app.py

import gradio as gr
import spaces
import torch
from PIL import Image
from compel import Compel, DiffusersTextualInversionManager
from diffusers import DiffusionPipeline
from diffusers.utils import make_image_grid

from src.example import EXAMPLES

DIFFUSERS_MODEL_IDS = [
    # SD Models
    "stabilityai/stable-diffusion-3-medium-diffusers",
    "stabilityai/stable-diffusion-xl-base-1.0",
    "stabilityai/stable-diffusion-2-1",
    "runwayml/stable-diffusion-v1-5",

    # Other Models
    "Prgckwb/trpfrog-diffusion",
]
EXTERNAL_MODEL_MAPPING = {
    "Beautiful Realistic Asians": "checkpoints/diffusers/Beautiful Realistic Asians v7",
}
MODEL_CHOICES = DIFFUSERS_MODEL_IDS + list(EXTERNAL_MODEL_MAPPING.keys())

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


def load_pipeline(model_id, use_model_offload, safety_checker):
    # Diffusers リポジトリ内のモデル
    if model_id in DIFFUSERS_MODEL_IDS:
        pipe = DiffusionPipeline.from_pretrained(
            model_id,
            torch_dtype=torch.float16,
        )

    # CIVITAI 系列由来のモデル
    else:
        pipe = DiffusionPipeline.from_pretrained(
            EXTERNAL_MODEL_MAPPING[model_id],
            torch_dtype=torch.float16,
        )

        # Load Textual Inversion
        pipe.load_textual_inversion("checkpoints/embeddings/BadNegAnatomyV1 neg.pt", token='BadNegAnatomyV1-neg')
        pipe.load_textual_inversion("checkpoints/embeddings/Deep Negative V1 75T.pt", token='DeepNegative')
        pipe.load_textual_inversion("checkpoints/embeddings/easynegative.safetensors", token='EasyNegative')
        pipe.load_textual_inversion("checkpoints/embeddings/Negative Hand Embedding.pt", token='negative_hand-neg')

        # Load LoRA
        pipe.load_lora_weights("checkpoints/lora/detailed style SD1.5.safetensors", adapter_name='detail')
        pipe.load_lora_weights("checkpoints/lora/perfection style SD1.5.safetensors", adapter_name='perfection')
        pipe.load_lora_weights("checkpoints/lora/Hand v3 SD1.5.safetensors", adapter_name='hands')
        pipe.set_adapters(['detail', 'hands'], adapter_weights=[0.5, 0.5])

    # VRAM が少ないとき用の対策
    if use_model_offload:
        pipe.enable_model_cpu_offload()
    else:
        pipe = pipe.to(device)

    if not safety_checker:
        pipe.safety_checker = None

    return pipe


@spaces.GPU(duration=120)
@torch.inference_mode()
def inference(
        prompt: str,
        model_id: str = "stabilityai/stable-diffusion-3-medium-diffusers",
        negative_prompt: str = "",
        width: int = 512,
        height: int = 512,
        guidance_scale: float = 7.5,
        num_inference_steps: int = 50,
        num_images: int = 4,
        safety_checker: bool = True,
        use_model_offload: bool = False,
        seed: int = 8888,
        progress=gr.Progress(track_tqdm=True),
) -> Image.Image:
    progress(0, 'Loading pipeline...')
    pipe = load_pipeline(model_id, use_model_offload, safety_checker)

    # For Compel
    textual_inversion_manager = DiffusersTextualInversionManager(pipe)
    compel_procs = Compel(
        tokenizer=pipe.tokenizer,
        text_encoder=pipe.text_encoder,
        textual_inversion_manager=textual_inversion_manager,
        truncate_long_prompts=False,
    )
    prompt_embed = compel_procs(prompt)
    negative_prompt_embed = compel_procs(negative_prompt)

    prompt_embed, negative_prompt_embed = compel_procs.pad_conditioning_tensors_to_same_length(
        [prompt_embed, negative_prompt_embed]
    )

    generator = torch.Generator(device=device).manual_seed(seed)

    progress(0.3, 'Generating images...')
    images = pipe(
        prompt_embeds=prompt_embed,
        negative_prompt_embeds=negative_prompt_embed,
        width=width,
        height=height,
        guidance_scale=guidance_scale,
        num_inference_steps=num_inference_steps,
        num_images_per_prompt=num_images,
        generator=generator,
    ).images

    progress(0.9, f'Done generating {num_images} images')
    if num_images % 2 == 1:
        image = make_image_grid(images, rows=num_images, cols=1)
    else:
        image = make_image_grid(images, rows=2, cols=num_images // 2)

    return image


def build_interface():
    """Build Gradio Interface"""

    theme = gr.themes.Default(primary_hue=gr.themes.colors.emerald)

    with gr.Blocks(theme=theme) as interface:
        gr.Markdown(f"# Stable Diffusion Demo")

        with gr.Row():
            with gr.Column():
                prompt = gr.Text(label="Prompt", placeholder="Enter a prompt here")

                model_id = gr.Dropdown(
                    label="Model ID",
                    choices=MODEL_CHOICES,
                    value="stabilityai/stable-diffusion-3-medium-diffusers",
                )

                # Additional Input Settings
                with gr.Accordion("Additional Settings", open=False):
                    negative_prompt = gr.Text(label="Negative Prompt", value="", )

                    with gr.Row():
                        width = gr.Number(label="Width", value=512, step=64, minimum=64, maximum=2048)
                        height = gr.Number(label="Height", value=512, step=64, minimum=64, maximum=2048)
                        num_images = gr.Number(label="Num Images", value=4, minimum=1, maximum=10, step=1)
                        seed = gr.Number(label="Seed", value=8888, step=1)

                    guidance_scale = gr.Slider(label="Guidance Scale", value=7.5, step=0.5, minimum=0, maximum=10)
                    num_inference_step = gr.Slider(
                        label="Num Inference Steps", value=50, minimum=1, maximum=100, step=2
                    )

                    with gr.Row():
                        use_safety_checker = gr.Checkbox(value=True, label='Use Safety Checker')
                        use_model_offload = gr.Checkbox(value=False, label='Use Model Offload')

                with gr.Accordion(label='Notes', open=False):
                    # language=HTML
                    notes = gr.HTML(
                        """
                            <h2>Negative Embeddings</h2>
                            <p>If you want to use negative embedding, use the following tokens in the prompt.</p>
                            <ul>
                                <li><a href='https://civitai.com/models/59614/badneganatomy-textual-inversion'>BadNegAnatomyV1-neg</a></li>
                                <li><a href='https://civitai.com/models/4629/deep-negative-v1x'>DeepNegativeV1</a> </li>
                                <li><a href='https://civitai.com/models/7808/easynegative'>EasyNegative</a></li>
                                <li><a href='https://civitai.com/models/56519/negativehand-negative-embedding'>negative_hand-neg</a></li>
                            </ul>
                        """
                    )

            with gr.Column():
                output_image = gr.Image(label="Image", type="pil")

        inputs = [
            prompt,
            model_id,
            negative_prompt,
            width,
            height,
            guidance_scale,
            num_inference_step,
            num_images,
            use_safety_checker,
            use_model_offload,
            seed,
        ]

        btn = gr.Button("Generate", variant='primary')
        btn.click(
            fn=inference,
            inputs=inputs,
            outputs=output_image
        )

        gr.Examples(
            examples=EXAMPLES,
            inputs=inputs,
            outputs=output_image,
            fn=inference,
            cache_examples='lazy'
        )
    return interface


if __name__ == "__main__":
    iface = build_interface()
    iface.queue().launch()