dreamo/dreamo_pipeline.py

# Copyright (c) 2025 Bytedance Ltd. and/or its affiliates
# Copyright 2024 Black Forest Labs and The HuggingFace Team. All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from typing import Any, Callable, Dict, List, Optional, Union

import diffusers
import numpy as np
import torch
import torch.nn as nn
from diffusers import FluxPipeline
from diffusers.pipelines.flux.pipeline_flux import calculate_shift, retrieve_timesteps
from diffusers.pipelines.flux.pipeline_output import FluxPipelineOutput
from einops import repeat
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

from dreamo.transformer import flux_transformer_forward
from dreamo.utils import convert_flux_lora_to_diffusers

diffusers.models.transformers.transformer_flux.FluxTransformer2DModel.forward = flux_transformer_forward


def get_task_embedding_idx(task):
    return 0


class DreamOPipeline(FluxPipeline):
    def __init__(self, scheduler, vae, text_encoder, tokenizer, text_encoder_2, tokenizer_2, transformer):
        super().__init__(scheduler, vae, text_encoder, tokenizer, text_encoder_2, tokenizer_2, transformer)
        self.t5_embedding = nn.Embedding(10, 4096)
        self.task_embedding = nn.Embedding(2, 3072)
        self.idx_embedding = nn.Embedding(10, 3072)

    def load_dreamo_model(self, device, use_turbo=True, version='v1.1'):
        # download models and load file
        hf_hub_download(repo_id='ByteDance/DreamO', filename='dreamo.safetensors', local_dir='models')
        hf_hub_download(repo_id='ByteDance/DreamO', filename='dreamo_cfg_distill.safetensors', local_dir='models')
        if version == 'v1':
            hf_hub_download(repo_id='ByteDance/DreamO', filename='dreamo_quality_lora_pos.safetensors',
                            local_dir='models')
            hf_hub_download(repo_id='ByteDance/DreamO', filename='dreamo_quality_lora_neg.safetensors',
                            local_dir='models')
            quality_lora_pos = load_file('models/dreamo_quality_lora_pos.safetensors')
            quality_lora_neg = load_file('models/dreamo_quality_lora_neg.safetensors')
        elif version == 'v1.1':
            hf_hub_download(repo_id='ByteDance/DreamO', filename='v1.1/dreamo_sft_lora.safetensors', local_dir='models')
            hf_hub_download(repo_id='ByteDance/DreamO', filename='v1.1/dreamo_dpo_lora.safetensors', local_dir='models')
            sft_lora = load_file('models/v1.1/dreamo_sft_lora.safetensors')
            dpo_lora = load_file('models/v1.1/dreamo_dpo_lora.safetensors')
        else:
            raise ValueError(f'there is no {version}')
        dreamo_lora = load_file('models/dreamo.safetensors')
        cfg_distill_lora = load_file('models/dreamo_cfg_distill.safetensors')

        # load embedding
        self.t5_embedding.weight.data = dreamo_lora.pop('dreamo_t5_embedding.weight')[-10:]
        self.task_embedding.weight.data = dreamo_lora.pop('dreamo_task_embedding.weight')
        self.idx_embedding.weight.data = dreamo_lora.pop('dreamo_idx_embedding.weight')
        self._prepare_t5()

        # main lora
        dreamo_diffuser_lora = convert_flux_lora_to_diffusers(dreamo_lora)
        adapter_names = ['dreamo']
        adapter_weights = [1]
        self.load_lora_weights(dreamo_diffuser_lora, adapter_name='dreamo')

        # cfg lora to avoid true image cfg
        cfg_diffuser_lora = convert_flux_lora_to_diffusers(cfg_distill_lora)
        self.load_lora_weights(cfg_diffuser_lora, adapter_name='cfg')
        adapter_names.append('cfg')
        adapter_weights.append(1)

        # turbo lora to speed up (from 25+ step to 12 step)
        if use_turbo:
            self.load_lora_weights(
                hf_hub_download(
                    "alimama-creative/FLUX.1-Turbo-Alpha", "diffusion_pytorch_model.safetensors", local_dir='models'
                ),
                adapter_name='turbo',
            )
            adapter_names.append('turbo')
            adapter_weights.append(1)

        if version == 'v1':
            # quality loras, one pos, one neg
            quality_lora_pos = convert_flux_lora_to_diffusers(quality_lora_pos)
            self.load_lora_weights(quality_lora_pos, adapter_name='quality_pos')
            adapter_names.append('quality_pos')
            adapter_weights.append(0.15)
            quality_lora_neg = convert_flux_lora_to_diffusers(quality_lora_neg)
            self.load_lora_weights(quality_lora_neg, adapter_name='quality_neg')
            adapter_names.append('quality_neg')
            adapter_weights.append(-0.8)
        elif version == 'v1.1':
            self.load_lora_weights(sft_lora, adapter_name='sft_lora')
            adapter_names.append('sft_lora')
            adapter_weights.append(1)
            self.load_lora_weights(dpo_lora, adapter_name='dpo_lora')
            adapter_names.append('dpo_lora')
            adapter_weights.append(1.25)

        self.set_adapters(adapter_names, adapter_weights)
        self.fuse_lora(adapter_names=adapter_names, lora_scale=1)
        self.unload_lora_weights()

        self.t5_embedding = self.t5_embedding.to(device)
        self.task_embedding = self.task_embedding.to(device)
        self.idx_embedding = self.idx_embedding.to(device)

    def _prepare_t5(self):
        self.text_encoder_2.resize_token_embeddings(len(self.tokenizer_2))
        num_new_token = 10
        new_token_list = [f"[ref#{i}]" for i in range(1, 10)] + ["[res]"]
        self.tokenizer_2.add_tokens(new_token_list, special_tokens=False)
        self.text_encoder_2.resize_token_embeddings(len(self.tokenizer_2))
        input_embedding = self.text_encoder_2.get_input_embeddings().weight.data
        input_embedding[-num_new_token:] = self.t5_embedding.weight.data

    @staticmethod
    def _prepare_latent_image_ids(batch_size, height, width, device, dtype, start_height=0, start_width=0):
        latent_image_ids = torch.zeros(height // 2, width // 2, 3)
        latent_image_ids[..., 1] = latent_image_ids[..., 1] + torch.arange(height // 2)[:, None] + start_height
        latent_image_ids[..., 2] = latent_image_ids[..., 2] + torch.arange(width // 2)[None, :] + start_width

        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
        )

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

    @staticmethod
    def _prepare_style_latent_image_ids(batch_size, height, width, device, dtype, start_height=0, start_width=0):
        latent_image_ids = torch.zeros(height // 2, width // 2, 3)
        latent_image_ids[..., 1] = latent_image_ids[..., 1] + start_height
        latent_image_ids[..., 2] = latent_image_ids[..., 2] + start_width

        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
        )

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

    @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,
        true_cfg_start_step: int = 1,
        true_cfg_end_step: int = 1,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 28,
        sigmas: Optional[List[float]] = None,
        guidance_scale: float = 3.5,
        neg_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,
        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,
        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,
        ref_conds=None,
        first_step_guidance_scale=3.5,
    ):
        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.
            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 3.5):
                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.
            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.
        """

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

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

        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,
            )

        # 4. Prepare latent variables
        num_channels_latents = self.transformer.config.in_channels // 4
        latents, latent_image_ids = self.prepare_latents(
            batch_size * num_images_per_prompt,
            num_channels_latents,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )

        # 4.1 concat ref tokens to latent
        origin_img_len = latents.shape[1]
        embeddings = repeat(self.task_embedding.weight[1], "c -> n l c", n=batch_size, l=origin_img_len)
        ref_latents = []
        ref_latent_image_idss = []
        start_height = height // 16
        start_width = width // 16
        for ref_cond in ref_conds:
            img = ref_cond['img']  # [b, 3, h, w], range [-1, 1]
            task = ref_cond['task']
            idx = ref_cond['idx']

            # encode ref with VAE
            img = img.to(latents)
            ref_latent = self.vae.encode(img).latent_dist.sample()
            ref_latent = (ref_latent - self.vae.config.shift_factor) * self.vae.config.scaling_factor
            cur_height = ref_latent.shape[2]
            cur_width = ref_latent.shape[3]
            ref_latent = self._pack_latents(ref_latent, batch_size, num_channels_latents, cur_height, cur_width)
            ref_latent_image_ids = self._prepare_latent_image_ids(
                batch_size, cur_height, cur_width, device, prompt_embeds.dtype, start_height, start_width
            )
            start_height += cur_height // 2
            start_width += cur_width // 2

            # prepare task_idx_embedding
            task_idx = get_task_embedding_idx(task)
            cur_task_embedding = repeat(
                self.task_embedding.weight[task_idx], "c -> n l c", n=batch_size, l=ref_latent.shape[1]
            )
            cur_idx_embedding = repeat(
                self.idx_embedding.weight[idx], "c -> n l c", n=batch_size, l=ref_latent.shape[1]
            )
            cur_embedding = cur_task_embedding + cur_idx_embedding

            # concat ref to latent
            embeddings = torch.cat([embeddings, cur_embedding], dim=1)
            ref_latents.append(ref_latent)
            ref_latent_image_idss.append(ref_latent_image_ids)

        # 5. Prepare timesteps
        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,
            self.scheduler.config.get("base_image_seq_len", 256),
            self.scheduler.config.get("max_image_seq_len", 4096),
            self.scheduler.config.get("base_shift", 0.5),
            self.scheduler.config.get("max_shift", 1.15),
        )
        timesteps, num_inference_steps = retrieve_timesteps(
            self.scheduler,
            num_inference_steps,
            device,
            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
        neg_guidance = torch.full([1], neg_guidance_scale, device=device, dtype=torch.float32)
        neg_guidance = neg_guidance.expand(latents.shape[0])
        first_step_guidance = torch.full([1], first_step_guidance_scale, device=device, dtype=torch.float32)

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

        # 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
                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
                timestep = t.expand(latents.shape[0]).to(latents.dtype)

                noise_pred = self.transformer(
                    hidden_states=torch.cat((latents, *ref_latents), dim=1),
                    timestep=timestep / 1000,
                    guidance=guidance if i > 0 else first_step_guidance,
                    pooled_projections=pooled_prompt_embeds,
                    encoder_hidden_states=prompt_embeds,
                    txt_ids=text_ids,
                    img_ids=torch.cat((latent_image_ids, *ref_latent_image_idss), dim=1),
                    joint_attention_kwargs=self.joint_attention_kwargs,
                    return_dict=False,
                    embeddings=embeddings,
                )[0][:, :origin_img_len]

                if do_true_cfg and i >= true_cfg_start_step and i < true_cfg_end_step:
                    neg_noise_pred = self.transformer(
                        hidden_states=latents,
                        timestep=timestep / 1000,
                        guidance=neg_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 and 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()

        self._current_timestep = None

        if output_type == "latent":
            image = latents
        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 = self.image_processor.postprocess(image, output_type=output_type)

        # Offload all models
        self.maybe_free_model_hooks()

        if not return_dict:
            return (image,)

        return FluxPipelineOutput(images=image)