CompVis/stable-diffusion-v1-3

Recomendamos usar la biblioteca de Difusores de Hugging Face para ejecutar Stable Diffusion.

pip install --upgrade diffusers transformers scipy

import torch
from torch import autocast
from diffusers import StableDiffusionPipeline

model_id = "CompVis/stable-diffusion-v1-3"
device = "cuda"

pipe = StableDiffusionPipeline.from_pretrained(model_id)
pipe = pipe.to(device)

prompt = "una foto de un astronauta montando un caballo en Marte"
with autocast("cuda"):
    image = pipe(prompt, guidance_scale=7.5)["sample"][0]

image.save("astronaut_rides_horse.png")

# Si tienes menos de 10GB de RAM GPU, puedes cargar el modelo en precisión float16
pipe = StableDiffusionPipeline.from_pretrained(model_id, torch_dtype=torch.float16)
pipe = pipe.to(device)

with autocast("cuda"):
    image = pipe(prompt, guidance_scale=7.5)["sample"][0]

image.save("astronaut_rides_horse.png")

# Para cambiar el programador de ruido
from diffusers import StableDiffusionPipeline, LMSDiscreteScheduler

scheduler = LMSDiscreteScheduler(beta_start=0.00085, beta_end=0.012, beta_schedule="scaled_linear", num_train_timesteps=1000)
pipe = StableDiffusionPipeline.from_pretrained(model_id, scheduler=scheduler, use_auth_token=True)
pipe = pipe.to("cuda")

prompt = "una foto de un astronauta montando un caballo en Marte"
with autocast("cuda"):
    image = pipe(prompt, guidance_scale=7.5)["sample"][0]

image.save("astronaut_rides_horse.png")