app.py

import gradio as gr
from diffusers import StableDiffusionXLPipeline, UNet2DConditionModel, EulerDiscreteScheduler, LCMScheduler, AutoencoderKL,DiffusionPipeline
import torch
from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
import spaces


def save_image(img):
    unique_name = str(uuid.uuid4()) + ".png"
    img.save(unique_name)
    return unique_name

def randomize_seed_fn(seed: int, randomize_seed: bool) -> int:
    if randomize_seed:
        seed = random.randint(0, MAX_SEED)
    return seed

MAX_SEED = np.iinfo(np.int32).max
vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)

### RealVisXL V3 ###
RealVisXLv3_pipe = DiffusionPipeline.from_pretrained(
        "SG161222/RealVisXL_V3.0",
        torch_dtype=torch.float16,
        use_safetensors=True,
        add_watermarker=False,
        variant="fp16"
    )
RealVisXLv3_pipe.to("cuda")
### RealVisXL V4 ###
RealVisXLv4_pipe = DiffusionPipeline.from_pretrained(
        "SG161222/RealVisXL_V4.0",
        torch_dtype=torch.float16,
        use_safetensors=True,
        add_watermarker=False,
        variant="fp16"
    )
RealVisXLv4_pipe.to("cuda")
### SDXL Turbo #### 
pipe_turbo = StableDiffusionXLPipeline.from_pretrained("stabilityai/sdxl-turbo",
                                                       vae=vae,
                                                       torch_dtype=torch.float16,
                                                       variant="fp16"
                                                      )
pipe_turbo.to("cuda")

### SDXL Lightning ### 
base = "stabilityai/stable-diffusion-xl-base-1.0"
repo = "ByteDance/SDXL-Lightning"
ckpt = "sdxl_lightning_1step_unet_x0.safetensors" 

unet = UNet2DConditionModel.from_config(base, subfolder="unet").to(torch.float16)
unet.load_state_dict(load_file(hf_hub_download(repo, ckpt)))
pipe_lightning = StableDiffusionXLPipeline.from_pretrained(base,
                                                           unet=unet,
                                                           vae=vae,
                                                           text_encoder=pipe_turbo.text_encoder,
                                                           text_encoder_2=pipe_turbo.text_encoder_2,
                                                           tokenizer=pipe_turbo.tokenizer,
                                                           tokenizer_2=pipe_turbo.tokenizer_2,
                                                           torch_dtype=torch.float16,
                                                           variant="fp16"
                                                          )#.to("cuda")
del unet
pipe_lightning.scheduler = EulerDiscreteScheduler.from_config(pipe_lightning.scheduler.config, timestep_spacing="trailing", prediction_type="sample")
pipe_lightning.to("cuda")

### Hyper SDXL ### 
repo_name = "ByteDance/Hyper-SD"
ckpt_name = "Hyper-SDXL-1step-Unet.safetensors"

unet = UNet2DConditionModel.from_config(base, subfolder="unet").to(torch.float16)
unet.load_state_dict(load_file(hf_hub_download(repo_name, ckpt_name)))
pipe_hyper = StableDiffusionXLPipeline.from_pretrained(base,
                                                       unet=unet,
                                                       vae=vae,
                                                       text_encoder=pipe_turbo.text_encoder,
                                                       text_encoder_2=pipe_turbo.text_encoder_2,
                                                       tokenizer=pipe_turbo.tokenizer,
                                                       tokenizer_2=pipe_turbo.tokenizer_2,
                                                       torch_dtype=torch.float16,
                                                       variant="fp16"
                                                      )#.to("cuda")
pipe_hyper.scheduler = LCMScheduler.from_config(pipe_hyper.scheduler.config)
pipe_hyper.to("cuda")
del unet

@spaces.GPU
def run_comparison(prompt: str,
    negative_prompt: str = "",
    use_negative_prompt: bool = False,
    num_inference_steps: int = 30,
    num_images_per_prompt: int = 2,
    seed: int = 0,
    width: int = 1024,
    height: int = 1024,
    guidance_scale: float = 3,
    randomize_seed: bool = False,
    progress=gr.Progress(track_tqdm=True),
):
    seed = int(randomize_seed_fn(seed, randomize_seed))
    if not use_negative_prompt:
        negative_prompt = ""
    image_turbo=pipe_turbo(prompt=prompt,
        negative_prompt=negative_prompt,
        width=width,
        height=height,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps,
        num_images_per_prompt=num_images_per_prompt,
        cross_attention_kwargs={"scale": 0.65},
        output_type="pil",
        ).images
    image_paths = [save_image(img) for img in images]
    return image_paths, seed
    yield image_turbo, None, None, None, None
    image_lightning=pipe_lightning(prompt=prompt,
        negative_prompt=negative_prompt,
        width=width,
        height=height,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps,
        num_images_per_prompt=num_images_per_prompt,
        cross_attention_kwargs={"scale": 0.65},
        output_type="pil",
        ).images
    image_paths = [save_image(img) for img in images]
    return image_paths, seed
    yield image_turbo, image_lightning, None, None, None
    image_hyper=pipe_hyper(prompt=prompt,
        negative_prompt=negative_prompt,
        width=width,
        height=height,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps,
        num_images_per_prompt=num_images_per_prompt,
        cross_attention_kwargs={"scale": 0.65},
        output_type="pil",
        ).images
    image_paths = [save_image(img) for img in images]
    return image_paths, seed
    yield image_turbo, image_lightning, image_hyper, None, None
    image_r3=RealVisXLv3_pipe(prompt=prompt,
        negative_prompt=negative_prompt,
        width=width,
        height=height,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps,
        num_images_per_prompt=num_images_per_prompt,
        cross_attention_kwargs={"scale": 0.65},
        output_type="pil",
        ).images
    image_paths = [save_image(img) for img in images]
    return image_paths, seed
    yield image_turbo, image_lightning, image_hyper,image_r3, None
    image_r4=RealVisXLv4_pipe(prompt=prompt,
        negative_prompt=negative_prompt,
        width=width,
        height=height,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps,
        num_images_per_prompt=num_images_per_prompt,
        cross_attention_kwargs={"scale": 0.65},
        output_type="pil",
        ).images
    image_paths = [save_image(img) for img in images]
    return image_paths, seed
    yield image_turbo, image_lightning, image_hyper,image_r3, image_r4

examples = ["A dignified beaver wearing glasses, a vest, and colorful neck tie.",
"The spirit of a tamagotchi wandering in the city of Barcelona",
"an ornate, high-backed mahogany chair with a red cushion",
"a sketch of a camel next to a stream",
"a delicate porcelain teacup sits on a saucer, its surface adorned with intricate blue patterns",
"a baby swan grafitti",
"A bald eagle made of chocolate powder, mango, and whipped cream"
]

with gr.Blocks() as demo:
    gr.Markdown("## One step SDXL comparison 🦶")
    gr.Markdown('Compare SDXL variants and distillations able to generate images in a single diffusion step')
    prompt = gr.Textbox(label="Prompt")
    run = gr.Button("Run")
    with gr.Accordion("Advanced options", open=False):
        use_negative_prompt = gr.Checkbox(label="Use negative prompt", value=True)
        negative_prompt = gr.Text(
            label="Negative prompt",
            lines=4,
            max_lines=6,
            value="""(deformed, distorted, disfigured:1.3), poorly drawn, bad anatomy, wrong anatomy, extra limb, missing limb, floating limbs, (mutated hands and fingers:1.4), disconnected limbs, mutation, mutated, ugly, disgusting, blurry, amputation, (NSFW:1.25)""",
            placeholder="Enter a negative prompt",
            visible=True,
        )
        with gr.Row():
            num_inference_steps = gr.Slider(
                label="Steps",
                minimum=10,
                maximum=60,
                step=1,
                value=30,
            )
        with gr.Row():
            num_images_per_prompt = gr.Slider(
                label="Images",
                minimum=1,
                maximum=5,
                step=1,
                value=2,
            )
        seed = gr.Slider(
            label="Seed",
            minimum=0,
            maximum=MAX_SEED,
            step=1,
            value=0,
            visible=True
        )
        randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
        with gr.Row(visible=True):
            width = gr.Slider(
                label="Width",
                minimum=512,
                maximum=2048,
                step=8,
                value=1024,
            )
            height = gr.Slider(
                label="Height",
                minimum=512,
                maximum=2048,
                step=8,
                value=1024,
            )
        with gr.Row():
            guidance_scale = gr.Slider(
                label="Guidance Scale",
                minimum=0.1,
                maximum=20.0,
                step=0.1,
                value=6,
            )

    with gr.Row():
        with gr.Column():
            image_turbo = gr.Gallery(label="SDXL Turbo",columns=1, preview=True,)
            gr.Markdown("## [SDXL Turbo](https://huggingface.co/stabilityai/sdxl-turbo)")
        with gr.Column():
            image_lightning = gr.Gallery(label="SDXL Lightning",columns=1, preview=True,)
            gr.Markdown("## [SDXL Lightning](https://huggingface.co/ByteDance/SDXL-Lightning)")
        with gr.Column():
            image_hyper = gr.Gallery(label="Hyper SDXL",columns=1, preview=True,)
            gr.Markdown("## [Hyper SDXL](https://huggingface.co/ByteDance/Hyper-SD)")
        with gr.Column():
            image_r3 = gr.Gallery(label="RealVisXL V3",columns=1, preview=True,)
            gr.Markdown("## [RealVisXL V3](https://huggingface.co)")
        with gr.Column():
            image_r4 = gr.Gallery(label="RealVisXL V4",columns=1, preview=True,)
            gr.Markdown("## [RealVisXL V3](https://huggingface.co)")
    image_outputs = [image_turbo, image_lightning, image_hyper, image_r3, image_r4]
    gr.on(
        triggers=[
            prompt.submit,
            negative_prompt.submit,
            run_button.click,
        ],
        fn=run_comparison,
        inputs=[
            prompt,
            negative_prompt,
            use_negative_prompt,
            num_inference_steps,
            num_images_per_prompt,
            seed,
            width,
            height,
            guidance_scale,
            randomize_seed,
        ],
        outputs=[image_outputs, seed],
        api_name="run",
    )
    use_negative_prompt.change(
        fn=lambda x: gr.update(visible=x),
        inputs=use_negative_prompt,
        outputs=negative_prompt,
        api_name=False,
    )
    gr.Examples(
        examples=examples,
        fn=run_comparison,
        inputs=prompt,
        outputs=[image_outputs, seed],
        cache_examples=False,
        run_on_click=True
    )
demo.launch()