pipeline_newbie_img2img.py

# Copyright 2025 Alpha-VLLM 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 Any, Callable, Dict, List, Optional, Tuple, Union

import numpy as np
import torch

from transformers import (
    PreTrainedModel,
    Gemma3PreTrainedModel,
    GemmaTokenizer,
    GemmaTokenizerFast,
    XLMRobertaTokenizer,
    XLMRobertaTokenizerFast
)

from diffusers.pipelines.pipeline_utils import ImagePipelineOutput
from diffusers.image_processor import PipelineImageInput
from diffusers.pipelines.newbie.pipeline_newbie import NewbiePipeline

from diffusers.models import AutoencoderKL
from diffusers.models.transformers.transformer_lumina2 import Lumina2Transformer2DModel
from diffusers.schedulers import FlowMatchEulerDiscreteScheduler

from diffusers.utils import (
    is_torch_xla_available,
    logging,
    replace_example_docstring,
)
from diffusers.utils.torch_utils import randn_tensor


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


EXAMPLE_DOC_STRING = """
    Examples:
        ```py
        >>> import torch
        >>> from diffusers import NewbieImg2ImgPipeline
        >>> from diffusers.utils import load_image
        >>> from transformers import AutoModel

        >>> device = "cuda"
        >>> model_path = "Disty0/NewBie-image-Exp0.1-Diffusers"
        >>> text_encoder_2 = AutoModel.from_pretrained(model_path, subfolder="text_encoder_2", trust_remote_code=True, torch_dtype=torch.bfloat16)
        
        >>> pipe = NewbieImg2ImgPipeline.from_pretrained(model_path, text_encoder_2=text_encoder_2, torch_dtype=torch.bfloat16)
        >>> pipe.enable_model_cpu_offload(device=device)

        >>> url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
        >>> init_image = load_image(url).resize((1024, 1024))

        >>> prompt = "A fantasy landscape with mountains and a river, detailed, vibrant colors, anime style"
        >>> negative_prompt = "low quality, worst quality, blurry"

        >>> image = pipe(
        >>>     prompt,
        >>>     image=init_image,
        >>>     strength=0.6,
        >>>     negative_prompt=negative_prompt,
        >>>     guidance_scale=2.5,
        >>>     num_inference_steps=30,
        >>>     generator=torch.manual_seed(42),
        >>> ).images[0]
        ```
"""


# 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


# Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion_img2img.retrieve_latents
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 NewbieImg2ImgPipeline(NewbiePipeline):
    r"""
    Pipeline for image-to-image generation using Lumina-T2I / Newbie model.

    This model inherits from [`NewbiePipeline`]. Check the superclass documentation for the generic methods the
    library implements for all the pipelines (such as downloading or saving, running on a particular device, etc.)

    Args:
        vae ([`AutoencoderKL`]):
            Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations.
        text_encoder ([`Gemma3PreTrainedModel`]):
            Frozen Gemma3 text-encoder.
        text_encoder_2 ([`PreTrainedModel`]):
            Frozen JinaCLIPTextModel text-encoder. Requires `trust_remote_code=True`.
        tokenizer (`GemmaTokenizer` or `GemmaTokenizerFast`):
            Gemma tokenizer.
        tokenizer_2 (`XLMRobertaTokenizer` or `XLMRobertaTokenizerFast`):
            XLMRoberta tokenizer.
        transformer ([`Transformer2DModel`]):
            A text conditioned `Transformer2DModel` to denoise the encoded image latents.
        scheduler ([`SchedulerMixin`]):
            A scheduler to be used in combination with `transformer` to denoise the encoded image latents.
    """
    
    # Copied from diffusers.pipelines.stable_diffusion_3.pipeline_stable_diffusion_3_img2img.StableDiffusion3Img2ImgPipeline.get_timesteps
    def get_timesteps(self, num_inference_steps, strength, device):
        # get the original timestep using init_timestep
        init_timestep = min(num_inference_steps * strength, num_inference_steps)

        t_start = int(max(num_inference_steps - init_timestep, 0))
        timesteps = self.scheduler.timesteps[t_start * self.scheduler.order :]
        if hasattr(self.scheduler, "set_begin_index"):
            self.scheduler.set_begin_index(t_start * self.scheduler.order)

        return timesteps, num_inference_steps - t_start

    def prepare_latents(
        self,
        image,
        timestep,
        batch_size,
        num_channels_latents,
        height,
        width,
        dtype,
        device,
        generator,
        latents=None,
    ):
        if latents is not None:
            return latents.to(device=device, dtype=dtype)
        
        # 1. Encode the input image
        image = image.to(device=device, dtype=dtype)
        
        if image.shape[1] == num_channels_latents:
            image_latents = image
        else:
            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)

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

        # 2. Handle batch size expansion for num_images_per_prompt
        if batch_size > image_latents.shape[0] and batch_size % image_latents.shape[0] == 0:
            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."
            )
        
        # 3. Add noise to latents
        shape = image_latents.shape
        noise = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
        latents = self.scheduler.scale_noise(image_latents, timestep, noise)

        return latents

    @torch.no_grad()
    @replace_example_docstring(EXAMPLE_DOC_STRING)
    def __call__(
        self,
        prompt: Union[str, List[str]] = None,
        image: PipelineImageInput = None,
        strength: float = 0.6,
        width: Optional[int] = None,
        height: Optional[int] = None,
        num_inference_steps: int = 30,
        guidance_scale: float = 4.0,
        negative_prompt: Union[str, List[str]] = None,
        sigmas: List[float] = None,
        num_images_per_prompt: Optional[int] = 1,
        generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
        latents: Optional[torch.Tensor] = None,
        prompt_embeds: Optional[torch.Tensor] = None,
        pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        negative_prompt_embeds: Optional[torch.Tensor] = None,
        negative_pooled_prompt_embeds: Optional[torch.FloatTensor] = None,
        prompt_attention_mask: Optional[torch.Tensor] = None,
        negative_prompt_attention_mask: Optional[torch.Tensor] = None,
        output_type: Optional[str] = "pil",
        return_dict: bool = True,
        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"],
        system_prompt: Optional[str] = None,
        cfg_trunc_ratio: float = 1.0,
        cfg_normalization: bool = True,
        max_sequence_length: int = 512,
    ) -> Union[ImagePipelineOutput, Tuple]:
        """
        Function invoked when calling the pipeline for image-to-image 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.
            image (`torch.Tensor`, `PIL.Image.Image`, `np.ndarray`, `List[torch.Tensor]`, `List[PIL.Image.Image]`, or `List[np.ndarray]`):
                `Image`, numpy array or tensor representing an image batch to be used as the starting point.
            strength (`float`, *optional*, defaults to 0.6):
                Indicates extent to transform the reference `image`. Must be between 0 and 1. `image` is used as a
                starting point and more noise is added the higher the `strength`.
            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 `guidance_scale` is
                less than `1`).
            num_inference_steps (`int`, *optional*, defaults to 30):
                The number of denoising steps.
            guidance_scale (`float`, *optional*, defaults to 4.0):
                Guidance scale as defined in [Classifier-Free Diffusion
                Guidance](https://huggingface.co/papers/2207.12598).
            num_images_per_prompt (`int`, *optional*, defaults to 1):
                The number of images to generate per prompt.
            height (`int`, *optional*, defaults to self.unet.config.sample_size):
                The height in pixels of the generated image. If not provided, it is inferred from input image.
            width (`int`, *optional*, defaults to self.unet.config.sample_size):
                The width in pixels of the generated image. If not provided, it is inferred from input image.
            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.Tensor`, *optional*):
                Pre-generated noisy latents.
            prompt_embeds (`torch.Tensor`, *optional*):
                Pre-generated text embeddings.
            pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated pooled text embeddings.
            prompt_attention_mask (`torch.Tensor`, *optional*): Pre-generated attention mask for text embeddings.
            negative_prompt_embeds (`torch.Tensor`, *optional*):
                Pre-generated negative text embeddings.
            negative_prompt_attention_mask (`torch.Tensor`, *optional*):
                Pre-generated attention mask for negative text embeddings.
            negative_pooled_prompt_embeds (`torch.FloatTensor`, *optional*):
                Pre-generated negative pooled text embeddings.
            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.ImagePipelineOutput`] instead of a plain tuple.
            attention_kwargs:
                A kwargs dictionary that if specified is passed along to the `AttentionProcessor`.
            callback_on_step_end (`Callable`, *optional*):
                A function that calls at the end of each denoising steps during the inference.
            callback_on_step_end_tensor_inputs (`List`, *optional*):
                The list of tensor inputs for the `callback_on_step_end` function.
            system_prompt (`str`, *optional*):
                The system prompt to use for the image generation.
            cfg_trunc_ratio (`float`, *optional*, defaults to `1.0`):
                The ratio of the timestep interval to apply normalization-based guidance scale.
            cfg_normalization (`bool`, *optional*, defaults to `True`):
                Whether to apply normalization-based guidance scale.
            max_sequence_length (`int`, defaults to `512`):
                Maximum sequence length to use with the `prompt`.

        Examples:

        Returns:
            [`~pipelines.ImagePipelineOutput`] or `tuple`:
                If `return_dict` is `True`, [`~pipelines.ImagePipelineOutput`] is returned, otherwise a `tuple` is
                returned where the first element is a list with the generated images
        """
        # 1. Check strength
        if strength < 0 or strength > 1:
            raise ValueError(f"The value of strength should be in [0.0, 1.0] but is {strength}")

        # 2. Preprocess image
        init_image = self.image_processor.preprocess(image)
        init_image = init_image.to(dtype=torch.float32)

        # Get dimensions from image if not specified
        if height is None:
            height = init_image.shape[-2]
        if width is None:
            width = init_image.shape[-1]
            
        self._guidance_scale = guidance_scale
        self._attention_kwargs = attention_kwargs

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

        # 4. 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

        # 5. Encode input prompt
        (
            prompt_embeds,
            pooled_prompt_embeds,
            prompt_attention_mask,
            negative_prompt_embeds,
            negative_pooled_prompt_embeds,
            negative_prompt_attention_mask,
        ) = self.encode_prompt(
            prompt,
            self.do_classifier_free_guidance,
            negative_prompt=negative_prompt,
            num_images_per_prompt=num_images_per_prompt,
            device=device,
            prompt_embeds=prompt_embeds,
            pooled_prompt_embeds=pooled_prompt_embeds,
            negative_prompt_embeds=negative_prompt_embeds,
            negative_pooled_prompt_embeds=negative_pooled_prompt_embeds,
            prompt_attention_mask=prompt_attention_mask,
            negative_prompt_attention_mask=negative_prompt_attention_mask,
            max_sequence_length=max_sequence_length,
            system_prompt=system_prompt,
        )

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

        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=full_sigmas,
            mu=mu,
        )
        
        # 7. Adjust timesteps based on strength
        timesteps, num_inference_steps = self.get_timesteps(num_inference_steps, strength, device)
        if num_inference_steps < 1:
            raise ValueError(
                f"After adjusting the num_inference_steps by strength parameter: {strength}, the number of pipeline "
                f"steps is {num_inference_steps} which is < 1 and not appropriate for this pipeline."
            )
        
        latent_timestep = timesteps[:1].repeat(batch_size * num_images_per_prompt)
        
        # 8. Prepare latents
        latents = self.prepare_latents(
            init_image,
            latent_timestep,
            batch_size * num_images_per_prompt,
            self.transformer.config.in_channels,
            height,
            width,
            prompt_embeds.dtype,
            device,
            generator,
            latents,
        )

        num_warmup_steps = max(len(timesteps) - num_inference_steps * self.scheduler.order, 0)
        self._num_timesteps = len(timesteps)

        # 9. Denoising loop
        with self.progress_bar(total=num_inference_steps) as progress_bar:
            for i, t in enumerate(timesteps):
                # compute whether apply classifier-free truncation on this timestep
                do_classifier_free_truncation = (i + 1) / num_inference_steps > cfg_trunc_ratio
                
                # reverse the timestep since Lumina uses t=0 as the noise and t=1 as the image
                current_timestep = 1 - t / self.scheduler.config.num_train_timesteps
                # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
                current_timestep = current_timestep.expand(latents.shape[0])

                noise_pred_cond = self.transformer(
                    hidden_states=latents,
                    timestep=current_timestep,
                    encoder_hidden_states=prompt_embeds,
                    pooled_projections=pooled_prompt_embeds,
                    encoder_attention_mask=prompt_attention_mask,
                    return_dict=False,
                    attention_kwargs=self.attention_kwargs,
                )[0]

                # perform normalization-based guidance scale on a truncated timestep interval
                if self.do_classifier_free_guidance and not do_classifier_free_truncation:
                    noise_pred_uncond = self.transformer(
                        hidden_states=latents,
                        timestep=current_timestep,
                        encoder_hidden_states=negative_prompt_embeds,
                        pooled_projections=negative_pooled_prompt_embeds,
                        encoder_attention_mask=negative_prompt_attention_mask,
                        return_dict=False,
                        attention_kwargs=self.attention_kwargs,
                    )[0]
                    noise_pred = noise_pred_uncond + guidance_scale * (noise_pred_cond - noise_pred_uncond)
                    # apply normalization after classifier-free guidance
                    if cfg_normalization:
                        cond_norm = torch.norm(noise_pred_cond, dim=-1, keepdim=True)
                        noise_norm = torch.norm(noise_pred, dim=-1, keepdim=True)
                        noise_pred = noise_pred * (cond_norm / noise_norm)
                else:
                    noise_pred = noise_pred_cond

                # compute the previous noisy sample x_t -> x_t-1
                latents_dtype = latents.dtype
                noise_pred = -noise_pred
                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)
                    pooled_prompt_embeds = callback_outputs.pop("pooled_prompt_embeds", pooled_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()

        if not output_type == "latent":
            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)
        else:
            image = latents

        # Offload all models
        self.maybe_free_model_hooks()

        if not return_dict:
            return (image,)

        return ImagePipelineOutput(images=image)