cjm-diffusers-utils

Some utility functions I frequently use with 🤗 diffusers.

Install

pip install cjm_diffusers_utils

How to use

import torch
from cjm_pytorch_utils.core import get_torch_device
device = get_torch_device()
dtype = torch.float16 if device == 'cuda' else torch.float32
device, dtype

('cuda', torch.float16)

pil_to_latent

from cjm_diffusers_utils.core import pil_to_latent
from PIL import Image
from diffusers import AutoencoderKL

model_name = "stabilityai/stable-diffusion-2-1"
vae = AutoencoderKL.from_pretrained(model_name, subfolder="vae").to(device=device, dtype=dtype)

img_path = img_path = '../images/cat.jpg'
src_img = Image.open(img_path).convert('RGB')
print(f"Source Image Size: {src_img.size}")

img_latents = pil_to_latent(src_img, vae)
print(f"Latent Dimensions: {img_latents.shape}")

Source Image Size: (768, 512)
Latent Dimensions: torch.Size([1, 4, 64, 96])

latent_to_pil

from cjm_diffusers_utils.core import latent_to_pil

decoded_img = latent_to_pil(img_latents, vae)
print(f"Decoded Image Size: {decoded_img.size}")

Decoded Image Size: (768, 512)

text_to_emb

from cjm_diffusers_utils.core import text_to_emb
from transformers import CLIPTextModel, CLIPTokenizer

# Load the tokenizer for the specified model
tokenizer = CLIPTokenizer.from_pretrained(model_name, subfolder="tokenizer")
# Load the text encoder for the specified model
text_encoder = CLIPTextModel.from_pretrained(model_name, subfolder="text_encoder").to(device=device, dtype=dtype)

prompt = "A cat sitting on the floor."
text_emb = text_to_emb(prompt, tokenizer, text_encoder)
text_emb.shape

torch.Size([2, 77, 1024])

prepare_noise_scheduler

from cjm_diffusers_utils.core import prepare_noise_scheduler
from diffusers import DEISMultistepScheduler

noise_scheduler = DEISMultistepScheduler.from_pretrained(model_name, subfolder='scheduler')
print(f"Number of timesteps: {len(noise_scheduler.timesteps)}")
print(noise_scheduler.timesteps[:10])

noise_scheduler = prepare_noise_scheduler(noise_scheduler, 70, 1.0)
print(f"Number of timesteps: {len(noise_scheduler.timesteps)}")
print(noise_scheduler.timesteps[:10])

Number of timesteps: 1000
tensor([999., 998., 997., 996., 995., 994., 993., 992., 991., 990.])
Number of timesteps: 70
tensor([999, 985, 970, 956, 942, 928, 913, 899, 885, 871])

prepare_depth_mask

from cjm_diffusers_utils.core import prepare_depth_mask

depth_map_path = '../images/depth-cat.png'
depth_map = Image.open(depth_map_path)
print(f"Depth map size: {depth_map.size}")

depth_mask = prepare_depth_mask(depth_map).to(device=device, dtype=dtype)
depth_mask.shape, depth_mask.min(), depth_mask.max()

Depth map size: (768, 512)

(torch.Size([1, 1, 64, 96]),
 tensor(-1., device='cuda:0', dtype=torch.float16),
 tensor(1., device='cuda:0', dtype=torch.float16))