ostris/Flex.2-preview

Instalación básica con Diffusers:

pip install -U diffusers transformers accelerate

Ejemplo rápido:

import torch
from diffusers import DiffusionPipeline

# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("ostris/Flex.2-preview", dtype=torch.bfloat16, device_map="cuda")

prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
image = pipe(prompt).images[0]

Uso recomendado con AutoPipelineForText2Image, incluyendo inpainting y control:

pip install --upgrade torch accelerate transformers diffusers

import torch
from diffusers import AutoPipelineForText2Image
from diffusers.utils import load_image

name_or_path = "ostris/Flex.2-preview"
inpaint_image = load_image("https://ostris.com/wp-content/uploads/2025/04/dog.jpg")
inpaint_mask = load_image("https://ostris.com/wp-content/uploads/2025/04/dog_mask.jpg")
control_image = load_image("https://ostris.com/wp-content/uploads/2025/04/dog_depth.jpg")
dtype = torch.bfloat16

pipe = AutoPipelineForText2Image.from_pretrained(
    name_or_path,
    custom_pipeline=name_or_path,
    torch_dtype=dtype
).to("cuda")

image = pipe(
    prompt="A white friendly robotic dog sitting on a bench",
    inpaint_image=inpaint_image,
    inpaint_mask=inpaint_mask,
    control_image=control_image,
    control_strength=0.5,
    control_stop=0.33,
    height=1024,
    width=1024,
    guidance_scale=3.5,
    num_inference_steps=50,
    generator=torch.Generator("cpu").manual_seed(42)
).images[0]

image.save(f"robot_dog.png")

Para tarjetas de consumo con menos de 24 GB de VRAM, la página recomienda cuantización con torchao:

pip install --upgrade torchao

También puede usarse en ComfyUI mediante el nodo Flex2 Conditioner de ComfyUI-FlexTools, junto con comfyui_controlnet_aux para generar imágenes de control como pose y profundidad.