app.py

import spaces
from diffusers import (
    StableDiffusionXLPipeline,
    EulerDiscreteScheduler,
    UNet2DConditionModel,
    AutoencoderTiny,
)
import torch
import os
from huggingface_hub import hf_hub_download


from PIL import Image
import gradio as gr
import time
from safetensors.torch import load_file
import time
import tempfile
from pathlib import Path

# Constants
BASE = "stabilityai/stable-diffusion-xl-base-1.0"
REPO = "ByteDance/SDXL-Lightning"
# 1-step
CHECKPOINT = "sdxl_lightning_2step_unet.safetensors"
taesd_model = "madebyollin/taesdxl"

# {
#     "1-Step": ["sdxl_lightning_1step_unet_x0.safetensors", 1],
#     "2-Step": ["sdxl_lightning_2step_unet.safetensors", 2],
#     "4-Step": ["sdxl_lightning_4step_unet.safetensors", 4],
#     "8-Step": ["sdxl_lightning_8step_unet.safetensors", 8],
# }


SFAST_COMPILE = os.environ.get("SFAST_COMPILE", "0") == "1"
SAFETY_CHECKER = os.environ.get("SAFETY_CHECKER", "0") == "1"
USE_TAESD = os.environ.get("USE_TAESD", "0") == "1"

# check if MPS is available OSX only M1/M2/M3 chips

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
torch_device = device
torch_dtype = torch.float16

print(f"SAFETY_CHECKER: {SAFETY_CHECKER}")
print(f"SFAST_COMPILE: {SFAST_COMPILE}")
print(f"USE_TAESD: {USE_TAESD}")
print(f"device: {device}")


unet = UNet2DConditionModel.from_config(BASE, subfolder="unet")
# .to( "cuda", torch.float16)
# unet.load_state_dict(load_file(hf_hub_download(REPO, CHECKPOINT), device="cuda"))
unet.load_state_dict(load_file(hf_hub_download(REPO, CHECKPOINT)))
pipe = StableDiffusionXLPipeline.from_pretrained(
    BASE, unet=unet, torch_dtype=torch.float16, variant="fp16", safety_checker=False
)
# .to("cuda")
unet = unet.to(dtype=torch.float16)

if USE_TAESD:
    pipe.vae = AutoencoderTiny.from_pretrained(
        taesd_model, torch_dtype=torch_dtype, use_safetensors=True
    ).to(device)


# Ensure sampler uses "trailing" timesteps.
pipe.scheduler = EulerDiscreteScheduler.from_config(
    pipe.scheduler.config, timestep_spacing="trailing"
)
pipe.set_progress_bar_config(disable=True)
if SAFETY_CHECKER:
    from safety_checker import StableDiffusionSafetyChecker
    from transformers import CLIPFeatureExtractor

    safety_checker = StableDiffusionSafetyChecker.from_pretrained(
        "CompVis/stable-diffusion-safety-checker"
    ).to(device)
    feature_extractor = CLIPFeatureExtractor.from_pretrained(
        "openai/clip-vit-base-patch32"
    )

    def check_nsfw_images(
        images: list[Image.Image],
    ) -> tuple[list[Image.Image], list[bool]]:
        safety_checker_input = feature_extractor(images, return_tensors="pt").to(device)
        has_nsfw_concepts = safety_checker(
            images=[images],
            clip_input=safety_checker_input.pixel_values.to(torch_device),
        )

        return images, has_nsfw_concepts


if SFAST_COMPILE:
    from sfast.compilers.diffusion_pipeline_compiler import compile, CompilationConfig

    # sfast compilation
    config = CompilationConfig.Default()
    try:
        import xformers

        config.enable_xformers = True
    except ImportError:
        print("xformers not installed, skip")
    try:
        import triton

        config.enable_triton = True
    except ImportError:
        print("Triton not installed, skip")
    # CUDA Graph is suggested for small batch sizes and small resolutions to reduce CPU overhead.
    # But it can increase the amount of GPU memory used.
    # For StableVideoDiffusionPipeline it is not needed.
    config.enable_cuda_graph = True

    pipe = compile(pipe, config)


@spaces.GPU
def predict(prompt, seed=1231231):
    generator = torch.manual_seed(seed)
    last_time = time.time()
    results = pipe(
        prompt=prompt,
        generator=generator,
        num_inference_steps=2,
        guidance_scale=0.0,
        # width=768,
        # height=768,
        output_type="pil",
    )
    print(f"Pipe took {time.time() - last_time} seconds")
    if SAFETY_CHECKER:
        images, has_nsfw_concepts = check_nsfw_images(results.images)
        if any(has_nsfw_concepts):
            gr.Warning("NSFW content detected.")
            return Image.new("RGB", (512, 512))
    image = results.images[0]
    with tempfile.NamedTemporaryFile(suffix=".jpg", delete=False) as tmpfile:
        image.save(tmpfile, "JPEG", quality=80, optimize=True, progressive=True)
        return Path(tmpfile.name)


css = """
#container{
    margin: 0 auto;
    max-width: 40rem;
}
#intro{
    max-width: 100%;
    margin: 0 auto;
}
"""
with gr.Blocks(css=css) as demo:
    with gr.Column(elem_id="container"):
        gr.Markdown(
            """
# SDXL-Lightning- Text To Image 2-Steps
**Model**: https://huggingface.co/ByteDance/SDXL-Lightning
            """,
            elem_id="intro",
        )
        with gr.Row():
            with gr.Row():
                prompt = gr.Textbox(
                    placeholder="Insert your prompt here:", scale=5, container=False
                )
                generate_bt = gr.Button("Generate", scale=1)

        image = gr.Image(type="filepath")
        with gr.Accordion("Advanced options", open=False):
            seed = gr.Slider(
                randomize=True, minimum=0, maximum=12013012031030, label="Seed", step=1
            )
        with gr.Accordion("Run with diffusers"):
            gr.Markdown(
                """## Running SDXL-Lightning with `diffusers`
```py
import torch
from diffusers import StableDiffusionXLPipeline, UNet2DConditionModel, EulerDiscreteScheduler
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file

base = "stabilityai/stable-diffusion-xl-base-1.0"
repo = "ByteDance/SDXL-Lightning"
ckpt = "sdxl_lightning_2step_unet.safetensors" # Use the correct ckpt for your step setting!

# Load model.
unet = UNet2DConditionModel.from_config(base, subfolder="unet").to("cuda", torch.float16)
unet.load_state_dict(load_file(hf_hub_download(repo, ckpt), device="cuda"))
pipe = StableDiffusionXLPipeline.from_pretrained(base, unet=unet, torch_dtype=torch.float16, variant="fp16").to("cuda")

# Ensure sampler uses "trailing" timesteps.
pipe.scheduler = EulerDiscreteScheduler.from_config(pipe.scheduler.config, timestep_spacing="trailing")

# Ensure using the same inference steps as the loaded model and CFG set to 0.
pipe("A girl smiling", num_inference_steps=2, guidance_scale=0).images[0].save("output.png")
```
            """
            )

        inputs = [prompt, seed]
        outputs = [image]
        generate_bt.click(
            fn=predict, inputs=inputs, outputs=outputs, show_progress=False
        )
        prompt.input(fn=predict, inputs=inputs, outputs=outputs, show_progress=False)
        seed.change(fn=predict, inputs=inputs, outputs=outputs, show_progress=False)

demo.queue()
demo.launch()