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 import os import spaces import torch from diffusers import AutoencoderKLWan, WanPipeline, WanImageToVideoPipeline, UniPCMultistepScheduler from diffusers.utils import export_to_video import gradio as gr import tempfile from huggingface_hub import hf_hub_download import numpy as np from PIL import Image import random HF_MODEL = os.environ.get("HF_UPLOAD_REPO", "rahul7star/qwen-edit-img-repo") # --- CPU-only upload function --- def upload_image_and_prompt_cpu(input_image, prompt_text) -> str: from datetime import datetime import tempfile, os, uuid, shutil from huggingface_hub import HfApi # Instantiate the HfApi class api = HfApi() print(prompt_text) today_str = datetime.now().strftime("%Y-%m-%d") unique_subfolder = f"Upload-Image-{uuid.uuid4().hex[:8]}" hf_folder = f"{today_str}/{unique_subfolder}" # Save image temporarily with tempfile.NamedTemporaryFile(suffix=".png", delete=False) as tmp_img: if isinstance(input_image, str): shutil.copy(input_image, tmp_img.name) else: input_image.save(tmp_img.name, format="PNG") tmp_img_path = tmp_img.name # Upload image using HfApi instance api.upload_file( path_or_fileobj=tmp_img_path, path_in_repo=f"{hf_folder}/input_image.png", repo_id=HF_MODEL, repo_type="model", token=os.environ.get("HUGGINGFACE_HUB_TOKEN") ) # Save prompt as summary.txt summary_file = tempfile.NamedTemporaryFile(delete=False, suffix=".txt").name with open(summary_file, "w", encoding="utf-8") as f: f.write(prompt_text) api.upload_file( path_or_fileobj=summary_file, path_in_repo=f"{hf_folder}/summary.txt", repo_id=HF_MODEL, repo_type="model", token=os.environ.get("HUGGINGFACE_HUB_TOKEN") ) # Cleanup os.remove(tmp_img_path) os.remove(summary_file) return hf_folder # --- 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( "rahul7star/qwen-char-lora", weight_name="qwen_lora/Qwen-Image-Edit-2509-Lightning-4steps-V1.0-bf16_dim1.safetensors" ) pipe.fuse_lora(lora_scale=1.0) # pipe.load_lora_weights( # "rahul7star/qwen-char-lora", # weight_name="qwen_lora/qwen-multiple-angle.safetensors", # ) # pipe.fuse_lora(lora_scale=1.0) pipe.load_lora_weights( "rahul7star/qwen-char-lora", weight_name="qwen_lora/qwen-multiple-char.safetensors", ) pipe.fuse_lora(lora_scale=1.0) 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. The character stands with both arms extended slightly downward and close to the thighs, keeping the body evenly balanced on both sides. The legs are positioned symmetrically with a narrow stance. The 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." } # NEW: 出力解像度プリセット RESOLUTIONS = { "1:4": (512, 2048), "1:3": (576, 1728), "nealy 9:16": (768, 1344), "nealy 2:3": (832, 1216), "3:4": (896, 1152), } def _append_prompt(base: str, extra: str) -> str: extra = (extra or "").strip() return (base if not extra else f"{base} {extra}").strip() def generate_single_view(input_images, prompt, seed, num_inference_steps, true_guidance_scale): generator = torch.Generator(device=device).manual_seed(seed) print(prompt) 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 try: upload_image_and_prompt_cpu(result[0], prompt) except Exception as e: print("Upload failed:", e) return result[0] def concat_images_horizontally(images, bg_color=(255, 255, 255)): 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 # NEW: リサイズユーティリティ def resize_to_preset(img: Image.Image, preset_key: str) -> Image.Image: w, h = RESOLUTIONS[preset_key] return img.resize((w, h), Image.LANCZOS) @spaces.GPU() def generate_turnaround( image, extra_prompt="", preset_key="nealy 9:16", # NEW: デフォルト seed=42, randomize_seed=False, true_guidance_scale=1.0, num_inference_steps=4, progress=gr.Progress(track_tqdm=True), ): print(extra_prompt) try: upload_image_and_prompt_cpu(image, extra_prompt) except Exception as e: print("Upload failed:", e) 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] # 各プロンプト末尾に追記 p_front = _append_prompt(PROMPTS["front"], extra_prompt) p_back = _append_prompt(PROMPTS["back"], extra_prompt) p_left = _append_prompt(PROMPTS["left"], extra_prompt) p_right = _append_prompt(PROMPTS["right"], extra_prompt) progress(0.25, desc="正面生成中...") front = generate_single_view(pil_images, p_front, seed, num_inference_steps, true_guidance_scale) progress(0.5, desc="背面生成中...") back = generate_single_view([front], p_back, seed+1, num_inference_steps, true_guidance_scale) progress(0.75, desc="左側面生成中...") left = generate_single_view([front], p_left, seed+2, num_inference_steps, true_guidance_scale) progress(1.0, desc="右側面生成中...") right = generate_single_view([front], p_right, seed+3, num_inference_steps, true_guidance_scale) # NEW: ここで指定プリセットにリサイズ front_r = resize_to_preset(front, preset_key) back_r = resize_to_preset(back, preset_key) left_r = resize_to_preset(left, preset_key) right_r = resize_to_preset(right, preset_key) # NEW: リサイズ後を連結（横：正面→右→背面→左） concat = concat_images_horizontally([front_r, right_r, back_r, left_r]) return front_r, back_r, left_r, right_r, concat, seed, f"✅ {preset_key} にリサイズして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%;} /* 追加: 注意ボックスのスタイル */ .notice { background: #fff5f5; border: 1px solid #fca5a5; color: #7f1d1d; padding: 12px 14px; border-radius: 10px; font-weight: 600; line-height: 1.5; margin-bottom: 10px; } """ with gr.Blocks(css=css) as demo: with gr.Column(elem_id="col-container"): input_image = gr.Image(label="入力画像", type="pil", height=500) # 追記プロンプト欄 extra_prompt = gr.Textbox( label="追加プロンプト（各視点プロンプトの末尾に追記）", placeholder="例: high detail, anime style, soft lighting, 4k, pastel colors", lines=2 ) # NEW: 出力解像度プリセットのプルダウン preset_dropdown = gr.Dropdown( label="出力解像度プリセット", choices=list(RESOLUTIONS.keys()), value="nealy 9:16" ) 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) # NEW: クリック時に preset_dropdown を引数として渡す run_button.click( fn=generate_turnaround, inputs=[input_image, extra_prompt, preset_dropdown, 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()