flux_concept_attention/flux_with_concept_attention_pipeline.py

"""
    Here we make various wrapper classes for the FluxPipeline from diffusers
    to add the concept attention functionality. 

    We opt for a wrapper functionality
"""
import inspect

import torch
import numpy as np
from typing import List, Union, Optional, Dict, Any, Callable
import PIL.Image
import einops
import matplotlib.pyplot as plt

from diffusers import DiffusionPipeline
from diffusers.image_processor import PipelineImageInput
from diffusers.pipelines.flux.pipeline_flux import retrieve_timesteps, calculate_shift
from diffusers.utils import is_torch_xla_available, BaseOutput, logging, USE_PEFT_BACKEND, \
     scale_lora_layers, unscale_lora_layers

from diffusers.utils.torch_utils import randn_tensor

from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
from diffusers.loaders import FluxIPAdapterMixin, FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin
from diffusers.models.autoencoders import AutoencoderKL
from diffusers.models.transformers import FluxTransformer2DModel
from diffusers.schedulers import FlowMatchEulerDiscreteScheduler

from transformers import (
    CLIPImageProcessor,
    CLIPTextModel,
    CLIPTokenizer,
    CLIPVisionModelWithProjection,
    T5EncoderModel,
    T5TokenizerFast,
)

if is_torch_xla_available():
    import torch_xla.core.xla_model as xm

    XLA_AVAILABLE = True
else:
    XLA_AVAILABLE = False


logger = logging.get_logger(__name__)  # pylint: disable=invalid-name


def retrieve_timesteps(
        scheduler,
        num_inference_steps: Optional[int] = None,
        device: Optional[Union[str, torch.device]] = None,
        timesteps: Optional[List[int]] = None,
        sigmas: Optional[List[float]] = None,
        **kwargs,
):
    """
    Calls the scheduler's `set_timesteps` method and retrieves timesteps from the scheduler after the call. Handles
    custom timesteps. Any kwargs will be supplied to `scheduler.set_timesteps`.

    Args:
        scheduler (`SchedulerMixin`):
            The scheduler to get timesteps from.
        num_inference_steps (`int`):
            The number of diffusion steps used when generating samples with a pre-trained model. If used, `timesteps`
            must be `None`.
        device (`str` or `torch.device`, *optional*):
            The device to which the timesteps should be moved to. If `None`, the timesteps are not moved.
        timesteps (`List[int]`, *optional*):
            Custom timesteps used to override the timestep spacing strategy of the scheduler. If `timesteps` is passed,
            `num_inference_steps` and `sigmas` must be `None`.
        sigmas (`List[float]`, *optional*):
            Custom sigmas used to override the timestep spacing strategy of the scheduler. If `sigmas` is passed,
            `num_inference_steps` and `timesteps` must be `None`.

    Returns:
        `Tuple[torch.Tensor, int]`: A tuple where the first element is the timestep schedule from the scheduler and the
        second element is the number of inference steps.
    """
    if timesteps is not None and sigmas is not None:
        raise ValueError("Only one of `timesteps` or `sigmas` can be passed. Please choose one to set custom values")
    if timesteps is not None:
        accepts_timesteps = "timesteps" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
        if not accepts_timesteps:
            raise ValueError(
                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
                f" timestep schedules. Please check whether you are using the correct scheduler."
            )
        scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
        timesteps = scheduler.timesteps
        num_inference_steps = len(timesteps)
    elif sigmas is not None:
        accept_sigmas = "sigmas" in set(inspect.signature(scheduler.set_timesteps).parameters.keys())
        if not accept_sigmas:
            raise ValueError(
                f"The current scheduler class {scheduler.__class__}'s `set_timesteps` does not support custom"
                f" sigmas schedules. Please check whether you are using the correct scheduler."
            )
        scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
        timesteps = scheduler.timesteps
        num_inference_steps = len(timesteps)
    else:
        scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
        timesteps = scheduler.timesteps
    return timesteps, num_inference_steps


def retrieve_latents(
    encoder_output: torch.Tensor, generator: Optional[torch.Generator] = None, sample_mode: str = "sample"
):
    if hasattr(encoder_output, "latent_dist") and sample_mode == "sample":
        return encoder_output.latent_dist.sample(generator)
    elif hasattr(encoder_output, "latent_dist") and sample_mode == "argmax":
        return encoder_output.latent_dist.mode()
    elif hasattr(encoder_output, "latents"):
        return encoder_output.latents
    else:
        raise AttributeError("Could not access latents of provided encoder_output")


class FluxConceptAttentionOutput(BaseOutput):
    """
    Output class for the FluxPipeline with concept attention functionality.

    Args:
        images (`List[PIL.Image.Image]` or `np.ndarray`)
            The generated images.
        concept_attention_maps (`List[PIL.Image.Image]` or `np.ndarray`)
            The concept attention maps.
    """
    images: Union[List[PIL.Image.Image], np.ndarray]
    concept_attention_maps: Union[List[PIL.Image.Image], np.ndarray]

class FluxWithConceptAttentionPipeline(
    DiffusionPipeline,
    FluxLoraLoaderMixin,
    FromSingleFileMixin,
    TextualInversionLoaderMixin,
    FluxIPAdapterMixin,
):
    r"""
    The Flux pipeline for text-to-image generation with added Concept Attention. 

    Reference: https://blackforestlabs.ai/announcing-black-forest-labs/

    Args:
        transformer ([`FluxTransformer2DModel`]):
            Conditional Transformer (MMDiT) architecture to denoise the encoded image latents.
        scheduler ([`FlowMatchEulerDiscreteScheduler`]):
            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
        vae ([`AutoencoderKL`]):
            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
        text_encoder ([`CLIPTextModel`]):
            [CLIP](https://huggingface.co/docs/transformers/model_doc/clip#transformers.CLIPTextModel), specifically
            the [clip-vit-large-patch14](https://huggingface.co/openai/clip-vit-large-patch14) variant.
        text_encoder_2 ([`T5EncoderModel`]):
            [T5](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5EncoderModel), specifically
            the [google/t5-v1_1-xxl](https://huggingface.co/google/t5-v1_1-xxl) variant.
        tokenizer (`CLIPTokenizer`):
            Tokenizer of class
            [CLIPTokenizer](https://huggingface.co/docs/transformers/en/model_doc/clip#transformers.CLIPTokenizer).
        tokenizer_2 (`T5TokenizerFast`):
            Second Tokenizer of class
            [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast).
    """

    model_cpu_offload_seq = "text_encoder->text_encoder_2->image_encoder->transformer->vae"
    _optional_components = ["image_encoder", "feature_extractor"]
    _callback_tensor_inputs = ["latents", "prompt_embeds"]

    def __init__(
        self,
        scheduler: FlowMatchEulerDiscreteScheduler,
        vae: AutoencoderKL,
        text_encoder: CLIPTextModel,
        tokenizer: CLIPTokenizer,
        text_encoder_2: T5EncoderModel,
        tokenizer_2: T5TokenizerFast,
        transformer: FluxTransformer2DModel,
        image_encoder: CLIPVisionModelWithProjection = None,
        feature_extractor: CLIPImageProcessor = None,
    ):
        super().__init__()

        self.register_modules(
            vae=vae,
            text_encoder=text_encoder,
            text_encoder_2=text_encoder_2,
            tokenizer=tokenizer,
            tokenizer_2=tokenizer_2,
            transformer=transformer,
            scheduler=scheduler,
            image_encoder=image_encoder,
            feature_extractor=feature_extractor,
        )
        self.vae_scale_factor = (
            2 ** (len(self.vae.config.block_out_channels)) if hasattr(self, "vae") and self.vae is not None else 16
        )
        self.latent_channels = self.vae.config.latent_channels if getattr(self, "vae", None) else 16
        self.image_processor = VaeImageProcessor(
            vae_scale_factor=self.vae_scale_factor * 2, vae_latent_channels=self.latent_channels
        )
        self.tokenizer_max_length = (
            self.tokenizer.model_max_length if hasattr(self, "tokenizer") and self.tokenizer is not None else 77
        )
        self.default_sample_size = 64
        self.latent_width = self.default_sample_size
        self.latent_height = self.default_sample_size

    def _get_t5_prompt_embeds(
        self,
        prompt: Union[str, List[str]] = None,
        num_images_per_prompt: int = 1,
        max_sequence_length: int = 512,
        device: Optional[torch.device] = None,
        dtype: Optional[torch.dtype] = None,
    ):
        device = device or self._execution_device
        dtype = dtype or self.text_encoder.dtype

        prompt = [prompt] if isinstance(prompt, str) else prompt
        batch_size = len(prompt)

        if isinstance(self, TextualInversionLoaderMixin):
            prompt = self.maybe_convert_prompt(prompt, self.tokenizer_2)

        text_inputs = self.tokenizer_2(
            prompt,
            padding="max_length",
            max_length=max_sequence_length,
            truncation=True,
            return_length=False,
            return_overflowing_tokens=False,
            return_tensors="pt",
        )
        text_input_ids = text_inputs.input_ids
        untruncated_ids = self.tokenizer_2(prompt, padding="longest", return_tensors="pt").input_ids

        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
            removed_text = self.tokenizer_2.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1])
            logger.warning(
                "The following part of your input was truncated because `max_sequence_length` is set to "
                f" {max_sequence_length} tokens: {removed_text}"
            )

        prompt_embeds = self.text_encoder_2(text_input_ids.to(device), output_hidden_states=False)[0]

        dtype = self.text_encoder_2.dtype
        prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)

        _, seq_len, _ = prompt_embeds.shape

        # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)

        return prompt_embeds

    def _get_clip_prompt_embeds(
        self,
        prompt: Union[str, List[str]],
        num_images_per_prompt: int = 1,
        device: Optional[torch.device] = None,
    ):
        device = device or self._execution_device

        prompt = [prompt] if isinstance(prompt, str) else prompt
        batch_size = len(prompt)

        if isinstance(self, TextualInversionLoaderMixin):
            prompt = self.maybe_convert_prompt(prompt, self.tokenizer)

        text_inputs = self.tokenizer(
            prompt,
            padding="max_length",
            max_length=self.tokenizer_max_length,
            truncation=True,
            return_overflowing_tokens=False,
            return_length=False,
            return_tensors="pt",
        )

        text_input_ids = text_inputs.input_ids
        untruncated_ids = self.tokenizer(prompt, padding="longest", return_tensors="pt").input_ids
        if untruncated_ids.shape[-1] >= text_input_ids.shape[-1] and not torch.equal(text_input_ids, untruncated_ids):
            removed_text = self.tokenizer.batch_decode(untruncated_ids[:, self.tokenizer_max_length - 1 : -1])
            logger.warning(
                "The following part of your input was truncated because CLIP can only handle sequences up to"
                f" {self.tokenizer_max_length} tokens: {removed_text}"
            )
        prompt_embeds = self.text_encoder(text_input_ids.to(device), output_hidden_states=False)

        # Use pooled output of CLIPTextModel
        prompt_embeds = prompt_embeds.pooler_output
        prompt_embeds = prompt_embeds.to(dtype=self.text_encoder.dtype, device=device)

        # duplicate text embeddings for each generation per prompt, using mps friendly method
        prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt)
        prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1)

        return prompt_embeds

    def encode_prompt(
        self,
        prompt: Union[str, List[str]],
        prompt_2: Union[str, List[str]],
        device: Optional[torch.device] = None,
        num_images_per_prompt: int = 1,
        prompt_embeds: Optional[torch.FloatTensor] = None,
        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        max_sequence_length: int = 512,
        lora_scale: Optional[float] = None,
    ):
        r"""

        Args:
            prompt (`str` or `List[str]`, *optional*):
                prompt to be encoded
            prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts to be sent to the `tokenizer_2` and `text_encoder_2`. If not defined, `prompt` is
                used in all text-encoders
            device: (`torch.device`):
                torch device
            num_images_per_prompt (`int`):
                number of images that should be generated per prompt
            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.
            lora_scale (`float`, *optional*):
                A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
        """
        device = device or self._execution_device

        # set lora scale so that monkey patched LoRA
        # function of text encoder can correctly access it
        if lora_scale is not None and isinstance(self, FluxLoraLoaderMixin):
            self._lora_scale = lora_scale

            # dynamically adjust the LoRA scale
            if self.text_encoder is not None and USE_PEFT_BACKEND:
                scale_lora_layers(self.text_encoder, lora_scale)
            if self.text_encoder_2 is not None and USE_PEFT_BACKEND:
                scale_lora_layers(self.text_encoder_2, lora_scale)

        prompt = [prompt] if isinstance(prompt, str) else prompt

        if prompt_embeds is None:
            prompt_2 = prompt_2 or prompt
            prompt_2 = [prompt_2] if isinstance(prompt_2, str) else prompt_2

            # We only use the pooled prompt output from the CLIPTextModel
            pooled_prompt_embeds = self._get_clip_prompt_embeds(
                prompt=prompt,
                device=device,
                num_images_per_prompt=num_images_per_prompt,
            )
            prompt_embeds = self._get_t5_prompt_embeds(
                prompt=prompt_2,
                num_images_per_prompt=num_images_per_prompt,
                max_sequence_length=max_sequence_length,
                device=device,
            )

        if self.text_encoder is not None:
            if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND:
                # Retrieve the original scale by scaling back the LoRA layers
                unscale_lora_layers(self.text_encoder, lora_scale)

        if self.text_encoder_2 is not None:
            if isinstance(self, FluxLoraLoaderMixin) and USE_PEFT_BACKEND:
                # Retrieve the original scale by scaling back the LoRA layers
                unscale_lora_layers(self.text_encoder_2, lora_scale)

        dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype
        text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)

        return prompt_embeds, pooled_prompt_embeds, text_ids
    
    def encode_concepts(self, concepts: List[str], device: Optional[torch.device] = None):
        """
            Encodes our concept vectors using the T5 Encoder.  
        """
        """
            # Utils for concept encoding
            def embed_concepts(
                clip,
                t5,
                concepts: list[str],
                batch_size=1
            ):
                # Code pulled from concept_attention.flux/sampling.py: prepare()
                # Embed each concept separately
                concept_embeddings = []
                for concept in concepts:
                    concept_embedding = t5(concept)
                    # Pull out the first token
                    token_embedding = concept_embedding[0, 0, :] # First token of first prompt
                    concept_embeddings.append(token_embedding)
                concept_embeddings = torch.stack(concept_embeddings).unsqueeze(0)
                # Add filler tokens of zeros
                concept_ids = torch.zeros(batch_size, concept_embeddings.shape[1], 3)

                # Embed the concepts to a clip vector
                prompt = " ".join(concepts)
                vec = clip(prompt)
                vec = torch.zeros_like(vec).to(vec.device)

                return concept_embeddings, concept_ids, vec
        """

        concept_embeds = self._get_t5_prompt_embeds(
            prompt=concepts,
            num_images_per_prompt=1,
            max_sequence_length=64,
            device=device,
        )
        # Pull out the first token of each embedded concept to get the concept embeddings
        concept_embeds = concept_embeds[:, 0, :]
        concept_embeds = concept_embeds.unsqueeze(0)
        # Make the CLIP vector for the concepts 
        clip_vec = self._get_clip_prompt_embeds(
            prompt=" ".join(concepts),
            num_images_per_prompt=1,
            device=device,
        )   
        # # Set the vec to zero
        # clip_vec = torch.zeros_like(clip_vec).to(clip_vec.device)
        # # Add filler tokens of zeros
        concept_ids = torch.zeros(concept_embeds.shape[1], 3).to(device=device, dtype=concept_embeds.dtype)

        return concept_embeds, clip_vec, concept_ids

    def encode_image(self, image, device, num_images_per_prompt):
        dtype = next(self.image_encoder.parameters()).dtype

        if not isinstance(image, torch.Tensor):
            image = self.feature_extractor(image, return_tensors="pt").pixel_values

        image = image.to(device=device, dtype=dtype)
        image_embeds = self.image_encoder(image).image_embeds
        image_embeds = image_embeds.repeat_interleave(num_images_per_prompt, dim=0)
        return image_embeds

    def prepare_ip_adapter_image_embeds(
        self, ip_adapter_image, ip_adapter_image_embeds, device, num_images_per_prompt
    ):
        image_embeds = []
        if ip_adapter_image_embeds is None:
            if not isinstance(ip_adapter_image, list):
                ip_adapter_image = [ip_adapter_image]

            if len(ip_adapter_image) != len(self.transformer.encoder_hid_proj.image_projection_layers):
                raise ValueError(
                    f"`ip_adapter_image` must have same length as the number of IP Adapters. Got {len(ip_adapter_image)} images and {len(self.transformer.encoder_hid_proj.image_projection_layers)} IP Adapters."
                )

            for single_ip_adapter_image, image_proj_layer in zip(
                ip_adapter_image, self.transformer.encoder_hid_proj.image_projection_layers
            ):
                single_image_embeds = self.encode_image(single_ip_adapter_image, device, 1)

                image_embeds.append(single_image_embeds[None, :])
        else:
            for single_image_embeds in ip_adapter_image_embeds:
                image_embeds.append(single_image_embeds)

        ip_adapter_image_embeds = []
        for i, single_image_embeds in enumerate(image_embeds):
            single_image_embeds = torch.cat([single_image_embeds] * num_images_per_prompt, dim=0)
            single_image_embeds = single_image_embeds.to(device=device)
            ip_adapter_image_embeds.append(single_image_embeds)

        return ip_adapter_image_embeds

    def check_inputs(
        self,
        prompt,
        prompt_2,
        height,
        width,
        negative_prompt=None,
        negative_prompt_2=None,
        prompt_embeds=None,
        negative_prompt_embeds=None,
        pooled_prompt_embeds=None,
        negative_pooled_prompt_embeds=None,
        callback_on_step_end_tensor_inputs=None,
        max_sequence_length=None,
    ):
        if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0:
            logger.warning(
                f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly"
            )

        if callback_on_step_end_tensor_inputs is not None and not all(
            k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs
        ):
            raise ValueError(
                f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}"
            )

        if prompt is not None and prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
                " only forward one of the two."
            )
        elif prompt_2 is not None and prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `prompt_2`: {prompt_2} and `prompt_embeds`: {prompt_embeds}. Please make sure to"
                " only forward one of the two."
            )
        elif prompt is None and prompt_embeds is None:
            raise ValueError(
                "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined."
            )
        elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)):
            raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}")
        elif prompt_2 is not None and (not isinstance(prompt_2, str) and not isinstance(prompt_2, list)):
            raise ValueError(f"`prompt_2` has to be of type `str` or `list` but is {type(prompt_2)}")

        if negative_prompt is not None and negative_prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `negative_prompt`: {negative_prompt} and `negative_prompt_embeds`:"
                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
            )
        elif negative_prompt_2 is not None and negative_prompt_embeds is not None:
            raise ValueError(
                f"Cannot forward both `negative_prompt_2`: {negative_prompt_2} and `negative_prompt_embeds`:"
                f" {negative_prompt_embeds}. Please make sure to only forward one of the two."
            )

        if prompt_embeds is not None and negative_prompt_embeds is not None:
            if prompt_embeds.shape != negative_prompt_embeds.shape:
                raise ValueError(
                    "`prompt_embeds` and `negative_prompt_embeds` must have the same shape when passed directly, but"
                    f" got: `prompt_embeds` {prompt_embeds.shape} != `negative_prompt_embeds`"
                    f" {negative_prompt_embeds.shape}."
                )

        if prompt_embeds is not None and pooled_prompt_embeds is None:
            raise ValueError(
                "If `prompt_embeds` are provided, `pooled_prompt_embeds` also have to be passed. Make sure to generate `pooled_prompt_embeds` from the same text encoder that was used to generate `prompt_embeds`."
            )
        if negative_prompt_embeds is not None and negative_pooled_prompt_embeds is None:
            raise ValueError(
                "If `negative_prompt_embeds` are provided, `negative_pooled_prompt_embeds` also have to be passed. Make sure to generate `negative_pooled_prompt_embeds` from the same text encoder that was used to generate `negative_prompt_embeds`."
            )

        if max_sequence_length is not None and max_sequence_length > 512:
            raise ValueError(f"`max_sequence_length` cannot be greater than 512 but is {max_sequence_length}")

    @staticmethod
    def _prepare_latent_image_ids(batch_size, height, width, device, dtype):
        latent_image_ids = torch.zeros(height // 2, width // 2, 3)
        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, :]

        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
        )

        return latent_image_ids.to(device=device, dtype=dtype)

    @staticmethod
    def _pack_latents(latents, batch_size, num_channels_latents, height, width):
        latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2)
        latents = latents.permute(0, 2, 4, 1, 3, 5)
        latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4)

        return latents

    @staticmethod
    def _unpack_latents(latents, height, width, vae_scale_factor):
        batch_size, num_patches, channels = latents.shape

        height = height // vae_scale_factor
        width = width // vae_scale_factor

        latents = latents.view(batch_size, height, width, channels // 4, 2, 2)
        latents = latents.permute(0, 3, 1, 4, 2, 5)

        latents = latents.reshape(batch_size, channels // (2 * 2), height * 2, width * 2)

        return latents
    
    def enable_vae_slicing(self):
        r"""
        Enable sliced VAE decoding. When this option is enabled, the VAE will split the input tensor in slices to
        compute decoding in several steps. This is useful to save some memory and allow larger batch sizes.
        """
        self.vae.enable_slicing()

    def disable_vae_slicing(self):
        r"""
        Disable sliced VAE decoding. If `enable_vae_slicing` was previously enabled, this method will go back to
        computing decoding in one step.
        """
        self.vae.disable_slicing()

    def enable_vae_tiling(self):
        r"""
        Enable tiled VAE decoding. When this option is enabled, the VAE will split the input tensor into tiles to
        compute decoding and encoding in several steps. This is useful for saving a large amount of memory and to allow
        processing larger images.
        """
        self.vae.enable_tiling()

    def disable_vae_tiling(self):
        r"""
        Disable tiled VAE decoding. If `enable_vae_tiling` was previously enabled, this method will go back to
        computing decoding in one step.
        """
        self.vae.disable_tiling()

    def _encode_vae_image(self, image: torch.Tensor, generator: torch.Generator):
        if isinstance(generator, list):
            image_latents = [
                retrieve_latents(self.vae.encode(image[i: i + 1]), generator=generator[i])
                for i in range(image.shape[0])
            ]
            image_latents = torch.cat(image_latents, dim=0)
        else:
            image_latents = retrieve_latents(self.vae.encode(image), generator=generator)

        image_latents = (image_latents - self.vae.config.shift_factor) * self.vae.config.scaling_factor

        return image_latents

    def prepare_latents(
        self,
        image,
        timestep,
        batch_size,
        num_channels_latents,
        height,
        width,
        dtype,
        device,
        generator,
        latents=None,
    ):
        height = 2 * (int(height) // self.vae_scale_factor)
        width = 2 * (int(width) // self.vae_scale_factor)

        shape = (batch_size, num_channels_latents, height, width)

        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

        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."
            )
        
        if image is None:
            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
            latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width)

            latent_image_ids = self._prepare_latent_image_ids(batch_size, height, width, device, dtype)

            return latents, latent_image_ids

        shape = (batch_size, num_channels_latents, height, width)
        latent_image_ids = self._prepare_latent_image_ids(batch_size, height, width, device, dtype)
        image = image.to(device=device, dtype=dtype)
        if image.shape[1] != self.latent_channels:
            image_latents = self._encode_vae_image(image=image, generator=generator)
        else:
            image_latents = image
        if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0:
            # expand init_latents for batch_size
            additional_image_per_prompt = batch_size // image_latents.shape[0]
            image_latents = torch.cat([image_latents] * additional_image_per_prompt, dim=0)
        elif batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] != 0:
            raise ValueError(
                f"Cannot duplicate `image` of batch size {image_latents.shape[0]} to {batch_size} text prompts."
            )
        else:
            image_latents = torch.cat([image_latents], dim=0)

        noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
        latents = self.scheduler.scale_noise(image_latents, timestep, noise)
        latents = self._pack_latents(latents, batch_size, num_channels_latents, height, width)
        return latents, latent_image_ids

    @property
    def guidance_scale(self):
        return self._guidance_scale

    @property
    def joint_attention_kwargs(self):
        return self._joint_attention_kwargs

    @property
    def num_timesteps(self):
        return self._num_timesteps

    @property
    def interrupt(self):
        return self._interrupt

    @torch.no_grad()
    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
        prompt_2: Optional[Union[str, List[str]]] = None,
        negative_prompt: Union[str, List[str]] = None,
        negative_prompt_2: Optional[Union[str, List[str]]] = None,
        true_cfg_scale: float = 1.0,
        height: Optional[int] = None,
        width: Optional[int] = None,
        image: PipelineImageInput = None,
        timesteps: List[int] = None,
        num_inference_steps: int = 28,
        sigmas: Optional[List[float]] = None,
        guidance_scale: 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,
        ip_adapter_image: Optional[PipelineImageInput] = None,
        ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
        negative_ip_adapter_image: Optional[PipelineImageInput] = None,
        negative_ip_adapter_image_embeds: Optional[List[torch.Tensor]] = None,
        negative_prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
        concept_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,
    ):
        r"""
        Function invoked when calling the pipeline for generation.

        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.
            negative_prompt (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation. If not defined, one has to pass
                `negative_prompt_embeds` instead. Ignored when not using guidance (i.e., ignored if `true_cfg_scale` is
                not greater than `1`).
            negative_prompt_2 (`str` or `List[str]`, *optional*):
                The prompt or prompts not to guide the image generation to be sent to `tokenizer_2` and
                `text_encoder_2`. If not defined, `negative_prompt` is used in all the text-encoders.
            true_cfg_scale (`float`, *optional*, defaults to 1.0):
                When > 1.0 and a provided `negative_prompt`, enables true classifier-free guidance.
            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.
            image (`torch.FloatTensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.FloatTensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`, *optional*):
                Input image for img2img generation. If provided, the pipeline will perform image-to-image generation
                instead of text-to-image generation.
            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.
            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.
            sigmas (`List[float]`, *optional*):
                Custom sigmas to use for the denoising process with schedulers which support a `sigmas` argument in
                their `set_timesteps` method. If not defined, the default behavior when `num_inference_steps` is passed
                will be used.
            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.
            ip_adapter_image: (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
            ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*):
                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of
                IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not
                provided, embeddings are computed from the `ip_adapter_image` input argument.
            negative_ip_adapter_image:
                (`PipelineImageInput`, *optional*): Optional image input to work with IP Adapters.
            negative_ip_adapter_image_embeds (`List[torch.Tensor]`, *optional*):
                Pre-generated image embeddings for IP-Adapter. It should be a list of length same as number of
                IP-adapters. Each element should be a tensor of shape `(batch_size, num_images, emb_dim)`. If not
                provided, embeddings are computed from the `ip_adapter_image` input argument.
            negative_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated negative text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, negative_prompt_embeds will be generated from `negative_prompt` input
                argument.
            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated negative pooled text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt
                weighting. If not provided, pooled negative_prompt_embeds will be generated from `negative_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`.

        Examples:

        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.
        """
        # Verify the concept kwargs inputs
        if concept_attention_kwargs is not None:
            assert "concepts" in concept_attention_kwargs, "Concepts must be passed in the concept_attention_kwargs"
            assert isinstance(concept_attention_kwargs["concepts"], list), "Concepts must be a list of strings"
            assert len(concept_attention_kwargs["concepts"]) > 0, "Concepts must not be an empty list"
            assert "timesteps" in concept_attention_kwargs, "Timesteps must be passed in the concept_attention_kwargs"
            assert isinstance(concept_attention_kwargs["timesteps"], list), "Timesteps must be a list of integers"
            assert len(concept_attention_kwargs["timesteps"]) > 0, "Timesteps must not be an empty list"
            assert "layers" in concept_attention_kwargs, "Layers must be passed in the concept_attention_kwargs"
            assert isinstance(concept_attention_kwargs["layers"], list), "Layers must be a list of integers"
            assert len(concept_attention_kwargs["layers"]) > 0, "Layers must not be an empty list"

        height = height or self.default_sample_size * self.vae_scale_factor
        width = width or self.default_sample_size * self.vae_scale_factor

        init_image = None
        if image is not None:
            width, height = image.size
            init_image = self.image_processor.preprocess(image, height=height, width=width)
            init_image = init_image.to(dtype=torch.float32)

        # 1. Check inputs. Raise error if not correct
        self.check_inputs(
            prompt,
            prompt_2,
            height,
            width,
            negative_prompt=negative_prompt,
            negative_prompt_2=negative_prompt_2,
            prompt_embeds=prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
            callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs,
            max_sequence_length=max_sequence_length,
        )

        self._guidance_scale = guidance_scale
        self._joint_attention_kwargs = joint_attention_kwargs
        self._current_timestep = None
        self._interrupt = False

        # 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]

        device = self._execution_device

        lora_scale = (
            self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None
        )
        has_neg_prompt = negative_prompt is not None or (
            negative_prompt_embeds is not None and negative_pooled_prompt_embeds is not None
        )
        do_true_cfg = true_cfg_scale > 1 and has_neg_prompt
        (
            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,
        )
        if do_true_cfg:
            (
                negative_prompt_embeds,
                negative_pooled_prompt_embeds,
                _,
            ) = self.encode_prompt(
                prompt=negative_prompt,
                prompt_2=negative_prompt_2,
                prompt_embeds=negative_prompt_embeds,
                pooled_prompt_embeds=negative_pooled_prompt_embeds,
                device=device,
                num_images_per_prompt=num_images_per_prompt,
                max_sequence_length=max_sequence_length,
                lora_scale=lora_scale,
            )

        # Embed concepts
        concept_embeddings, pooled_concept_embeds, concept_ids = self.encode_concepts(
            concept_attention_kwargs["concepts"], 
            device=device
        )
        # Add the concept embeddings to the concept_attention_kwargs
        # if concept_attention_kwargs is not None:
        #     concept_attention_kwargs["concept_embeddings"] = concept_embeddings
        #     concept_attention_kwargs["concept_vec"] = concept_vec

        # 4. Prepare timesteps
        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if timesteps is None else None
        image_seq_len = (int(height) // self.vae_scale_factor // 2) * (int(width) // self.vae_scale_factor // 2)
        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)

        # 5. Prepare latent variables
        num_channels_latents = self.transformer.config.in_channels // 4
        latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
        latents, latent_image_ids = self.prepare_latents(
            init_image,
            latent_timestep,
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )

        # 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

        if (ip_adapter_image is not None or ip_adapter_image_embeds is not None) and (
            negative_ip_adapter_image is None and negative_ip_adapter_image_embeds is None
        ):
            negative_ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8)
        elif (ip_adapter_image is None and ip_adapter_image_embeds is None) and (
            negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None
        ):
            ip_adapter_image = np.zeros((width, height, 3), dtype=np.uint8)

        if self.joint_attention_kwargs is None:
            self._joint_attention_kwargs = {}

        image_embeds = None
        negative_image_embeds = None
        if ip_adapter_image is not None or ip_adapter_image_embeds is not None:
            image_embeds = self.prepare_ip_adapter_image_embeds(
                ip_adapter_image,
                ip_adapter_image_embeds,
                device,
                batch_size * num_images_per_prompt,
            )
        if negative_ip_adapter_image is not None or negative_ip_adapter_image_embeds is not None:
            negative_image_embeds = self.prepare_ip_adapter_image_embeds(
                negative_ip_adapter_image,
                negative_ip_adapter_image_embeds,
                device,
                batch_size * num_images_per_prompt,
            )

        # Make concept attention maps
        all_concept_attention_maps = []

        # 6. Denoising loop
        with self.progress_bar(total=num_inference_steps) as progress_bar:
            for i, t in enumerate(timesteps):
                if self.interrupt:
                    continue

                self._current_timestep = t
                if image_embeds is not None:
                    self._joint_attention_kwargs["ip_adapter_image_embeds"] = image_embeds
                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
                timestep = t.expand(latents.shape[0]).to(latents.dtype)

                # Don't do concept attention if the timestep is not in the concept_attention_kwargs
                if concept_attention_kwargs is not None and i not in concept_attention_kwargs["timesteps"]:
                    current_concept_embeddings = None
                elif concept_attention_kwargs is None:
                    # Use concept embeddings for all timesteps if no specific config
                    current_concept_embeddings = concept_embeddings
                else:
                    # Use concept embeddings for specified timesteps
                    current_concept_embeddings = concept_embeddings


                transformer_output = self.transformer(
                    hidden_states=latents,
                    timestep=timestep / 1000,
                    guidance=guidance,
                    pooled_projections=pooled_prompt_embeds,
                    pooled_concept_embeds=pooled_concept_embeds,
                    encoder_hidden_states=prompt_embeds,
                    concept_hidden_states=current_concept_embeddings,
                    txt_ids=text_ids,
                    img_ids=latent_image_ids,
                    concept_ids=concept_ids,
                    joint_attention_kwargs=self.joint_attention_kwargs,
                    concept_attention_kwargs=concept_attention_kwargs,
                    return_dict=False,
                )
                noise_pred, current_concept_attention_maps = transformer_output
                
                # Process attention maps immediately with softmax (critical!)
                if concept_attention_kwargs is not None and i in concept_attention_kwargs["timesteps"] and current_concept_attention_maps is not None:
                    if isinstance(current_concept_attention_maps, list):
                        # Transformer now returns list of dictionaries (one per layer)
                        for layer_dict in current_concept_attention_maps:
                            if isinstance(layer_dict, dict):
                                all_concept_attention_maps.append(layer_dict)
                    elif isinstance(current_concept_attention_maps, dict):
                        # Collect vector dictionaries for proper postprocessing (like reference)
                        all_concept_attention_maps.append(current_concept_attention_maps)

                if do_true_cfg:
                    if negative_image_embeds is not None:
                        self._joint_attention_kwargs["ip_adapter_image_embeds"] = negative_image_embeds
                    neg_noise_pred = self.transformer(
                        hidden_states=latents,
                        timestep=timestep / 1000,
                        guidance=guidance,
                        pooled_projections=negative_pooled_prompt_embeds,
                        encoder_hidden_states=negative_prompt_embeds,
                        txt_ids=text_ids,
                        img_ids=latent_image_ids,
                        joint_attention_kwargs=self.joint_attention_kwargs,
                        return_dict=False,
                    )[0]
                    noise_pred = neg_noise_pred + true_cfg_scale * (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()

        self._current_timestep = None

        if output_type == "latent":
            image = latents
        elif output_type == "visualization":
            latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
            latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
            image = self.vae.decode(latents, return_dict=False)[0]
            image = image.detach()
            # Force PIL for visualization
            image = self.image_processor.postprocess(image, output_type="pil")
        else:
            latents = self._unpack_latents(latents, height, width, self.vae_scale_factor)
            latents = (latents / self.vae.config.scaling_factor) + self.vae.config.shift_factor
            image = self.vae.decode(latents, return_dict=False)[0]
            image = image.detach()
            image = self.image_processor.postprocess(image, output_type=output_type)


        # Process final attention maps
        if all_concept_attention_maps:
            # Extract and stack vectors across timesteps (like reference)
            concept_vectors = []
            image_vectors = []
            
            for timestep_data in all_concept_attention_maps:
                concept_vectors.append(timestep_data['concept_vectors'].detach().cpu())
                image_vectors.append(timestep_data['image_vectors'].detach().cpu())
            
            # Stack across timesteps: (timesteps, batch, heads, tokens, dim)
            concept_vectors = torch.stack(concept_vectors, dim=0)
            image_vectors = torch.stack(image_vectors, dim=0)
            
            concept_vectors = concept_vectors / (concept_vectors.norm(dim=-1, keepdim=True) + 1e-8)
            
            concept_attention_maps = einops.einsum(
                image_vectors,
                concept_vectors,
                "time batch patches dim, time batch concepts dim -> time batch concepts patches"
            )
            
            concept_attention_maps = torch.softmax(concept_attention_maps, dim=-2)
            
            concept_attention_maps = concept_attention_maps.mean(dim=0)
            
            # Convert to numpy for further processing
            concept_attention_maps = concept_attention_maps.float().numpy()
            
            # Reshape to spatial format (64x64 for FLUX)
            concept_attention_maps = einops.rearrange(
                concept_attention_maps,
                "batch concepts (h w) -> batch concepts h w",
                h=height // 16,
                w=width // 16,
            )

            # Normalize concept maps - always return raw normalized maps (not colored PIL images)
            # Global min-max normalization per batch for consistency
            processed = []
            for b in range(concept_attention_maps.shape[0]):
                maps = concept_attention_maps[b]  # (concepts, H, W)
                vmin, vmax = maps.min(), maps.max()
                
                if vmax > vmin:
                    maps = (maps - vmin) / (vmax - vmin)
                else:
                    maps = np.zeros_like(maps)
                
                # If visualization PIL images are needed, create them separately
                if output_type == "visualization":  # New output type for colored visualizations
                    batch_imgs = []
                    for i, m in enumerate(maps):
                        colored = plt.get_cmap("plasma")(m)
                        rgb = (colored[:, :, :3] * 255).astype(np.uint8)
                        batch_imgs.append(PIL.Image.fromarray(rgb))
                    processed.append(batch_imgs)
                else:
                    # Return raw normalized arrays as list of numpy arrays per concept
                    processed.append([maps[i] for i in range(maps.shape[0])])
            
            concept_attention_maps = processed
        else:
            concept_attention_maps = []

        # Offload all models
        self.maybe_free_model_hooks()

        if not return_dict:
            return (image, concept_attention_maps)

        return FluxConceptAttentionOutput(
            images=image,
            concept_attention_maps=concept_attention_maps,
        )
        

    def encode_image(
        self,
        image: Union[PIL.Image.Image, torch.Tensor],
        concepts: List[str],
        prompt: str = "",  # Optional prompt context
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_samples: int = 1,
        num_inference_steps: int = 28,
        noise_timestep: int = 15,  # How much noise to add (higher = more noise)
        guidance_scale: float = 3.5,
        generator: Optional[torch.Generator] = None,
        layers: List[int] = [15, 16, 17, 18],
        timesteps_to_analyze: List[int] = [24, 25, 26, 27],
        device: Optional[torch.device] = None,
        return_dict: bool = True,
        max_sequence_length: int = 512,
    ) -> Union[torch.Tensor, Dict]:
        """
        Encode an existing image to analyze concept attention patterns.
        
        This method adds controlled noise to an input image and runs it through
        the denoising process to capture how the model attends to different concepts.
        
        Args:
            image: Input image to analyze (PIL Image or tensor)
            concepts: List of concept strings to analyze
            prompt: Optional text prompt for context
            height, width: Output dimensions (inferred from image if not provided)
            num_samples: Number of noise samples to average over
            num_inference_steps: Total denoising steps
            noise_timestep: Which timestep to add noise at (higher = more noise)
            guidance_scale: Guidance scale for generation
            generator: Random generator for reproducibility
            layers: Which transformer layers to analyze
            timesteps_to_analyze: Which denoising timesteps to capture attention from
            device: Compute device
            return_dict: Whether to return dict or tuple
            max_sequence_length: Max sequence length for text encoding
            
        Returns:
            Dictionary containing original image and concept attention maps
        """
        device = device or self._execution_device
        
        # Process input image
        if isinstance(image, PIL.Image.Image):
            if height is None or width is None:
                width, height = image.size
            # Ensure dimensions are compatible
            height = height - (height % (self.vae_scale_factor * 2))
            width = width - (width % (self.vae_scale_factor * 2))
            
            # Preprocess image
            processed_image = self.image_processor.preprocess(
                image, height=height, width=width
            ).to(device=device, dtype=self.vae.dtype)
        else:
            processed_image = image.to(device=device, dtype=self.vae.dtype)
            if height is None or width is None:
                _, _, height, width = processed_image.shape
        
        # Encode image to latent space
        with torch.no_grad():
            if processed_image.shape[1] != self.latent_channels:
                image_latents = self._encode_vae_image(processed_image, generator)
            else:
                image_latents = processed_image
        
        # Setup text embeddings
        prompt_embeds, pooled_prompt_embeds, text_ids = self.encode_prompt(
            prompt=prompt or "",
            prompt_2=None,
            device=device,
            num_images_per_prompt=1,
            max_sequence_length=max_sequence_length,
        )
        
        # Setup concept embeddings
        concept_embeddings, pooled_concept_embeds, concept_ids = self.encode_concepts(
            concepts, device=device
        )
        
        # Prepare timesteps
        image_seq_len = (height // self.vae_scale_factor // 2) * (width // self.vae_scale_factor // 2)
        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,
        )
        
        # Get full timestep schedule
        timesteps, _ = retrieve_timesteps(
            self.scheduler,
            num_inference_steps,
            device,
            timesteps=None,
            sigmas=None,
            mu=mu,
        )
        
        # Prepare latent image IDs
        latent_image_ids = self._prepare_latent_image_ids(
            1, 
            2 * (height // self.vae_scale_factor),
            2 * (width // self.vae_scale_factor), 
            device, 
            image_latents.dtype
        )
        
        # Handle guidance
        if self.transformer.config.guidance_embeds:
            guidance = torch.full([1], guidance_scale, device=device, dtype=torch.float32)
        else:
            guidance = None
        
        # Collect attention maps across samples
        all_concept_attention_maps = []

        for sample_idx in range(num_samples):
            # Add noise at specified timestep
            if generator is not None:
                # Use different seed for each sample
                sample_generator = torch.Generator(device=device).manual_seed(
                    generator.initial_seed() + sample_idx
                )
                noise = torch.randn(
                    image_latents.shape, 
                    generator=sample_generator, 
                    device=device, 
                    dtype=image_latents.dtype
                )
            else:
                noise = torch.randn_like(image_latents)
            
            # Get the timestep tensor for noise addition
            t_noise = timesteps[noise_timestep]
            noisy_latents = self.scheduler.scale_noise(
                image_latents, t_noise.unsqueeze(0), noise
            )
            
            # Pack latents for transformer
            packed_latents = self._pack_latents(
                noisy_latents, 
                1, 
                self.transformer.config.in_channels // 4,
                2 * (height // self.vae_scale_factor),
                2 * (width // self.vae_scale_factor)
            )
            
            # Run single denoising step with concept attention
            timestep_tensor = t_noise.expand(1).to(packed_latents.dtype) / 1000
            
            concept_attention_kwargs = {
                "concepts": concepts,
                "timesteps": [0],  # We're only doing one step
                "layers": layers
            }
            
            with torch.no_grad():
                transformer_output = self.transformer(
                    hidden_states=packed_latents,
                    timestep=timestep_tensor,
                    guidance=guidance,
                    pooled_projections=pooled_prompt_embeds,
                    pooled_concept_embeds=pooled_concept_embeds,
                    encoder_hidden_states=prompt_embeds,
                    concept_hidden_states=concept_embeddings,
                    txt_ids=text_ids,
                    img_ids=latent_image_ids,
                    concept_ids=concept_ids,
                    joint_attention_kwargs=None,
                    concept_attention_kwargs=concept_attention_kwargs,
                    return_dict=False,
                )
                
                _, sample_concept_attention_maps = transformer_output
                
                if sample_concept_attention_maps:
                    all_concept_attention_maps.extend(sample_concept_attention_maps)
        
        # Process collected attention maps
        concept_attention_maps = []
        if all_concept_attention_maps:
            # Extract and stack vectors across samples
            concept_vectors = []
            image_vectors = []
            
            for sample_data in all_concept_attention_maps:
                concept_vectors.append(sample_data['concept_vectors'].detach().cpu())
                image_vectors.append(sample_data['image_vectors'].detach().cpu())
            
            if concept_vectors and image_vectors:
                # Stack and average across samples
                concept_vectors = torch.stack(concept_vectors, dim=0).mean(dim=0)
                image_vectors = torch.stack(image_vectors, dim=0).mean(dim=0)
                
                # Normalize concept vectors
                concept_vectors = concept_vectors / (concept_vectors.norm(dim=-1, keepdim=True) + 1e-8)
                
                # Compute attention maps
                attention_maps = einops.einsum(
                    image_vectors,
                    concept_vectors,
                    "batch patches dim, batch concepts dim -> batch concepts patches"
                )
                
                # Apply softmax normalization
                attention_maps = torch.softmax(attention_maps, dim=-2)
                
                # Convert to numpy and reshape to spatial format
                attention_maps = attention_maps.float().numpy()
                spatial_maps = einops.rearrange(
                    attention_maps,
                    "batch concepts (h w) -> batch concepts h w",
                    h=height // 16,
                    w=width // 16,
                )
                
                # Process maps per batch (usually just 1 batch for encode_image)
                for b in range(spatial_maps.shape[0]):
                    maps = spatial_maps[b]  # (concepts, H, W)
                    
                    # Normalize maps
                    vmin, vmax = maps.min(), maps.max()
                    if vmax > vmin:
                        maps = (maps - vmin) / (vmax - vmin)
                    else:
                        maps = np.zeros_like(maps)
                    
                    # Convert to list of arrays per concept
                    concept_maps = [maps[i] for i in range(maps.shape[0])]
                    concept_attention_maps.append(concept_maps)
        
        # Prepare output
        result = {
            'image': image,  # Original input image
            'concept_attention_maps': concept_attention_maps,
            'concepts': concepts,
            'height': height,
            'width': width
        }
        
        if not return_dict:
            return (image, concept_attention_maps)
        
        return result