Parallel Sampling of Diffusion Models
Last updated
Last updated
is by Andy Shih, Suneel Belkhale, Stefano Ermon, Dorsa Sadigh, Nima Anari.
The abstract from the paper is:
Diffusion models are powerful generative models but suffer from slow sampling, often taking 1000 sequential denoising steps for one sample. As a result, considerable efforts have been directed toward reducing the number of denoising steps, but these methods hurt sample quality. Instead of reducing the number of denoising steps (trading quality for speed), in this paper we explore an orthogonal approach: can we run the denoising steps in parallel (trading compute for speed)? In spite of the sequential nature of the denoising steps, we show that surprisingly it is possible to parallelize sampling via Picard iterations, by guessing the solution of future denoising steps and iteratively refining until convergence. With this insight, we present ParaDiGMS, a novel method to accelerate the sampling of pretrained diffusion models by denoising multiple steps in parallel. ParaDiGMS is the first diffusion sampling method that enables trading compute for speed and is even compatible with existing fast sampling techniques such as DDIM and DPMSolver. Using ParaDiGMS, we improve sampling speed by 2-4x across a range of robotics and image generation models, giving state-of-the-art sampling speeds of 0.2s on 100-step DiffusionPolicy and 16s on 1000-step StableDiffusion-v2 with no measurable degradation of task reward, FID score, or CLIP score.
The original codebase can be found at , and the pipeline was contributed by . β€οΈ
This pipeline improves sampling speed by running denoising steps in parallel, at the cost of increased total FLOPs. Therefore, it is better to call this pipeline when running on multiple GPUs. Otherwise, without enough GPU bandwidth sampling may be even slower than sequential sampling.
The two parameters to play with are parallel (batch size) and tolerance.
If it fits in memory, for a 1000-step DDPM you can aim for a batch size of around 100 (for example, 8 GPUs and batch_per_device=12 to get parallel=96). A higher batch size may not fit in memory, and lower batch size gives less parallelism.
For tolerance, using a higher tolerance may get better speedups but can risk sample quality degradation. If there is quality degradation with the default tolerance, then use a lower tolerance like 0.001.
For a 1000-step DDPM on 8 A100 GPUs, you can expect around a 3x speedup from compared to the by setting parallel=80 and tolerance=0.1.
π Diffusers offers for generating multiple prompts in parallel on multiple GPUs. But is designed for speeding up sampling of a single prompt by using multiple GPUs.
Make sure to check out the Schedulers to learn how to explore the tradeoff between scheduler speed and quality, and see the section to learn how to efficiently load the same components into multiple pipelines.
( vae: AutoencoderKLtext_encoder: CLIPTextModeltokenizer: CLIPTokenizerunet: UNet2DConditionModelscheduler: KarrasDiffusionSchedulerssafety_checker: StableDiffusionSafetyCheckerfeature_extractor: CLIPImageProcessorrequires_safety_checker: bool = True )
Parameters
tokenizer (CLIPTokenizer) β A CLIPTokenizer to tokenize text.
feature_extractor (CLIPImageProcessor) β A CLIPImageProcessor to extract features from generated images; used as inputs to the safety_checker.
Pipeline for text-to-image generation using a parallelized version of Stable Diffusion.
The pipeline also inherits the following loading methods:
__call__
Parameters
prompt (str or List[str], optional) β The prompt or prompts to guide image generation. If not defined, you need to pass prompt_embeds.
height (int, optional, defaults to self.unet.config.sample_size * self.vae_scale_factor) β The height in pixels of the generated image.
width (int, optional, defaults to self.unet.config.sample_size * self.vae_scale_factor) β The width in pixels of the generated image.
num_inference_steps (int, optional, defaults to 50) β The number of denoising steps. More denoising steps usually lead to a higher quality image at the expense of slower inference.
parallel (int, optional, defaults to 10) β The batch size to use when doing parallel sampling. More parallelism may lead to faster inference but requires higher memory usage and can also require more total FLOPs.
tolerance (float, optional, defaults to 0.1) β The error tolerance for determining when to slide the batch window forward for parallel sampling. Lower tolerance usually leads to less or no degradation. Higher tolerance is faster but can risk degradation of sample quality. The tolerance is specified as a ratio of the schedulerβs noise magnitude.
guidance_scale (float, optional, defaults to 7.5) β A higher guidance scale value encourages the model to generate images closely linked to the text prompt at the expense of lower image quality. Guidance scale is enabled when guidance_scale > 1.
negative_prompt (str or List[str], optional) β The prompt or prompts to guide what to not include in image generation. If not defined, you need to pass negative_prompt_embeds instead. Ignored when not using guidance (guidance_scale < 1).
num_images_per_prompt (int, optional, defaults to 1) β The number of images to generate per prompt.
latents (torch.FloatTensor, optional) β 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 is generated by sampling using the supplied random generator.
prompt_embeds (torch.FloatTensor, optional) β Pre-generated text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not provided, text embeddings are generated from the prompt input argument.
negative_prompt_embeds (torch.FloatTensor, optional) β Pre-generated negative text embeddings. Can be used to easily tweak text inputs (prompt weighting). If not provided, negative_prompt_embeds are generated from the negative_prompt input argument.
output_type (str, optional, defaults to "pil") β The output format of the generated image. Choose between PIL.Image or np.array.
callback (Callable, optional) β A function that calls every callback_steps steps during inference. The function is called with the following arguments: callback(step: int, timestep: int, latents: torch.FloatTensor).
callback_steps (int, optional, defaults to 1) β The frequency at which the callback function is called. If not specified, the callback is called at every step.
debug (bool, optional, defaults to False) β Whether or not to run in debug mode. In debug mode, torch.cumsum is evaluated using the CPU.
Returns
The call function to the pipeline for generation.
Examples:
Copied
disable_vae_slicing
( )
Disable sliced VAE decoding. If enable_vae_slicing was previously enabled, this method will go back to computing decoding in one step.
disable_vae_tiling
( )
Disable tiled VAE decoding. If enable_vae_tiling was previously enabled, this method will go back to computing decoding in one step.
enable_vae_slicing
( )
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.
enable_vae_tiling
( )
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.
encode_prompt
( promptdevicenum_images_per_promptdo_classifier_free_guidancenegative_prompt = Noneprompt_embeds: typing.Optional[torch.FloatTensor] = Nonenegative_prompt_embeds: typing.Optional[torch.FloatTensor] = Nonelora_scale: typing.Optional[float] = None )
Parameters
prompt (str or List[str], optional) β prompt to be encoded device β (torch.device): torch device
num_images_per_prompt (int) β number of images that should be generated per prompt
do_classifier_free_guidance (bool) β whether to use classifier free guidance or not
negative_prompt (str or List[str], optional) β 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).
prompt_embeds (torch.FloatTensor, optional) β 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) β 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.
lora_scale (float, optional) β A lora scale that will be applied to all LoRA layers of the text encoder if LoRA layers are loaded.
Encodes the prompt into text encoder hidden states.
( images: typing.Union[typing.List[PIL.Image.Image], numpy.ndarray]nsfw_content_detected: typing.Optional[typing.List[bool]] )
Parameters
images (List[PIL.Image.Image] or np.ndarray) β List of denoised PIL images of length batch_size or NumPy array of shape (batch_size, height, width, num_channels).
nsfw_content_detected (List[bool]) β List indicating whether the corresponding generated image contains βnot-safe-for-workβ (nsfw) content or None if safety checking could not be performed.
Output class for Stable Diffusion pipelines.
vae () β Variational Auto-Encoder (VAE) model to encode and decode images to and from latent representations.
text_encoder (CLIPTextModel) β Frozen text-encoder ().
unet () β A UNet2DConditionModel to denoise the encoded image latents.
scheduler () β A scheduler to be used in combination with unet to denoise the encoded image latents. Can be one of , , or .
safety_checker (StableDiffusionSafetyChecker) β Classification module that estimates whether generated images could be considered offensive or harmful. Please refer to the for more details about a modelβs potential harms.
This model inherits from . Check the superclass documentation for the generic methods implemented for all pipelines (downloading, saving, running on a particular device, etc.).
for loading textual inversion embeddings
for loading LoRA weights
for saving LoRA weights
for loading .ckpt files
( prompt: typing.Union[str, typing.List[str]] = Noneheight: typing.Optional[int] = Nonewidth: typing.Optional[int] = Nonenum_inference_steps: int = 50parallel: int = 10tolerance: float = 0.1guidance_scale: float = 7.5negative_prompt: typing.Union[str, typing.List[str], NoneType] = Nonenum_images_per_prompt: typing.Optional[int] = 1eta: float = 0.0generator: typing.Union[torch._C.Generator, typing.List[torch._C.Generator], NoneType] = Nonelatents: typing.Optional[torch.FloatTensor] = Noneprompt_embeds: typing.Optional[torch.FloatTensor] = Nonenegative_prompt_embeds: typing.Optional[torch.FloatTensor] = Noneoutput_type: typing.Optional[str] = 'pil'return_dict: bool = Truecallback: typing.Union[typing.Callable[[int, int, torch.FloatTensor], NoneType], NoneType] = Nonecallback_steps: int = 1cross_attention_kwargs: typing.Union[typing.Dict[str, typing.Any], NoneType] = Nonedebug: bool = False ) β or tuple
eta (float, optional, defaults to 0.0) β Corresponds to parameter eta (Ξ·) from the paper. Only applies to the , and is ignored in other schedulers.
generator (torch.Generator or List[torch.Generator], optional) β A to make generation deterministic.
return_dict (bool, optional, defaults to True) β Whether or not to return a instead of a plain tuple.
cross_attention_kwargs (dict, optional) β A kwargs dictionary that if specified is passed along to the AttentionProcessor as defined in .
or tuple
If return_dict is True, is returned, otherwise a tuple is returned where the first element is a list with the generated images and the second element is a list of bools indicating whether the corresponding generated image contains βnot-safe-for-workβ (nsfw) content.