app.py

import spaces
from diffusers import (
    StableDiffusionXLPipeline,
    EulerDiscreteScheduler,
    UNet2DConditionModel,
    AutoencoderTiny,
)
import torch
import os
from huggingface_hub import hf_hub_download
from compel import Compel, ReturnedEmbeddingsType
from gradio_promptweighting import PromptWeighting


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"


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"))
pipe = StableDiffusionXLPipeline.from_pretrained(
    BASE, unet=unet, torch_dtype=torch.float16, variant="fp16", safety_checker=False
).to("cuda")
unet = unet.to(dtype=torch.float16)

compel = Compel(
    tokenizer=[pipe.tokenizer, pipe.tokenizer_2],
    text_encoder=[pipe.text_encoder, pipe.text_encoder_2],
    returned_embeddings_type=ReturnedEmbeddingsType.PENULTIMATE_HIDDEN_STATES_NON_NORMALIZED,
    requires_pooled=[False, True],
)


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, prompt_w, guidance_scale, seed=1231231):
    generator = torch.manual_seed(seed)
    last_time = time.time()
    prompt_w = " ".join(
        [f"({p['prompt']}){p['scale']}" for p in prompt_w if p["prompt"]]
    )

    conditioning, pooled = compel([prompt + " " + prompt_w, ""])

    results = pipe(
        prompt_embeds=conditioning[0:1],
        pooled_prompt_embeds=pooled[0:1],
        negative_prompt_embeds=conditioning[1:2],
        negative_pooled_prompt_embeds=pooled[1:2],
        generator=generator,
        num_inference_steps=2,
        guidance_scale=guidance_scale,
        # 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: 80rem;
}
#intro{
    max-width: 100%;
    margin: 0 auto;
}
.generating {
    display: none
}
"""
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.Column():
                with gr.Group():
                    prompt = gr.Textbox(
                        placeholder="Insert your prompt here:",
                        max_lines=1,
                        label="Prompt",
                    )
                    prompt_w = PromptWeighting(
                        min=0,
                        max=3,
                        step=0.005,
                        show_label=False,
                    )

                with gr.Accordion("Advanced options", open=True):
                    seed = gr.Slider(
                        minimum=0,
                        maximum=12013012031030,
                        label="Seed",
                        step=1,
                    )
                    guidance_scale = gr.Slider(
                        minimum=0.0,
                        maximum=20.0,
                        label="Guidance scale",
                        value=0.0,
                        step=0.1,
                    )
                generate_bt = gr.Button("Generate")
            with gr.Column():
                image = gr.Image(type="filepath")

        inputs = [
            prompt,
            prompt_w,
            guidance_scale,
            seed,
        ]
        outputs = [image]

        gr.on(
            triggers=[
                prompt.input,
                prompt_w.input,
                generate_bt.click,
                guidance_scale.input,
                seed.input,
            ],
            fn=predict,
            inputs=inputs,
            outputs=outputs,
            show_progress="hidden",
            show_api=False,
            trigger_mode="always_last",
        )

demo.queue(api_open=False)
demo.launch()