src/inference.py

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

from src.const import DIFFUSERS_MODEL_IDS, EXTERNAL_MODEL_MAPPING, DEVICE


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)

    # Seed 固定
    generator = torch.Generator(device=DEVICE).manual_seed(seed)

    if isinstance(pipe, StableDiffusionPipeline):
        # 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]
        )

        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
    else:
        progress(0.3, 'Generating images...')
        images = 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,
            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