🤗 Diffusers provides pretrained diffusion models across multiple modalities, such as vision and audio, and serves as a modular toolbox for inference and training of diffusion models.

More precisely, 🤗 Diffusers offers:

• State-of-the-art diffusion pipelines that can be run in inference with just a couple of lines of code (see src/diffusers/pipelines).
• Various noise schedulers that can be used interchangeably for the prefered speed vs. quality trade-off in inference (see src/diffusers/schedulers).
• Multiple types of models, such as UNet, that can be used as building blocks in an end-to-end diffusion system (see src/diffusers/models).
• Training examples to show how to train the most popular diffusion models (see examples).

Definitions

Models: Neural network that models p_\theta(\mathbf{x}_{t-1}|\mathbf{x}_t) (see image below) and is trained end-to-end to denoise a noisy input to an image. Examples: UNet, Conditioned UNet, 3D UNet, Transformer UNet

Figure from DDPM paper (https://arxiv.org/abs/2006.11239).

Schedulers: Algorithm class for both inference and training. The class provides functionality to compute previous image according to alpha, beta schedule as well as predict noise for training. Examples: DDPM, DDIM, PNDM, DEIS

Sampling and training algorithms. Figure from DDPM paper (https://arxiv.org/abs/2006.11239).

Diffusion Pipeline: End-to-end pipeline that includes multiple diffusion models, possible text encoders, ... Examples: Glide, Latent-Diffusion, Imagen, DALL-E 2

Figure from ImageGen (https://imagen.research.google/).

Philosophy

• Readability and clarity is prefered over highly optimized code. A strong importance is put on providing readable, intuitive and elementary code design. E.g., the provided schedulers are separated from the provided models and provide well-commented code that can be read alongside the original paper.
• Diffusers is modality independent and focusses on providing pretrained models and tools to build systems that generate continous outputs, e.g. vision and audio.
• Diffusion models and schedulers are provided as consise, elementary building blocks whereas diffusion pipelines are a collection of end-to-end diffusion systems that can be used out-of-the-box, should stay as close as possible to their original implementation and can include components of other library, such as text-encoders. Examples for diffusion pipelines are Glide and Latent Diffusion.

Quickstart

Installation

pip install diffusers  # should install diffusers 0.0.4

1. diffusers as a toolbox for schedulers and models

diffusers is more modularized than transformers. The idea is that researchers and engineers can use only parts of the library easily for the own use cases. It could become a central place for all kinds of models, schedulers, training utils and processors that one can mix and match for one's own use case. Both models and schedulers should be load- and saveable from the Hub.

For more examples see schedulers and models

Example for DDPM:

import torch
from diffusers import UNetModel, DDPMScheduler
import PIL
import numpy as np
import tqdm

generator = torch.manual_seed(0)
torch_device = "cuda" if torch.cuda.is_available() else "cpu"

# 1. Load models
noise_scheduler = DDPMScheduler.from_config("fusing/ddpm-lsun-church", tensor_format="pt")
unet = UNetModel.from_pretrained("fusing/ddpm-lsun-church").to(torch_device)

# 2. Sample gaussian noise
image = torch.randn(
    (1, unet.in_channels, unet.resolution, unet.resolution),
    generator=generator,
)
image = image.to(torch_device)

# 3. Denoise
num_prediction_steps = len(noise_scheduler)
for t in tqdm.tqdm(reversed(range(num_prediction_steps)), total=num_prediction_steps):
    # predict noise residual
    with torch.no_grad():
        residual = unet(image, t)

    # predict previous mean of image x_t-1
    pred_prev_image = noise_scheduler.step(residual, image, t)

    # optionally sample variance
    variance = 0
    if t > 0:
        noise = torch.randn(image.shape, generator=generator).to(image.device)
        variance = noise_scheduler.get_variance(t).sqrt() * noise

    # set current image to prev_image: x_t -> x_t-1
    image = pred_prev_image + variance

# 5. process image to PIL
image_processed = image.cpu().permute(0, 2, 3, 1)
image_processed = (image_processed + 1.0) * 127.5
image_processed = image_processed.numpy().astype(np.uint8)
image_pil = PIL.Image.fromarray(image_processed[0])

# 6. save image
image_pil.save("test.png")

Example for DDIM:

import torch
from diffusers import UNetModel, DDIMScheduler
import PIL
import numpy as np
import tqdm

generator = torch.manual_seed(0)
torch_device = "cuda" if torch.cuda.is_available() else "cpu"

# 1. Load models
noise_scheduler = DDIMScheduler.from_config("fusing/ddpm-celeba-hq", tensor_format="pt")
unet = UNetModel.from_pretrained("fusing/ddpm-celeba-hq").to(torch_device)

# 2. Sample gaussian noise
image = torch.randn(
   (1, unet.in_channels, unet.resolution, unet.resolution),
   generator=generator,
)
image = image.to(torch_device)

# 3. Denoise                                                                                                                                           
num_inference_steps = 50
eta = 0.0  # <- deterministic sampling

for t in tqdm.tqdm(reversed(range(num_inference_steps)), total=num_inference_steps):
    # 1. predict noise residual
    orig_t = len(noise_scheduler) // num_inference_steps * t

    with torch.no_grad():
        residual = unet(image, orig_t)

    # 2. predict previous mean of image x_t-1
    pred_prev_image = noise_scheduler.step(residual, image, t, num_inference_steps, eta)

    # 3. optionally sample variance
    variance = 0
    if eta > 0:
        noise = torch.randn(image.shape, generator=generator).to(image.device)
        variance = noise_scheduler.get_variance(t).sqrt() * eta * noise

    # 4. set current image to prev_image: x_t -> x_t-1
    image = pred_prev_image + variance

# 5. process image to PIL
image_processed = image.cpu().permute(0, 2, 3, 1)
image_processed = (image_processed + 1.0) * 127.5
image_processed = image_processed.numpy().astype(np.uint8)
image_pil = PIL.Image.fromarray(image_processed[0])

# 6. save image
image_pil.save("test.png")

Examples for other modalities:

Diffuser for planning in reinforcement learning (currenlty only inference):

2. diffusers as a collection of popular Diffusion systems (Glide, Dalle, ...)

For more examples see pipelines.

Example image generation with PNDM

from diffusers import PNDM, UNetModel, PNDMScheduler
import PIL.Image
import numpy as np
import torch

model_id = "fusing/ddim-celeba-hq"

model = UNetModel.from_pretrained(model_id)
scheduler = PNDMScheduler()

# load model and scheduler
pndm = PNDM(unet=model, noise_scheduler=scheduler)

# run pipeline in inference (sample random noise and denoise)
with torch.no_grad():
    image = pndm()

# process image to PIL
image_processed = image.cpu().permute(0, 2, 3, 1)
image_processed = (image_processed + 1.0) / 2
image_processed = torch.clamp(image_processed, 0.0, 1.0)
image_processed = image_processed * 255
image_processed = image_processed.numpy().astype(np.uint8)
image_pil = PIL.Image.fromarray(image_processed[0])

# save image
image_pil.save("test.png")

Example 1024x1024 image generation with SDE VE

See paper for more information on SDE VE.

from diffusers import DiffusionPipeline
import torch
import PIL.Image
import numpy as np

torch.manual_seed(32)

score_sde_sv = DiffusionPipeline.from_pretrained("fusing/ffhq_ncsnpp")

# Note this might take up to 3 minutes on a GPU
image = score_sde_sv(num_inference_steps=2000)

image = image.permute(0, 2, 3, 1).cpu().numpy()
image = np.clip(image * 255, 0, 255).astype(np.uint8)
image_pil = PIL.Image.fromarray(image[0])

# save image
image_pil.save("test.png")

Example 32x32 image generation with SDE VP

See paper for more information on SDE VE.

from diffusers import DiffusionPipeline
import torch
import PIL.Image
import numpy as np

torch.manual_seed(32)

score_sde_sv = DiffusionPipeline.from_pretrained("fusing/cifar10-ddpmpp-deep-vp")

# Note this might take up to 3 minutes on a GPU
image = score_sde_sv(num_inference_steps=1000)

image = image.permute(0, 2, 3, 1).cpu().numpy()
image = np.clip(image * 255, 0, 255).astype(np.uint8)
image_pil = PIL.Image.fromarray(image[0])

# save image
image_pil.save("test.png")

Text to Image generation with Latent Diffusion

Note: To use latent diffusion install transformers from this branch.

from diffusers import DiffusionPipeline

ldm = DiffusionPipeline.from_pretrained("fusing/latent-diffusion-text2im-large")

generator = torch.manual_seed(42)

prompt = "A painting of a squirrel eating a burger"
image = ldm([prompt], generator=generator, eta=0.3, guidance_scale=6.0, num_inference_steps=50)

image_processed = image.cpu().permute(0, 2, 3, 1)
image_processed = image_processed  * 255.
image_processed = image_processed.numpy().astype(np.uint8)
image_pil = PIL.Image.fromarray(image_processed[0])

# save image
image_pil.save("test.png")

Text to speech with GradTTS and BDDMPipeline

import torch
from diffusers import BDDMPipeline, GradTTSPipeline

torch_device = "cuda"

# load grad tts and bddm pipelines
grad_tts = GradTTSPipeline.from_pretrained("fusing/grad-tts-libri-tts")
bddm = BDDMPipeline.from_pretrained("fusing/diffwave-vocoder-ljspeech")

text = "Hello world, I missed you so much."

# generate mel spectograms using text
mel_spec = grad_tts(text, torch_device=torch_device)

#  generate the speech by passing mel spectograms to BDDMPipeline pipeline
generator = torch.manual_seed(42)
audio = bddm(mel_spec, generator, torch_device=torch_device)

# save generated audio
from scipy.io.wavfile import write as wavwrite
sampling_rate = 22050
wavwrite("generated_audio.wav", sampling_rate, audio.squeeze().cpu().numpy())

TODO

• Create common API for models
• Add tests for models
• Adapt schedulers for training
• Write google colab for training
• Write docs / Think about how to structure docs
• Add tests to circle ci
• Add Diffusion LM models
• Add more vision models
• Add more speech models
• Add RL model
• Add FID and KID metrics