diffusers 0.30.3__py3-none-any.whl → 0.32.0__py3-none-any.whl

diffusers/pipelines/flux/pipeline_flux.py

@@ -17,10 +17,17 @@ from typing import Any, Callable, Dict, List, Optional, Union
 
 import numpy as np
 import torch
-from transformers import CLIPTextModel, CLIPTokenizer, T5EncoderModel, T5TokenizerFast
+from transformers import (
+    CLIPImageProcessor,
+    CLIPTextModel,
+    CLIPTokenizer,
+    CLIPVisionModelWithProjection,
+    T5EncoderModel,
+    T5TokenizerFast,
+)
 
-from ...image_processor import VaeImageProcessor
-from ...loaders import FluxLoraLoaderMixin
+from ...image_processor import PipelineImageInput, VaeImageProcessor
+from ...loaders import FluxIPAdapterMixin, FluxLoraLoaderMixin, FromSingleFileMixin, TextualInversionLoaderMixin
 from ...models.autoencoders import AutoencoderKL
 from ...models.transformers import FluxTransformer2DModel
 from ...schedulers import FlowMatchEulerDiscreteScheduler
@@ -86,7 +93,7 @@ def retrieve_timesteps(
     sigmas: Optional[List[float]] = None,
     **kwargs,
 ):
-    """
+    r"""
     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`.
 
@@ -137,7 +144,13 @@ def retrieve_timesteps(
     return timesteps, num_inference_steps
 
 
-class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
+class FluxPipeline(
+    DiffusionPipeline,
+    FluxLoraLoaderMixin,
+    FromSingleFileMixin,
+    TextualInversionLoaderMixin,
+    FluxIPAdapterMixin,
+):
     r"""
     The Flux pipeline for text-to-image generation.
 
@@ -164,8 +177,8 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
             [T5TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/t5#transformers.T5TokenizerFast).
     """
 
-    model_cpu_offload_seq = "text_encoder->text_encoder_2->transformer->vae"
-    _optional_components = []
+    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__(
@@ -177,6 +190,8 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         text_encoder_2: T5EncoderModel,
         tokenizer_2: T5TokenizerFast,
         transformer: FluxTransformer2DModel,
+        image_encoder: CLIPVisionModelWithProjection = None,
+        feature_extractor: CLIPImageProcessor = None,
     ):
         super().__init__()
 
@@ -188,15 +203,19 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
             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
+            2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8
         )
-        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor)
+        # Flux latents are turned into 2x2 patches and packed. This means the latent width and height has to be divisible
+        # by the patch size. So the vae scale factor is multiplied by the patch size to account for this
+        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)
         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.default_sample_size = 128
 
     def _get_t5_prompt_embeds(
         self,
@@ -212,6 +231,9 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         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",
@@ -255,6 +277,9 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         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",
@@ -331,10 +356,6 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
                 scale_lora_layers(self.text_encoder_2, lora_scale)
 
         prompt = [prompt] if isinstance(prompt, str) else prompt
-        if prompt is not None:
-            batch_size = len(prompt)
-        else:
-            batch_size = prompt_embeds.shape[0]
 
         if prompt_embeds is None:
             prompt_2 = prompt_2 or prompt
@@ -364,24 +385,71 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
                 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(batch_size, prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)
-        text_ids = text_ids.repeat(num_images_per_prompt, 1, 1)
+        text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)
 
         return prompt_embeds, pooled_prompt_embeds, text_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 % 8 != 0 or width % 8 != 0:
-            raise ValueError(f"`height` and `width` have to be divisible by 8 but are {height} and {width}.")
+        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
@@ -409,25 +477,47 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         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_ids = torch.zeros(height, width, 3)
+        latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height)[:, None]
+        latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width)[None, :]
 
         latent_image_id_height, latent_image_id_width, latent_image_id_channels = latent_image_ids.shape
 
-        latent_image_ids = latent_image_ids[None, :].repeat(batch_size, 1, 1, 1)
         latent_image_ids = latent_image_ids.reshape(
-            batch_size, latent_image_id_height * latent_image_id_width, latent_image_id_channels
+            latent_image_id_height * latent_image_id_width, latent_image_id_channels
         )
 
         return latent_image_ids.to(device=device, dtype=dtype)
@@ -444,16 +534,47 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
     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
+        # VAE applies 8x compression on images but we must also account for packing which requires
+        # latent height and width to be divisible by 2.
+        height = 2 * (int(height) // (vae_scale_factor * 2))
+        width = 2 * (int(width) // (vae_scale_factor * 2))
 
-        latents = latents.view(batch_size, height, width, channels // 4, 2, 2)
+        latents = latents.view(batch_size, height // 2, width // 2, 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)
+        latents = latents.reshape(batch_size, channels // (2 * 2), height, width)
 
         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 prepare_latents(
         self,
         batch_size,
@@ -465,13 +586,15 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         generator,
         latents=None,
     ):
-        height = 2 * (int(height) // self.vae_scale_factor)
-        width = 2 * (int(width) // self.vae_scale_factor)
+        # VAE applies 8x compression on images but we must also account for packing which requires
+        # latent height and width to be divisible by 2.
+        height = 2 * (int(height) // (self.vae_scale_factor * 2))
+        width = 2 * (int(width) // (self.vae_scale_factor * 2))
 
         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)
+            latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype)
             return latents.to(device=device, dtype=dtype), latent_image_ids
 
         if isinstance(generator, list) and len(generator) != batch_size:
@@ -483,7 +606,7 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         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)
+        latent_image_ids = self._prepare_latent_image_ids(batch_size, height // 2, width // 2, device, dtype)
 
         return latents, latent_image_ids
 
@@ -509,16 +632,25 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         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,
         num_inference_steps: int = 28,
-        timesteps: List[int] = None,
-        guidance_scale: float = 7.0,
+        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,
@@ -543,10 +675,10 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
             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.
+            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
@@ -568,6 +700,17 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
             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.
             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`.
@@ -605,8 +748,12 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
             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,
         )
@@ -628,6 +775,7 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         lora_scale = (
             self.joint_attention_kwargs.get("scale", None) if self.joint_attention_kwargs is not None else None
         )
+        do_true_cfg = true_cfg_scale > 1 and negative_prompt is not None
         (
             prompt_embeds,
             pooled_prompt_embeds,
@@ -642,6 +790,21 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
             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,
+            )
 
         # 4. Prepare latent variables
         num_channels_latents = self.transformer.config.in_channels // 4
@@ -657,7 +820,7 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
         )
 
         # 5. Prepare timesteps
-        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps)
+        sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas
         image_seq_len = latents.shape[1]
         mu = calculate_shift(
             image_seq_len,
@@ -670,32 +833,61 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
             self.scheduler,
             num_inference_steps,
             device,
-            timesteps,
-            sigmas,
+            sigmas=sigmas,
             mu=mu,
         )
         num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
         self._num_timesteps = len(timesteps)
 
+        # 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,
+            )
+
         # 6. Denoising loop
         with self.progress_bar(total=num_inference_steps) as progress_bar:
             for i, t in enumerate(timesteps):
                 if self.interrupt:
                     continue
 
+                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)
 
-                # handle guidance
-                if self.transformer.config.guidance_embeds:
-                    guidance = torch.tensor([guidance_scale], device=device)
-                    guidance = guidance.expand(latents.shape[0])
-                else:
-                    guidance = None
-
                 noise_pred = self.transformer(
                     hidden_states=latents,
-                    # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
                     timestep=timestep / 1000,
                     guidance=guidance,
                     pooled_projections=pooled_prompt_embeds,
@@ -706,6 +898,22 @@ class FluxPipeline(DiffusionPipeline, FluxLoraLoaderMixin):
                     return_dict=False,
                 )[0]
 
+                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]