import logging
from PIL import Image, PngImagePlugin
from datetime import datetime
from diffusers.models import AutoencoderKL
from diffusers import StableDiffusionXLPipeline, StableDiffusionXLImg2ImgPipeline
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
DESCRIPTION = "RealVis XL"
if not torch.cuda.is_available():
DESCRIPTION += "\nRunning on CPU 🥶 This demo does not work on CPU.
"
IS_COLAB = utils.is_google_colab() or os.getenv("IS_COLAB") == "1"
ENABLE_CPU_OFFLOAD = os.getenv("ENABLE_CPU_OFFLOAD") == "1"
OUTPUT_DIR = os.getenv("OUTPUT_DIR", "./outputs")
MODEL = os.getenv(
"MODEL",
"https://huggingface.co/SG161222/RealVisXL_V4.0/blob/main/RealVisXL_V4.0.safetensors",
)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
def load_pipeline(model_name):
vae = AutoencoderKL.from_pretrained(
"madebyollin/sdxl-vae-fp16-fix",
torch_dtype=torch.float16,
)
pipeline = (
StableDiffusionXLPipeline.from_single_file
if MODEL.endswith(".safetensors")
else StableDiffusionXLPipeline.from_pretrained
)
pipe = pipeline(
model_name,
vae=vae,
torch_dtype=torch.float16,
custom_pipeline="lpw_stable_diffusion_xl",
use_safetensors=True,
add_watermarker=False,
use_auth_token=HF_TOKEN,
variant="fp16",
)
pipe.to(device)
return pipe
@spaces.GPU
def generate(
custom_height: int = 1024,
guidance_scale: float = 7.0,
num_inference_steps: int = 30,
sampler: str = "DPM++ 2M SDE Karras",
aspect_ratio_selector: str = "1024 x 1024",
use_upscaler: bool = False,
upscaler_strength: float = 0.55,
upscale_by: float = 1.5,
progress=gr.Progress(track_tqdm=True),
) -> list:
generator = utils.seed_everything(seed)
width, height = utils.aspect_ratio_handler(
aspect_ratio_selector,
custom_width,
custom_height,
)
width, height = utils.preprocess_image_dimensions(width, height)
backup_scheduler = pipe.scheduler
pipe.scheduler = utils.get_scheduler(pipe.scheduler.config, sampler)
metadata = {
"prompt": prompt,
"negative_prompt": negative_prompt,
"resolution": f"{width} x {height}",
"guidance_scale": guidance_scale,
"num_inference_steps": num_inference_steps,
"seed": seed,
"sampler": sampler,
"use_upscaler": use_upscaler,
"upscaler_strength": upscaler_strength,
"upscale_by": upscale_by,
}
logger.info(json.dumps(metadata, indent=4))
try:
images = pipe(
prompt=prompt,
negative_prompt=negative_prompt,
width=width,
height=height,
guidance_scale=guidance_scale,
num_inference_steps=num_inference_steps,
generator=generator,
output_type="pil",
).images
if use_upscaler:
images = [image.resize((int(width * upscale_by), int(height * upscale_by))) for image in images]
return images, metadata
except Exception as e:
logger.exception(f"An error occurred: {e}")
raise
finally:
pipe.scheduler = backup_scheduler
utils.free_memory()
if torch.cuda.is_available():
pipe = load_pipeline(MODEL)
logger.info("Loaded on Device!")
else:
pipe = None
def postprocess_images(images):
return images # No caption, just return the images
with gr.Blocks(css="style.css") as demo:
title = gr.HTML(
f"""{DESCRIPTION}
""",
elem_id="title",
)
gr.Markdown(
f"""Gradio demo for ([RealVis XL]https://huggingface.co/SG161222/RealVisXL_V4.0/)""",
elem_id="subtitle",
)
gr.DuplicateButton(
placeholder="Enter a negative prompt",
value=""
)
aspect_ratio_selector = gr.Radio(
label="Aspect Ratio",
choices=config.aspect_ratios,
value="1024 x 1024",
container=True,
)
with gr.Group(visible=False) as custom_resolution:
with gr.Row():
custom_width = gr.Slider(
label="Width",
minimum=MIN_IMAGE_SIZE,
maximum=MAX_IMAGE_SIZE,
step=8,
value=1024,
)
custom_height = gr.Slider(
label="Height",
minimum=MIN_IMAGE_SIZE,
maximum=MAX_IMAGE_SIZE,
step=8,
value=1024,
)
use_upscaler = gr.Checkbox(label="Use Upscaler", value=False)
with gr.Row() as upscaler_row:
upscaler_strength = gr.Slider(
label="Strength",
minimum=0,
maximum=1,
step=0.05,
value=0.55,
visible=False,
)
upscale_by = gr.Slider(
label="Upscale by",
minimum=1,
maximum=1.5,
step=0.1,
value=1.5,
visible=False,
)
sampler = gr.Dropdown(
label="Sampler",
choices=config.sampler_list,
interactive=True,
value="DPM++ 2M SDE Karras",
)
with gr.Row():
seed = gr.Slider(
label="Seed", minimum=0, maximum=utils.MAX_SEED, step=1, value=0
)
randomize_seed = gr.Checkbox(label="Randomize seed", value=True)
with gr.Group():
with gr.Row():
guidance_scale = gr.Slider(
label="Guidance scale",
minimum=1,
maximum=12,
step=0.1,
value=7.0,
)
num_inference_steps = gr.Slider(
label="Number of inference steps",
minimum=1,
maximum=100,
step=1,
value=28,
)
with gr.Accordion(label="Generation Parameters", open=False):
gr_metadata = gr.JSON(label="Metadata", show_label=False)
gr.Examples(
examples=config.examples,
inputs=prompt,
outputs=[result, gr_metadata],
fn=lambda *args, **kwargs: generate(*args, use_upscaler=True, **kwargs),
cache_examples=CACHE_EXAMPLES,
)
use_upscaler.change(
fn=lambda x: [gr.update(visible=x), gr.update(visible=x)],
inputs=use_upscaler,
outputs=[upscaler_strength, upscale_by],
queue=False,
api_name=False,
)
aspect_ratio_selector.change(
fn=lambda x: gr.update(visible=x == "Custom"),
inputs=aspect_ratio_selector,
outputs=custom_resolution,
queue=False,
api_name=False,
)
inputs = [
prompt,
negative_prompt,
custom_height,
guidance_scale,
num_inference_steps,
sampler,
aspect_ratio_selector,
use_upscaler,
upscaler_strength,
upscale_by,
]
prompt.submit(
fn=utils.randomize_seed_fn,
inputs=[seed, randomize_seed],
outputs=seed,
queue=False,
api_name=False,
).then(
fn=generate,
inputs=inputs,
outputs=[result, gr_metadata],
api_name="run",
)
negative_prompt.submit(
fn=utils.randomize_seed_fn,
inputs=[seed, randomize_seed],
outputs=seed,
queue=False,
api_name=False,
).then(
fn=generate,
inputs=inputs,
outputs=[result, gr_metadata],
api_name=False,
)
run_button.click(
fn=utils.randomize_seed_fn,
inputs=[seed, randomize_seed],
outputs=seed,
queue=False,
api_name=False,
).then(
fn=generate,
inputs=inputs,
outputs=[result, gr_metadata],
api_name=False,
)
demo.queue(max_size=20).launch(debug=IS_COLAB, share=IS_COLAB, show_error=True)