, defaults to 7.0):
+                Guidance scale as defined in [Classifier-Free Diffusion Guidance](https://arxiv.org/abs/2207.12598).
+                `guidance_scale` is defined as `w` of equation 2. of [Imagen
+                Paper](https://arxiv.org/pdf/2205.11487.pdf). Guidance scale is enabled by setting `guidance_scale >
+                1`. Higher guidance scale encourages to generate images that are closely linked to the text `prompt`,
+                usually at the expense of lower image quality.
+            num_images_per_prompt (`int`, *optional*, defaults to 1):
+                The number of images to generate per prompt.
+            generator (`torch.Generator` or `List[torch.Generator]`, *optional*):
+                One or a list of [torch generator(s)](https://pytorch.org/docs/stable/generated/torch.Generator.html)
+                to make generation deterministic.
+            latents (`torch.FloatTensor`, *optional*):
+                Pre-generated noisy latents, sampled from a Gaussian distribution, to be used as inputs for image
+                generation. Can be used to tweak the same generation with different prompts. If not provided, a latents
+                tensor will ge generated by sampling using the supplied random `generator`.
+            prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not
+                provided, text embeddings will be generated from `prompt` input argument.
+            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
+                Pre-generated pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting.
+                If not provided, pooled text embeddings will be generated from `prompt` input argument.
+            output_type (`str`, *optional*, defaults to `"pil"`):
+                The output format of the generate image. Choose between
+                [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~pipelines.flux.FluxPipelineOutput`] instead of a plain tuple.
+            joint_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            callback_on_step_end (`Callable`, *optional*):
+                A function that calls at the end of each denoising steps during the inference. The function is called
+                with the following arguments: `callback_on_step_end(self: DiffusionPipeline, step: int, timestep: int,
+                callback_kwargs: Dict)`. `callback_kwargs` will include a list of all tensors as specified by
+                `callback_on_step_end_tensor_inputs`.
+            callback_on_step_end_tensor_inputs (`List`, *optional*):
+                The list of tensor inputs for the `callback_on_step_end` function. The tensors specified in the list
+                will be passed as `callback_kwargs` argument. You will only be able to include variables listed in the
+                `._callback_tensor_inputs` attribute of your pipeline class.
+            max_sequence_length (`int` defaults to 512): Maximum sequence length to use with the `prompt`.
 
-                x1, y1, x2, y2 = list_layer_box[layer_idx]
-                x1, y1, x2, y2 = x1 // 16, y1 // 16, x2 // 16, y2 // 16
-                latent_image_ids = latent_image_ids[y1:y2, x1:x2, :]
+        Examples:
 
-                latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape
-                latent_image_ids = latent_image_ids.reshape(
-                    latent_image_id_height * latent_image_id_width, latent_image_id_channels
-                )
+        Returns:
+            [`~pipelines.flux.FluxPipelineOutput`] or `tuple`: [`~pipelines.flux.FluxPipelineOutput`] if `return_dict`
+            is True, otherwise a `tuple`. When returning a tuple, the first element is a list with the generated
+            images.
+        """
 
-                latent_image_ids_list.append(latent_image_ids)
+        height = height or self.default_sample_size * self.vae_scale_factor
+        width = width or self.default_sample_size * self.vae_scale_factor
 
-        full_latent_image_ids = torch.cat(latent_image_ids_list, dim=0)
+        # 1. Check inputs. Raise error if not correct
+        self.check_inputs(
+            prompt,
+            prompt_2,
+            height,
+            width,
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
+            max_sequence_length=max_sequence_length,
+        )
 
-        return full_latent_image_ids.to(device=device, dtype=dtype)
+        self._guidance_scale = guidance_scale
+        self._joint_attention_kwargs = joint_attention_kwargs
+        self._interrupt = False
 
-    def prepare_latents(
-        self,
-        batch_size,
-        num_layers,
-        num_channels_latents,
-        height,
-        width,
-        list_layer_box,
-        dtype,
-        device,
-        generator,
-        latents=None,
-    ):
-        height = 2 * (int(height) // self.vae_scale_factor)
-        width = 2 * (int(width) // self.vae_scale_factor)
+        # 2. Define call parameters
+        if prompt is not None and isinstance(prompt, str):
+            batch_size = 1
+        elif prompt is not None and isinstance(prompt, list):
+            batch_size = len(prompt)
+        else:
+            batch_size = prompt_embeds.shape[0]
 
-        shape = (batch_size, num_layers, num_channels_latents, height, width) 
+        device = self._execution_device
 
-        if latents is not None:
-            latent_image_ids = self._prepare_latent_image_ids(batch_size, height, width, device, dtype)
-            return latents.to(device=device, dtype=dtype), latent_image_ids
+        lora_scale = (
+            self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None
+        )
+        (
+            prompt_embeds,
+            pooled_prompt_embeds,
+            text_ids,
+        ) = self.encode_prompt(
+            prompt=prompt,
+            prompt_2=prompt_2,
+            prompt_embeds=prompt_embeds,
+            pooled_prompt_embeds=pooled_prompt_embeds,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            max_sequence_length=max_sequence_length,
+            lora_scale=lora_scale,
+        )
+        (
+            neg_prompt_embeds,
+            neg_pooled_prompt_embeds,
+            neg_text_ids,
+        ) = self.encode_prompt(
+            prompt="",
+            prompt_2=None,
+            device=device,
+            num_images_per_prompt=num_images_per_prompt,
+            max_sequence_length=max_sequence_length,
+            lora_scale=lora_scale,
+        )
 
-        if isinstance(generator, list) and len(generator) != batch_size:
-            raise ValueError(
-                f"You have passed a list of generators of length {len(generator)}, but requested an effective batch"
-                f" size of {batch_size}. Make sure the batch size matches the length of the generators."
-            )
+        # 4. Prepare latent variables
+        num_channels_latents = self.transformer.config.in_channels // 4
+        latents, latent_image_ids = self.prepare_latents(
+            batch_size * num_images_per_prompt,
+            num_layers,
+            num_channels_latents,
+            height,
+            width,
+            validation_box,
+            prompt_embeds.dtype,
+            device,
+            generator,
+            latents,
+        )
 
-        latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype) # [bs, n_layers, c_latent, h, w]
+        # 5. Prepare timesteps
+        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
+        image_seq_len = latent_image_ids.shape[0] # ???
+        mu = calculate_shift(
+            image_seq_len,
+            self.scheduler.config.base_image_seq_len,
+            self.scheduler.config.max_image_seq_len,
+            self.scheduler.config.base_shift,
+            self.scheduler.config.max_shift,
+        )
+        timesteps, num_inference_steps = retrieve_timesteps(
+            self.scheduler,
+            num_inference_steps,
+            device,
+            timesteps,
+            sigmas,
+            mu=mu,
+        )
+        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
+        self._num_timesteps = len(timesteps)
 
-        latent_image_ids = self._prepare_latent_image_ids(height, width, list_layer_box, device, dtype)
+        # handle guidance
+        if self.transformer.config.guidance_embeds:
+            guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32)
+            guidance = guidance.expand(latents.shape[0])
+        else:
+            guidance = None
 
-        return latents, latent_image_ids
+        # 6. Denoising loop
+        with self.progress_bar(total=num_inference_steps) as progress_bar:
+            for i, t in enumerate(timesteps):
+                if self.interrupt:
+                    continue
+
+                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+                timestep = t.expand(latents.shape[0]).to(latents.dtype)
+
+                noise_pred = self.transformer(
+                    hidden_states=latents,
+                    list_layer_box=validation_box,
+                    timestep=timestep / 1000,
+                    guidance=guidance,
+                    pooled_projections=pooled_prompt_embeds,
+                    encoder_hidden_states=prompt_embeds,
+                    txt_ids=text_ids,
+                    img_ids=latent_image_ids,
+                    joint_attention_kwargs=self.joint_attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                neg_noise_pred = self.transformer(
+                    hidden_states=latents,
+                    list_layer_box=validation_box,
+                    timestep=timestep / 1000,
+                    guidance=guidance,
+                    pooled_projections=neg_pooled_prompt_embeds,
+                    encoder_hidden_states=neg_prompt_embeds,
+                    txt_ids=neg_text_ids,
+                    img_ids=latent_image_ids,
+                    joint_attention_kwargs=self.joint_attention_kwargs,
+                    return_dict=False,
+                )[0]
+
+                noise_pred = neg_noise_pred + true_gs * (noise_pred - neg_noise_pred)
+
+                # compute the previous noisy sample x_t -> x_t-1
+                latents_dtype = latents.dtype
+                latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
+
+                if latents.dtype != latents_dtype:
+                    if torch.backends.mps.is_available():
+                        # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
+                        latents = latents.to(latents_dtype)
+
+                if callback_on_step_end is not None:
+                    callback_kwargs = {}
+                    for k in callback_on_step_end_tensor_inputs:
+                        callback_kwargs[k] = locals()[k]
+                    callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
+
+                    latents = callback_outputs.pop("latents", latents)
+                    prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
+
+                # call the callback, if provided
+                if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
+                    progress_bar.update()
+
+                if XLA_AVAILABLE:
+                    xm.mark_step()
+
+        # create a grey latent
+        bs, n_frames, channel_latent, height, width = latents.shape
+
+        pixel_grey = torch.zeros(size=(bs*n_frames, 3, height*8, width*8), device=latents.device, dtype=latents.dtype)
+        latent_grey = self.vae.encode(pixel_grey).latent_dist.sample()
+        latent_grey = (latent_grey - self.vae.config.shift_factor) * self.vae.config.scaling_factor
+        latent_grey = latent_grey.view(bs, n_frames, channel_latent, height, width)  # [bs, f, c_latent, h, w]
+        
+        # fill in the latents
+        for layer_idx in range(latent_grey.shape[1]):
+            if validation_box[layer_idx] == None:
+                continue
+            x1, y1, x2, y2 = validation_box[layer_idx]
+            x1, y1, x2, y2 = x1 // 8, y1 // 8, x2 // 8, y2 // 8
+            latent_grey[:, layer_idx, :, y1:y2, x1:x2] = latents[:, layer_idx, :, y1:y2, x1:x2]
+        latents = latent_grey
+
+        if output_type == "latent":
+            image = latents
+
+        else:
+            latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
+            latents = latents.reshape(bs * n_frames, channel_latent, height, width)
+            latents_segs = torch.split(latents, 16, dim=0) ### split latents by 16 to avoid odd purple output
+            image_segs = [self.vae.decode(latents_seg, return_dict=False)[0] for latents_seg in latents_segs]
+            image = torch.cat(image_segs, dim=0)
+            if sdxl_vae is not None:
+                sdxl_vae = sdxl_vae.to(dtype=image.dtype, device=image.device)
+                
+                decoded_fg, decoded_alpha = sdxl_vae(latents, [validation_box])
+                decoded_alpha = (decoded_alpha + 1.0) / 2.0 #torch.Size([5, 1, 1024, 1024])
+                decoded_alpha = torch.clamp(decoded_alpha, min=0.0, max=1.0).permute(0, 2, 3, 1) #torch.Size([5, 1024, 1024, 1])
+
+                decoded_fg = (decoded_fg + 1.0) / 2.0
+                decoded_fg = torch.clamp(decoded_fg, min=0.0, max=1.0).permute(0, 2, 3, 1)#torch.Size([5, 1024, 1024, 3]))
+
+                vis_list = None
+                png = torch.cat([decoded_fg, decoded_alpha], dim=3)#[0] #torch.Size([1024, 1024, 4])
+                result_list = (png * 255.0).detach().cpu().float().numpy().clip(0, 255).astype(np.uint8)
+            else:
+                result_list, vis_list = None, None
+            image = self.image_processor.postprocess(image, output_type=output_type)
+
+        # Offload all models
+        self.maybe_free_model_hooks()
+
+        if not return_dict:
+            return (image, result_list, vis_list, latents)
+
+        return FluxPipelineOutput(images=image), result_list, vis_list, latents
+
+
+class CustomFluxPipelineCfgInpaint(CustomFluxPipeline):
 
     @torch.no_grad()
     def __call__(
         self,
+        image: Optional[List[torch.FloatTensor]] = None,
+        mask: Optional[torch.FloatTensor] = None,
         prompt: Union[str, List[str]] = None,
         prompt_2: Optional[Union[str, List[str]]] = None,
         validation_box: List[tuple] = None,
@@ -557,6 +814,7 @@ class CustomFluxPipelineCfg(FluxPipeline):
         callback_on_step_end_tensor_inputs: List[str] = ["latents"],
         max_sequence_length: int = 512,
         num_layers: int = 5,
+        sdxl_vae: nn.Module = None,
         transparent_decoder: nn.Module = None,
     ):
         r"""
@@ -703,9 +961,22 @@ class CustomFluxPipelineCfg(FluxPipeline):
             latents,
         )
 
+        # 4.1. Prepare image and mask
+        merged_pt, backgd_pt, list_layer_pt = image[0], image[1], image[2:]
+        # prepare RGB, Alpha
+        layer_pt_grey = [layer_pt[:, :3] * ((layer_pt[:, 3:4] + 1) / 2.) for layer_pt in list_layer_pt]
+        pixel_values_vae_input = torch.cat([merged_pt, backgd_pt] + layer_pt_grey, dim=0).to(device, dtype=self.vae.dtype)  # [bs*(l+2), c_img, H, W]
+        # Convert images to latent space
+        model_input = self.vae.encode(pixel_values_vae_input).latent_dist.sample()
+        model_input = (model_input - self.vae.config.shift_factor) * self.vae.config.scaling_factor
+        model_input = model_input.reshape(1, len(validation_box), model_input.shape[1], model_input.shape[2], model_input.shape[3])  # [bs, f, c_latent, h, w]
+        # copy latent and noise
+        orig_latents = model_input
+        noise = latents.clone()
+
         # 5. Prepare timesteps
         sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
-        image_seq_len = latent_image_ids.shape[0]
+        image_seq_len = latent_image_ids.shape[0] # ???
         mu = calculate_shift(
             image_seq_len,
             self.scheduler.config.base_image_seq_len,
@@ -772,6 +1043,16 @@ class CustomFluxPipelineCfg(FluxPipeline):
                 latents_dtype = latents.dtype
                 latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0]
 
+                # blend the latents with the original image
+                init_latents_proper = orig_latents.to(latents.dtype)
+                init_mask = mask.reshape(1, -1, 1, 1, 1).to(latents.dtype)
+                if i < len(timesteps) - 1:
+                    noise_timestep = timesteps[i + 1]
+                    init_latents_proper = self.scheduler.scale_noise(
+                        init_latents_proper, torch.tensor([noise_timestep]), noise
+                    )
+                latents = (1 - init_mask) * init_latents_proper + init_mask * latents
+
                 if latents.dtype != latents_dtype:
                     if torch.backends.mps.is_available():
                         # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272
@@ -794,12 +1075,12 @@ class CustomFluxPipelineCfg(FluxPipeline):
                     xm.mark_step()
 
         # create a grey latent
-        bs, n_layers, channel_latent, height, width = latents.shape
+        bs, n_frames, channel_latent, height, width = latents.shape
 
-        pixel_grey = torch.zeros(size=(bs*n_layers, 3, height*8, width*8), device=latents.device, dtype=latents.dtype)
+        pixel_grey = torch.zeros(size=(bs*n_frames, 3, height*8, width*8), device=latents.device, dtype=latents.dtype)
         latent_grey = self.vae.encode(pixel_grey).latent_dist.sample()
         latent_grey = (latent_grey - self.vae.config.shift_factor) * self.vae.config.scaling_factor
-        latent_grey = latent_grey.view(bs, n_layers, channel_latent, height, width)  # [bs, n_layers, c_latent, h, w]
+        latent_grey = latent_grey.view(bs, n_frames, channel_latent, height, width)  # [bs, f, c_latent, h, w]
         
         # fill in the latents
         for layer_idx in range(latent_grey.shape[1]):
@@ -815,22 +1096,22 @@ class CustomFluxPipelineCfg(FluxPipeline):
 
         else:
             latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
-            latents = latents.reshape(bs * n_layers, channel_latent, height, width)
+            latents = latents.reshape(bs * n_frames, channel_latent, height, width)
             latents_segs = torch.split(latents, 16, dim=0) ### split latents by 16 to avoid odd purple output
             image_segs = [self.vae.decode(latents_seg, return_dict=False)[0] for latents_seg in latents_segs]
             image = torch.cat(image_segs, dim=0)
-            if transparent_decoder is not None:
-                transparent_decoder = transparent_decoder.to(dtype=image.dtype, device=image.device)
+            if sdxl_vae is not None:
+                sdxl_vae = sdxl_vae.to(dtype=image.dtype, device=image.device)
                 
-                decoded_fg, decoded_alpha = transparent_decoder(latents, [validation_box])
-                decoded_alpha = (decoded_alpha + 1.0) / 2.0
-                decoded_alpha = torch.clamp(decoded_alpha, min=0.0, max=1.0).permute(0, 2, 3, 1)
+                decoded_fg, decoded_alpha = sdxl_vae(latents, [validation_box])
+                decoded_alpha = (decoded_alpha + 1.0) / 2.0 #torch.Size([5, 1, 1024, 1024])
+                decoded_alpha = torch.clamp(decoded_alpha, min=0.0, max=1.0).permute(0, 2, 3, 1) #torch.Size([5, 1024, 1024, 1])
 
                 decoded_fg = (decoded_fg + 1.0) / 2.0
-                decoded_fg = torch.clamp(decoded_fg, min=0.0, max=1.0).permute(0, 2, 3, 1)
+                decoded_fg = torch.clamp(decoded_fg, min=0.0, max=1.0).permute(0, 2, 3, 1)#torch.Size([5, 1024, 1024, 3]))
 
                 vis_list = None
-                png = torch.cat([decoded_fg, decoded_alpha], dim=3)
+                png = torch.cat([decoded_fg, decoded_alpha], dim=3)#[0] #torch.Size([1024, 1024, 4])
                 result_list = (png * 255.0).detach().cpu().float().numpy().clip(0, 255).astype(np.uint8)
             else:
                 result_list, vis_list = None, None

requirements.txt

@@ -43,4 +43,5 @@ pynvml==11.5.3                 # 新增明确版本（conda实际安装11.5.3）
 colorama==0.4.6                # 新增明确版本（conda实际安装0.4.6）
 click>=8.0.4,<9                # 保持约束（conda实际安装8.1.7符合要求）\
 
-sentencepiece
+sentencepiece
+random