diffusers_local/pipeline_z_image_control_unified.py

# Copyright 2025 Alibaba Z-Image Team 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.

import inspect
from typing import List, Optional, Union
import torch
from PIL import Image

from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler, DiffusionPipeline
from diffusers.loaders import FromSingleFileMixin, ZImageLoraLoaderMixin
from diffusers.image_processor import VaeImageProcessor 
from diffusers.utils import logging
from diffusers.pipelines.z_image.pipeline_z_image import calculate_shift
from diffusers.pipelines.z_image.pipeline_output import ZImagePipelineOutput
from diffusers_local.z_image_control_transformer_2d import ZImageControlTransformer2DModel
from transformers import AutoTokenizer, PreTrainedModel
from diffusers.utils.torch_utils import randn_tensor

logger = logging.get_logger(__name__)

# Copied from diffusers.pipelines.flux.pipeline_flux.calculate_shift
def calculate_shift(
    image_seq_len,
    base_seq_len: int = 256,
    max_seq_len: int = 4096,
    base_shift: float = 0.5,
    max_shift: float = 1.15,
):
    m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
    b = base_shift - m * base_seq_len
    mu = image_seq_len * m + b
    return mu


# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.retrieve_timesteps
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,
):
    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`.

    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


class ZImageControlUnifiedPipeline(DiffusionPipeline, ZImageLoraLoaderMixin, FromSingleFileMixin):
    _model_cpu_offload_seq = "text_encoder->transformer->vae"
    _optional_components = []
    _callback_tensor_inputs = ["latents", "prompt_embeds"]
    
    def __init__(
        self,
        scheduler: FlowMatchEulerDiscreteScheduler,
        vae: AutoencoderKL,
        text_encoder: PreTrainedModel,
        tokenizer: AutoTokenizer,
        transformer: ZImageControlTransformer2DModel,
    ):      
        self.register_modules(
            vae=vae, text_encoder=text_encoder, tokenizer=tokenizer,
            transformer=transformer, scheduler=scheduler
        )
        self.vae_scale_factor = 2 ** (len(self.vae.config.block_out_channels) - 1)
        self.image_processor = VaeImageProcessor(vae_scale_factor=self.vae_scale_factor * 2)

    def _encode_prompt(self, prompt: str, device: torch.device, max_sequence_length: int) -> torch.Tensor:
        messages = [{"role": "user", "content": prompt}]
        if hasattr(self.tokenizer, "apply_chat_template"):
            prompt_formatted = self.tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=True)
        else:
            prompt_formatted = prompt
            
        text_inputs = self.tokenizer(prompt_formatted, padding="max_length", max_length=max_sequence_length, truncation=True, return_tensors="pt").to(device)
        prompt_masks = text_inputs.attention_mask.bool()
        with torch.no_grad():
            prompt_embeds = self.text_encoder(input_ids=text_inputs.input_ids, attention_mask=prompt_masks, output_hidden_states=True).hidden_states[-2]
        return prompt_embeds[0][prompt_masks[0]]

    def prepare_latents(self, batch_size, num_channels, height, width, dtype, device, generator, latents=None):
        shape = (batch_size, num_channels, height // self.vae_scale_factor, width // self.vae_scale_factor)
        if latents is None:
            latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
        else:
            latents = latents.to(device)
        return latents * self.scheduler.init_noise_sigma if hasattr(self.scheduler, "init_noise_sigma") else latents
    
    def prepare_control_image(self, image, width, height, batch_size, num_images_per_prompt, device, dtype):
        image = self.image_processor.preprocess(image, height=height, width=width).to(device=device, dtype=dtype)
                
        image_batch_size = image.shape[0]
        if image_batch_size == 1:
            repeat_by = batch_size
        else:
            repeat_by = num_images_per_prompt
        image = image.repeat_interleave(repeat_by, dim=0)
        return image

    @torch.no_grad()
    def __call__(
        self,
        prompt: Union[str, List[str]],
        image: Union[torch.Tensor, Image.Image],
        negative_prompt: Optional[Union[str, List[str]]] = None,
        height: Optional[int] = None,
        width: Optional[int] = None,
        num_inference_steps: int = 50,
        guidance_scale: float = 0.0,
        controlnet_conditioning_scale: float = 1.0,
        num_images_per_prompt: int = 1,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        output_type: str = "pil",
        return_dict: bool = True,
        **kwargs,
    ):
        device = self._execution_device
        height = height or image.height
        width = width or image.width
        
        # 1. Prompt adjustment and batch size
        if isinstance(prompt, str): prompt = [prompt]
        if isinstance(negative_prompt, str): negative_prompt = [negative_prompt]
        
        batch_size = len(prompt) * num_images_per_prompt
        do_cfg = guidance_scale > 0.0

        # 2. Encode text
        # Repeat embeddings if num_images_per_prompt > 1
        prompt_embeds_list = []
        for p in prompt:
            embed = self._encode_prompt(p, device, 512)
            for _ in range(num_images_per_prompt):
                prompt_embeds_list.append(embed)
        
        if do_cfg:
            if negative_prompt is None: negative_prompt = [""] * len(prompt)
            neg_embeds_list = []
            for np in negative_prompt:
                embed = self._encode_prompt(np, device, 512)
                for _ in range(num_images_per_prompt):
                    neg_embeds_list.append(embed)
            
            prompt_input = neg_embeds_list + prompt_embeds_list
        else:
            prompt_input = prompt_embeds_list

       
        # 3. Control image preparation
        control_tensor = self.prepare_control_image(
            image, width, height, batch_size, num_images_per_prompt, device, self.vae.dtype
        )
        
        if len(control_tensor.shape) == 3:
            control_tensor = control_tensor.unsqueeze(0)
            
        with torch.no_grad():
            # Encode to latents
            control_latents = self.vae.encode(control_tensor).latent_dist.mode()
            control_latents = control_latents * self.vae.config.scaling_factor

        # Channel fix: 4 channels -> 16 channels
        if control_latents.shape[1] == 4 and self.transformer.in_channels == 16:
            control_latents = control_latents.repeat(1, 4, 1, 1)  # [B, 16, H, W]
        
        control_latents = control_latents.to(dtype=self.transformer.dtype)
        
        # Fix dimension: frame dimension [B, 16, 1, H, W]
        control_latents = control_latents.unsqueeze(2)                
        control_context = list(control_latents.unbind(0))
        
        # Expansion for CFG
        if do_cfg:
            control_context_input = control_context * 2
        else:
            control_context_input = control_context
       
        # 4. Initial latents
        latents = self.prepare_latents(
            batch_size, self.transformer.in_channels, height, width, 
            prompt_embeds_list[0].dtype, device, generator
        )
        latents = latents.to(self.transformer.dtype)

        # 5. Denoising loop
        image_seq_len = (height // (self.vae_scale_factor)) * (width // (self.vae_scale_factor))
        mu = calculate_shift(image_seq_len)
        self.scheduler.set_timesteps(num_inference_steps, device=device, mu=mu)
        
        for t in self.progress_bar(self.scheduler.timesteps):
            t_input = t.expand(len(prompt_input))
            timestep_norm = (1000.0 - t_input) / 1000.0
            
            latents_input = torch.cat([latents] * 2) if do_cfg else latents
            
            # List of [16, 1, H, W]
            latent_list = list(latents_input.unsqueeze(2).unbind(dim=0))
            
            model_out_list = self.transformer(
                x=latent_list,
                t=timestep_norm,
                cap_feats=prompt_input,
                control_context=control_context_input,
                conditioning_scale=controlnet_conditioning_scale,
            )[0]
            
            model_out = torch.stack(model_out_list, dim=0).squeeze(2)

            if do_cfg:
                neg_out, pos_out = model_out.chunk(2)
                noise_pred = neg_out + guidance_scale * (pos_out - neg_out)
            else:
                noise_pred = model_out
            
            noise_pred = -noise_pred
            latents = self.scheduler.step(noise_pred, t, latents).prev_sample
            
        # 6. Decode
        if not output_type == "latent":
            # Pass 16 channels to VAE
            latents_for_vae = latents.to(self.vae.dtype)
            latents_for_vae = (latents_for_vae / self.vae.config.scaling_factor) + self.vae.config.shift_factor
            
            image = self.vae.decode(latents_for_vae, return_dict=False)[0]
            image = self.image_processor.postprocess(image, output_type=output_type)
        else:
            image = latents
        
        self.maybe_free_model_hooks()
        return ZImagePipelineOutput(images=image)