# Copyright 2025 Alibaba Ovis-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 Any, Callable, Dict, List, Optional, Union import numpy as np import torch from transformers import Qwen2TokenizerFast, Qwen3Model from ...image_processor import VaeImageProcessor from ...models import AutoencoderKL, OvisImageTransformer2DModel from ...schedulers import FlowMatchEulerDiscreteScheduler from ...utils import is_torch_xla_available, logging, replace_example_docstring from ...utils.torch_utils import randn_tensor from ..pipeline_utils import DiffusionPipeline from .pipeline_output import OvisImagePipelineOutput 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 OvisImagePipeline >>> pipe = OvisImagePipeline.from_pretrained("AIDC-AI/Ovis-Image-7B", torch_dtype=torch.bfloat16) >>> pipe.to("cuda") >>> prompt = 'A creative 3D artistic render where the text "OVIS-IMAGE" is written in a bold, expressive handwritten brush style using thick, wet oil paint. The paint is a mix of vibrant rainbow colors (red, blue, yellow) swirling together like toothpaste or impasto art. You can see the ridges of the brush bristles and the glossy, wet texture of the paint. The background is a clean artist\'s canvas. Dynamic lighting creates soft shadows behind the floating paint strokes. Colorful, expressive, tactile texture, 4k detail.' >>> image = pipe(prompt, negative_prompt="", num_inference_steps=50, guidance_scale=5.0).images[0] >>> image.save("ovis_image.png") ``` """ 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 OvisImagePipeline( DiffusionPipeline, ): r""" The Ovis-Image pipeline for text-to-image generation. Reference: https://github.com/AIDC-AI/Ovis-Image Args: transformer ([`OvisImageTransformer2DModel`]): Conditional Transformer (MMDiT) architecture to denoise the encoded image latents. scheduler ([`FlowMatchEulerDiscreteScheduler`]): A scheduler to be used in combination with `transformer` to denoise the encoded image latents. vae ([`AutoencoderKL`]): Variational Auto-Encoder (VAE) Model to encode and decode images to and from latent representations. text_encoder ([`Qwen3Model`]): Text encoder of class [Qwen3Model](https://huggingface.co/docs/transformers/en/model_doc/qwen3#transformers.Qwen3Model). tokenizer (`Qwen2TokenizerFast`): Tokenizer of class [Qwen2TokenizerFast](https://huggingface.co/docs/transformers/en/model_doc/qwen2#transformers.Qwen2TokenizerFast). """ 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: Qwen3Model, tokenizer: Qwen2TokenizerFast, transformer: OvisImageTransformer2DModel, ): super().__init__() 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) if getattr(self, "vae", None) else 8 # Ovis-Image 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.system_prompt = "Describe the image by detailing the color, quantity, text, shape, size, texture, spatial relationships of the objects and background: " self.user_prompt_begin_id = 28 self.tokenizer_max_length = 256 + self.user_prompt_begin_id self.default_sample_size = 128 def _get_messages( self, prompt: Union[str, List[str]] = None, ): prompt = [prompt] if isinstance(prompt, str) else prompt messages = [] for each_prompt in prompt: message = [ { "role": "user", "content": self.system_prompt + each_prompt, } ] message = self.tokenizer.apply_chat_template( message, tokenize=False, add_generation_prompt=True, enable_thinking=False ) messages.append(message) return messages def _get_ovis_prompt_embeds( self, prompt: Union[str, List[str]] = None, num_images_per_prompt: int = 1, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None, ): device = device or self._execution_device dtype = dtype or self.text_encoder.dtype messages = self._get_messages(prompt) batch_size = len(messages) tokens = self.tokenizer( messages, padding="max_length", truncation=True, max_length=self.tokenizer_max_length, return_tensors="pt", add_special_tokens=False, ) input_ids = tokens.input_ids.to(device) attention_mask = tokens.attention_mask.to(device) outputs = self.text_encoder( input_ids=input_ids, attention_mask=attention_mask, ) prompt_embeds = outputs.last_hidden_state prompt_embeds = prompt_embeds * attention_mask[..., None] prompt_embeds = prompt_embeds[:, self.user_prompt_begin_id :, :] _, seq_len, _ = prompt_embeds.shape # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1) prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1) return prompt_embeds def encode_prompt( self, prompt: Union[str, List[str]], device: Optional[torch.device] = None, num_images_per_prompt: int = 1, prompt_embeds: Optional[torch.FloatTensor] = None, ): r""" Args: prompt (`str`, *optional*): prompt to be encoded device: (`torch.device`): torch device num_images_per_prompt (`int`): number of images that should be generated per prompt prompt_embeds (`torch.FloatTensor`, *optional*): Pre-generated text embeddings. Can be used to easily tweak text inputs, *e.g.* prompt weighting. If not provided, text embeddings will be generated from `prompt` input argument. """ device = device or self._execution_device if prompt_embeds is None: prompt_embeds = self._get_ovis_prompt_embeds( prompt=prompt, device=device, num_images_per_prompt=num_images_per_prompt, ) dtype = self.text_encoder.dtype if self.text_encoder is not None else self.transformer.dtype text_ids = torch.zeros(prompt_embeds.shape[1], 3) text_ids[..., 1] = text_ids[..., 1] + torch.arange(prompt_embeds.shape[1])[None, :] text_ids[..., 2] = text_ids[..., 2] + torch.arange(prompt_embeds.shape[1])[None, :] text_ids = text_ids.to(device=device, dtype=dtype) return prompt_embeds, text_ids def check_inputs( self, prompt, height, width, negative_prompt=None, prompt_embeds=None, negative_prompt_embeds=None, callback_on_step_end_tensor_inputs=None, max_sequence_length=None, ): if height % (self.vae_scale_factor * 2) != 0 or width % (self.vae_scale_factor * 2) != 0: logger.warning( f"`height` and `width` have to be divisible by {self.vae_scale_factor * 2} but are {height} and {width}. Dimensions will be resized accordingly" ) if callback_on_step_end_tensor_inputs is not None and not all( k in self._callback_tensor_inputs for k in callback_on_step_end_tensor_inputs ): raise ValueError( f"`callback_on_step_end_tensor_inputs` has to be in {self._callback_tensor_inputs}, but found {[k for k in callback_on_step_end_tensor_inputs if k not in self._callback_tensor_inputs]}" ) if prompt is not None and prompt_embeds is not None: raise ValueError( f"Cannot forward both `prompt`: {prompt} and `prompt_embeds`: {prompt_embeds}. Please make sure to" " only forward one of the two." ) elif prompt is None and prompt_embeds is None: raise ValueError( "Provide either `prompt` or `prompt_embeds`. Cannot leave both `prompt` and `prompt_embeds` undefined." ) elif prompt is not None and (not isinstance(prompt, str) and not isinstance(prompt, list)): raise ValueError(f"`prompt` has to be of type `str` or `list` but is {type(prompt)}") 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." ) if max_sequence_length is not None and max_sequence_length > 256: raise ValueError(f"`max_sequence_length` cannot be greater than 256 but is {max_sequence_length}") @staticmethod def _prepare_latent_image_ids(batch_size, height, width, device, dtype): 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.reshape( latent_image_id_height * latent_image_id_width, latent_image_id_channels ) return latent_image_ids.to(device=device, dtype=dtype) @staticmethod def _pack_latents(latents, batch_size, num_channels_latents, height, width): latents = latents.view(batch_size, num_channels_latents, height // 2, 2, width // 2, 2) latents = latents.permute(0, 2, 4, 1, 3, 5) latents = latents.reshape(batch_size, (height // 2) * (width // 2), num_channels_latents * 4) return latents @staticmethod def _unpack_latents(latents, height, width, vae_scale_factor): batch_size, num_patches, channels = latents.shape # 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 // 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, width) return latents def prepare_latents( self, batch_size, num_channels_latents, height, width, dtype, device, generator, latents=None, ): # 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 // 2, width // 2, device, dtype) return latents.to(device=device, dtype=dtype), latent_image_ids if isinstance(generator, list) and len(generator) != batch_size: raise ValueError( f"You have passed a list of generators of length {len(generator)}, but requested an effective batch" f" size of {batch_size}. Make sure the batch size matches the length of the generators." ) 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 // 2, width // 2, device, dtype) return latents, latent_image_ids @property def guidance_scale(self): return self._guidance_scale @property def joint_attention_kwargs(self): return self._joint_attention_kwargs @property def num_timesteps(self): return self._num_timesteps @property def current_timestep(self): return self._current_timestep @property def interrupt(self): return self._interrupt @torch.no_grad() @replace_example_docstring(EXAMPLE_DOC_STRING) def __call__( self, prompt: Union[str, List[str]] = None, negative_prompt: Union[str, List[str]] = "", guidance_scale: float = 5.0, height: Optional[int] = None, width: Optional[int] = None, num_inference_steps: int = 50, sigmas: Optional[List[float]] = None, 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, negative_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 = 256, ): 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. 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 not greater than `1`). guidance_scale (`float`, *optional*, defaults to 1.0): True classifier-free guidance (guidance scale) is enabled when `guidance_scale` > 1 and `negative_prompt` is provided. 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. 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 be 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. 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. 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.ovis_image.OvisImagePipelineOutput`] or `tuple`: [`~pipelines.ovis_image.OvisImagePipelineOutput`] 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, height, width, negative_prompt=negative_prompt, prompt_embeds=prompt_embeds, negative_prompt_embeds=negative_prompt_embeds, callback_on_step_end_tensor_inputs=callback_on_step_end_tensor_inputs, max_sequence_length=max_sequence_length, ) 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 do_classifier_free_guidance = guidance_scale > 1 ( prompt_embeds, text_ids, ) = self.encode_prompt( prompt=prompt, prompt_embeds=prompt_embeds, device=device, num_images_per_prompt=num_images_per_prompt, ) if do_classifier_free_guidance: ( negative_prompt_embeds, negative_text_ids, ) = self.encode_prompt( prompt=negative_prompt, prompt_embeds=negative_prompt_embeds, device=device, num_images_per_prompt=num_images_per_prompt, ) # 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, ) # 5. Prepare timesteps sigmas = np.linspace(1.0, 1 / num_inference_steps, num_inference_steps) if sigmas is None else sigmas if hasattr(self.scheduler.config, "use_flow_sigmas") and self.scheduler.config.use_flow_sigmas: sigmas = None 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) if self.joint_attention_kwargs is None: self._joint_attention_kwargs = {} # 6. Denoising loop # We set the index here to remove DtoH sync, helpful especially during compilation. # Check out more details here: https://github.com/huggingface/diffusers/pull/11696 self.scheduler.set_begin_index(0) 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) with self.transformer.cache_context("cond"): noise_pred = self.transformer( hidden_states=latents, timestep=timestep / 1000, encoder_hidden_states=prompt_embeds, txt_ids=text_ids, img_ids=latent_image_ids, return_dict=False, )[0] if do_classifier_free_guidance: with self.transformer.cache_context("uncond"): neg_noise_pred = self.transformer( hidden_states=latents, timestep=timestep / 1000, encoder_hidden_states=negative_prompt_embeds, txt_ids=negative_text_ids, img_ids=latent_image_ids, return_dict=False, )[0] noise_pred = neg_noise_pred + guidance_scale * (noise_pred - neg_noise_pred) # compute the previous noisy sample x_t -> x_t-1 latents_dtype = latents.dtype latents = self.scheduler.step(noise_pred, t, latents, return_dict=False)[0] if latents.dtype != latents_dtype: if torch.backends.mps.is_available(): # some platforms (eg. apple mps) misbehave due to a pytorch bug: https://github.com/pytorch/pytorch/pull/99272 latents = latents.to(latents_dtype) if callback_on_step_end is not None: callback_kwargs = {} for k in callback_on_step_end_tensor_inputs: callback_kwargs[k] = locals()[k] callback_outputs = callback_on_step_end(self, i, t, callback_kwargs) latents = callback_outputs.pop("latents", latents) prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds) # call the callback, if provided if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0): progress_bar.update() if XLA_AVAILABLE: xm.mark_step() self._current_timestep = None if output_type == "latent": image = latents 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 OvisImagePipelineOutput(images=image)