Lambda Diffusers

Additional models and pipelines for 🤗 Diffusers created by Lambda Labs

Stable Diffusion Image Variations
Pokemon text to image

🦄 Other exciting ML projects at Lambda: ML Times, Distributed Training Guide, Text2Video, GPU Benchmark.

Installation

git clone https://github.com/LambdaLabsML/lambda-diffusers.git
cd lambda-diffusers
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Stable Diffusion Image Variations

A fine-tuned version of Stable Diffusion conditioned on CLIP image embeddings to enabel Image Variations.

Download the weights ported to 🤗 Diffusers here.
See the original training repo here.

Usage

from diffusers import StableDiffusionImageVariationPipeline
from PIL import Image

device = "cuda:0"
sd_pipe = StableDiffusionImageVariationPipeline.from_pretrained(
  "lambdalabs/sd-image-variations-diffusers",
  revision="v2.0",
  )
sd_pipe = sd_pipe.to(device)

im = Image.open("path/to/image.jpg")
tform = transforms.Compose([
    transforms.ToTensor(),
    transforms.Resize(
        (224, 224),
        interpolation=transforms.InterpolationMode.BICUBIC,
        antialias=False,
        ),
    transforms.Normalize(
      [0.48145466, 0.4578275, 0.40821073],
      [0.26862954, 0.26130258, 0.27577711]),
])
inp = tform(im).to(device)

out = sd_pipe(inp, guidance_scale=3)
out["images"][0].save("result.jpg")

Pokemon text to image

Stable Diffusion fine tuned on Pokémon by Lambda Labs.

Put in a text prompt and generate your own Pokémon character, no "prompt engineering" required!

If you want to find out how to train your own Stable Diffusion variants, see this example from Lambda Labs.

Girl with a pearl earring, Cute Obama creature, Donald Trump, Boris Johnson, Totoro, Hello Kitty

Model description

Trained on BLIP captioned Pokémon images using 2xA6000 GPUs on Lambda GPU Cloud for around 15,000 step (about 6 hours, at a cost of about $10).

Usage

import torch
from diffusers import StableDiffusionPipeline
from torch import autocast

pipe = StableDiffusionPipeline.from_pretrained("lambdalabs/sd-pokemon-diffusers", torch_dtype=torch.float16)
pipe = pipe.to("cuda")

prompt = "Yoda"
scale = 10
n_samples = 4

# Sometimes the nsfw checker is confused by the Pokémon images, you can disable
# it at your own risk here
disable_safety = False

if disable_safety:
  def null_safety(images, **kwargs):
      return images, False
  pipe.safety_checker = null_safety

with autocast("cuda"):
  images = pipe(n_samples*[prompt], guidance_scale=scale).images

for idx, im in enumerate(images):
  im.save(f"{idx:06}.png")

Benchmarking inference

We have updated the original benchmark using xformers and a newer version of Diffusers, see the new results here (original results can still be found here).

Usage

Ensure that NVIDIA container toolkit is installed on your system and then run the following:

git clone https://github.com/LambdaLabsML/lambda-diffusers.git
cd lambda-diffusers/scripts
make bench

Currently xformers does not support H100. The "without xformers" results below are generated by running the benchmark with --xformers no (can be set in scripts/Makefile)