import gradio as gr import numpy as np import random import torch import spaces from PIL import Image from diffusers import FlowMatchEulerDiscreteScheduler from optimization import optimize_pipeline_ from qwenimage.pipeline_qwenimage_edit_plus import QwenImageEditPlusPipeline from qwenimage.transformer_qwenimage import QwenImageTransformer2DModel from qwenimage.qwen_fa3_processor import QwenDoubleStreamAttnProcessorFA3 import math import os # --- Model Loading --- dtype = torch.bfloat16 device = "cuda" if torch.cuda.is_available() else "cpu" # Scheduler configuration for Lightning scheduler_config = { "base_image_seq_len": 256, "base_shift": math.log(3), "invert_sigmas": False, "max_image_seq_len": 8192, "max_shift": math.log(3), "num_train_timesteps": 1000, "shift": 1.0, "shift_terminal": None, "stochastic_sampling": False, "time_shift_type": "exponential", "use_beta_sigmas": False, "use_dynamic_shifting": True, "use_exponential_sigmas": False, "use_karras_sigmas": False, } # Initialize scheduler scheduler = FlowMatchEulerDiscreteScheduler.from_config(scheduler_config) # Load model pipe = QwenImageEditPlusPipeline.from_pretrained( "Qwen/Qwen-Image-Edit-2509", scheduler=scheduler, torch_dtype=dtype ).to(device) pipe.load_lora_weights( "2vXpSwA7/iroiro-lora", weight_name="qwen_lora/Qwen-Image-Edit-2509-Lightning-4steps-V1.0-bf16_dim1.safetensors" ) pipe.fuse_lora(lora_scale=0.8) pipe.transformer.__class__ = QwenImageTransformer2DModel pipe.transformer.set_attn_processor(QwenDoubleStreamAttnProcessorFA3()) optimize_pipeline_(pipe, image=[Image.new("RGB", (1024, 1024)), Image.new("RGB", (1024, 1024))], prompt="prompt") # --- Constants --- MAX_SEED = np.iinfo(np.int32).max PROMPTS = { "front": "Move the camera to a front-facing position so the full body of the character is visible. Background is plain white.", "back": "Move the camera to a back-facing position so the full body of the character is visible. Background is plain white.", "left": "Move the camera to a side view (profile) from the left so the full body of the character is visible. Background is plain white.", "right": "Move the camera to a side view (profile) from the right so the full body of the character is visible. Background is plain white." } def generate_single_view(input_images, prompt, seed, num_inference_steps, true_guidance_scale): """単一視点画像生成""" generator = torch.Generator(device=device).manual_seed(seed) result = pipe( image=input_images if input_images else None, prompt=prompt, negative_prompt=" ", num_inference_steps=num_inference_steps, generator=generator, true_cfg_scale=true_guidance_scale, num_images_per_prompt=1, ).images return result[0] def concat_images_horizontally(images, bg_color=(255, 255, 255)): """画像を横に連結""" images = [img.convert("RGB") for img in images if img is not None] if not images: return None h = max(img.height for img in images) resized = [] for img in images: if img.height != h: w = int(img.width * (h / img.height)) img = img.resize((w, h), Image.LANCZOS) resized.append(img) w_total = sum(img.width for img in resized) canvas = Image.new("RGB", (w_total, h), bg_color) x = 0 for img in resized: canvas.paste(img, (x, 0)) x += img.width return canvas @spaces.GPU(duration=300) def generate_turnaround( image, seed=42, randomize_seed=False, true_guidance_scale=1.0, num_inference_steps=4, progress=gr.Progress(track_tqdm=True), ): """4視点＋横連結PNG生成""" if randomize_seed: seed = random.randint(0, MAX_SEED) if image is None: return None, None, None, None, None, seed, "エラー: 入力画像をアップロードしてください" if isinstance(image, Image.Image): input_image = image.convert("RGB") else: input_image = Image.open(image).convert("RGB") pil_images = [input_image] progress(0.25, desc="正面生成中...") front = generate_single_view(pil_images, PROMPTS["front"], seed, num_inference_steps, true_guidance_scale) progress(0.5, desc="背面生成中...") back = generate_single_view([front], PROMPTS["back"], seed+1, num_inference_steps, true_guidance_scale) progress(0.75, desc="左側面生成中...") left = generate_single_view([front], PROMPTS["left"], seed+2, num_inference_steps, true_guidance_scale) progress(1.0, desc="右側面生成中...") right = generate_single_view([front], PROMPTS["right"], seed+3, num_inference_steps, true_guidance_scale) concat = concat_images_horizontally([front, right, back, left]) return front, back, left, right, concat, seed, "✅ PNG形式で4視点＋連結画像を生成しました" # --- UI --- css = """ #col-container {margin: 0 auto; max-width: 1400px;} .image-container img {object-fit: contain !important; max-width: 100%; max-height: 100%;} """ with gr.Blocks(css=css) as demo: gr.Markdown("# キャラクター4視点立ち絵自動生成（PNG出力対応）") gr.Markdown("アップロードしたキャラクター画像から正面・背面・左右側面、さらに4枚連結のPNG画像を出力します。") with gr.Column(elem_id="col-container"): input_image = gr.Image(label="入力画像", type="pil", height=500) run_button = gr.Button("🎨 生成開始", variant="primary") status_text = gr.Textbox(label="ステータス", interactive=False) with gr.Row(): result_front = gr.Image(label="正面", type="pil", format="png", height=400, show_download_button=True) result_back = gr.Image(label="背面", type="pil", format="png", height=400, show_download_button=True) with gr.Row(): result_left = gr.Image(label="左側面", type="pil", format="png", height=400, show_download_button=True) result_right = gr.Image(label="右側面", type="pil", format="png", height=400, show_download_button=True) # PNG連結出力 result_concat = gr.Image(label="連結画像（正面→右→背面→左）", type="pil", format="png", height=400, show_download_button=True) with gr.Accordion("⚙️ 詳細設定", open=False): seed = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=0) randomize_seed = gr.Checkbox(label="ランダムシード", value=True) true_guidance_scale = gr.Slider(label="True guidance scale", minimum=1.0, maximum=10.0, step=0.1, value=1.0) num_inference_steps = gr.Slider(label="生成ステップ数", minimum=1, maximum=40, step=1, value=4) run_button.click( fn=generate_turnaround, inputs=[input_image, seed, randomize_seed, true_guidance_scale, num_inference_steps], outputs=[result_front, result_back, result_left, result_right, result_concat, seed, status_text], ) if __name__ == "__main__": demo.launch()