Spaces:
Runtime error
Runtime error
| import spaces | |
| import torch | |
| from diffusers.pipelines.wan.pipeline_wan_i2v import WanImageToVideoPipeline | |
| from diffusers.models.transformers.transformer_wan import WanTransformer3DModel | |
| from diffusers.utils.export_utils import export_to_video | |
| import gradio as gr | |
| import tempfile | |
| import numpy as np | |
| from PIL import Image | |
| import random | |
| import gc | |
| from torchao.quantization import quantize_ | |
| from torchao.quantization import Float8DynamicActivationFloat8WeightConfig | |
| from torchao.quantization import Int8WeightOnlyConfig | |
| import aoti | |
| MODEL_ID = "Wan-AI/Wan2.2-I2V-A14B-Diffusers" | |
| MAX_DIM = 832 | |
| MIN_DIM = 480 | |
| SQUARE_DIM = 640 | |
| MULTIPLE_OF = 16 | |
| MAX_SEED = np.iinfo(np.int32).max | |
| FIXED_FPS = 16 | |
| MIN_FRAMES_MODEL = 8 | |
| MAX_FRAMES_MODEL = 80 | |
| MIN_DURATION = round(MIN_FRAMES_MODEL/FIXED_FPS,1) | |
| MAX_DURATION = round(MAX_FRAMES_MODEL/FIXED_FPS,1) | |
| pipe = WanImageToVideoPipeline.from_pretrained(MODEL_ID, | |
| transformer=WanTransformer3DModel.from_pretrained('cbensimon/Wan2.2-I2V-A14B-bf16-Diffusers', | |
| subfolder='transformer', | |
| torch_dtype=torch.bfloat16, | |
| device_map='cuda', | |
| ), | |
| transformer_2=WanTransformer3DModel.from_pretrained('cbensimon/Wan2.2-I2V-A14B-bf16-Diffusers', | |
| subfolder='transformer_2', | |
| torch_dtype=torch.bfloat16, | |
| device_map='cuda', | |
| ), | |
| torch_dtype=torch.bfloat16, | |
| ).to('cuda') | |
| pipe.load_lora_weights( | |
| "Kijai/WanVideo_comfy", | |
| weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors", | |
| adapter_name="lightx2v" | |
| ) | |
| kwargs_lora = {} | |
| kwargs_lora["load_into_transformer_2"] = True | |
| pipe.load_lora_weights( | |
| "Kijai/WanVideo_comfy", | |
| weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors", | |
| adapter_name="lightx2v_2", **kwargs_lora | |
| ) | |
| pipe.set_adapters(["lightx2v", "lightx2v_2"], adapter_weights=[1., 1.]) | |
| pipe.fuse_lora(adapter_names=["lightx2v"], lora_scale=3., components=["transformer"]) | |
| pipe.fuse_lora(adapter_names=["lightx2v_2"], lora_scale=1., components=["transformer_2"]) | |
| pipe.unload_lora_weights() | |
| quantize_(pipe.text_encoder, Int8WeightOnlyConfig()) | |
| quantize_(pipe.transformer, Float8DynamicActivationFloat8WeightConfig()) | |
| quantize_(pipe.transformer_2, Float8DynamicActivationFloat8WeightConfig()) | |
| aoti.aoti_blocks_load(pipe.transformer, 'zerogpu-aoti/Wan2', variant='fp8da') | |
| aoti.aoti_blocks_load(pipe.transformer_2, 'zerogpu-aoti/Wan2', variant='fp8da') | |
| default_prompt_i2v = "์ด ์ด๋ฏธ์ง์ ์๋๊ฐ์ ๋ถ์ฌํ๊ณ , ์ํ ๊ฐ์ ์์ง์๊ณผ ๋ถ๋๋ฌ์ด ์ ๋๋ฉ์ด์ ์ ์ ์ฉ" | |
| default_negative_prompt = "์์กฐ ์ ๋ช , ๊ณผ๋ค ๋ ธ์ถ, ์ ์ , ์ธ๋ถ ํ๋ฆผ, ์๋ง, ์คํ์ผ, ์ํ, ๊ทธ๋ฆผ, ํ๋ฉด, ์ ์ง, ํ์์กฐ, ์ต์ ํ์ง, ์ ํ์ง, JPEG ์์ถ, ์ถํจ, ๋ถ์์ , ์ถ๊ฐ ์๊ฐ๋ฝ, ์๋ชป ๊ทธ๋ ค์ง ์, ์๋ชป ๊ทธ๋ ค์ง ์ผ๊ตด, ๊ธฐํ, ๋ณํ, ํํ ๋ถ๋ ์ฌ์ง, ์๊ฐ๋ฝ ์ตํฉ, ์ ์ง ํ๋ฉด, ์ง์ ๋ถํ ๋ฐฐ๊ฒฝ, ์ธ ๊ฐ์ ๋ค๋ฆฌ, ๋ฐฐ๊ฒฝ ์ฌ๋ ๋ง์, ๋ค๋ก ๊ฑท๊ธฐ" | |
| def resize_image(image: Image.Image) -> Image.Image: | |
| width, height = image.size | |
| if width == height: | |
| return image.resize((SQUARE_DIM, SQUARE_DIM), Image.LANCZOS) | |
| aspect_ratio = width / height | |
| MAX_ASPECT_RATIO = MAX_DIM / MIN_DIM | |
| MIN_ASPECT_RATIO = MIN_DIM / MAX_DIM | |
| image_to_resize = image | |
| if aspect_ratio > MAX_ASPECT_RATIO: | |
| target_w, target_h = MAX_DIM, MIN_DIM | |
| crop_width = int(round(height * MAX_ASPECT_RATIO)) | |
| left = (width - crop_width) // 2 | |
| image_to_resize = image.crop((left, 0, left + crop_width, height)) | |
| elif aspect_ratio < MIN_ASPECT_RATIO: | |
| target_w, target_h = MIN_DIM, MAX_DIM | |
| crop_height = int(round(width / MIN_ASPECT_RATIO)) | |
| top = (height - crop_height) // 2 | |
| image_to_resize = image.crop((0, top, width, top + crop_height)) | |
| else: | |
| if width > height: | |
| target_w = MAX_DIM | |
| target_h = int(round(target_w / aspect_ratio)) | |
| else: | |
| target_h = MAX_DIM | |
| target_w = int(round(target_h * aspect_ratio)) | |
| final_w = round(target_w / MULTIPLE_OF) * MULTIPLE_OF | |
| final_h = round(target_h / MULTIPLE_OF) * MULTIPLE_OF | |
| final_w = max(MIN_DIM, min(MAX_DIM, final_w)) | |
| final_h = max(MIN_DIM, min(MAX_DIM, final_h)) | |
| return image_to_resize.resize((final_w, final_h), Image.LANCZOS) | |
| def get_num_frames(duration_seconds: float): | |
| return 1 + int(np.clip( | |
| int(round(duration_seconds * FIXED_FPS)), | |
| MIN_FRAMES_MODEL, | |
| MAX_FRAMES_MODEL, | |
| )) | |
| def get_duration( | |
| input_image, | |
| prompt, | |
| steps, | |
| negative_prompt, | |
| duration_seconds, | |
| guidance_scale, | |
| guidance_scale_2, | |
| seed, | |
| randomize_seed, | |
| progress, | |
| ): | |
| BASE_FRAMES_HEIGHT_WIDTH = 81 * 832 * 624 | |
| BASE_STEP_DURATION = 15 | |
| width, height = resize_image(input_image).size | |
| frames = get_num_frames(duration_seconds) | |
| factor = frames * width * height / BASE_FRAMES_HEIGHT_WIDTH | |
| step_duration = BASE_STEP_DURATION * factor ** 1.5 | |
| return 10 + int(steps) * step_duration | |
| def generate_video( | |
| input_image, | |
| prompt, | |
| steps = 4, | |
| negative_prompt=default_negative_prompt, | |
| duration_seconds = MAX_DURATION, | |
| guidance_scale = 1, | |
| guidance_scale_2 = 1, | |
| seed = 42, | |
| randomize_seed = False, | |
| progress=gr.Progress(track_tqdm=True), | |
| ): | |
| if input_image is None: | |
| raise gr.Error("์ด๋ฏธ์ง๋ฅผ ์ ๋ก๋ํด์ฃผ์ธ์.") | |
| num_frames = get_num_frames(duration_seconds) | |
| current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed) | |
| resized_image = resize_image(input_image) | |
| output_frames_list = pipe( | |
| image=resized_image, | |
| prompt=prompt, | |
| negative_prompt=negative_prompt, | |
| height=resized_image.height, | |
| width=resized_image.width, | |
| num_frames=num_frames, | |
| guidance_scale=float(guidance_scale), | |
| guidance_scale_2=float(guidance_scale_2), | |
| num_inference_steps=int(steps), | |
| generator=torch.Generator(device="cuda").manual_seed(current_seed), | |
| ).frames[0] | |
| with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmpfile: | |
| video_path = tmpfile.name | |
| export_to_video(output_frames_list, video_path, fps=FIXED_FPS) | |
| return video_path, current_seed | |
| # ์ธ๋ จ๋ ํ๊ธ UI | |
| with gr.Blocks(theme=gr.themes.Soft()) as demo: | |
| gr.Markdown("# ๐ฌ WAN_VO: WAN ๊ธฐ๋ฐ ์ด๊ณ ์ ์ด๋ฏธ์ง to ๋น๋์ค ๋ฌด๋ฃ ์์ฑ ์คํ์์ค") | |
| gr.Markdown("** WAN 2.2 14B + FAST + ํ๊ธํ + ํ๋ ** - 4~8๋จ๊ณ๋ก ๋น ๋ฅธ ์์ ์์ฑ") | |
| with gr.Row(): | |
| with gr.Column(scale=1): | |
| input_image_component = gr.Image(type="pil", label="์ ๋ ฅ ์ด๋ฏธ์ง") | |
| prompt_input = gr.Textbox(label="ํ๋กฌํํธ", value=default_prompt_i2v, lines=2) | |
| duration_seconds_input = gr.Slider( | |
| minimum=MIN_DURATION, | |
| maximum=MAX_DURATION, | |
| step=0.1, | |
| value=3.5, | |
| label="์์ ๊ธธ์ด (์ด)" | |
| ) | |
| with gr.Accordion("๊ณ ๊ธ ์ค์ ", open=False): | |
| negative_prompt_input = gr.Textbox(label="๋ค๊ฑฐํฐ๋ธ ํ๋กฌํํธ", value=default_negative_prompt, lines=2) | |
| steps_slider = gr.Slider(minimum=1, maximum=30, step=1, value=6, label="์์ฑ ๋จ๊ณ") | |
| guidance_scale_input = gr.Slider(minimum=0.0, maximum=10.0, step=0.5, value=1, label="๊ฐ์ด๋์ค ์ค์ผ์ผ 1") | |
| guidance_scale_2_input = gr.Slider(minimum=0.0, maximum=10.0, step=0.5, value=1, label="๊ฐ์ด๋์ค ์ค์ผ์ผ 2") | |
| seed_input = gr.Slider(label="์๋", minimum=0, maximum=MAX_SEED, step=1, value=42) | |
| randomize_seed_checkbox = gr.Checkbox(label="๋๋ค ์๋ ์ฌ์ฉ", value=True) | |
| generate_button = gr.Button("๐ฅ ์์ ์์ฑ", variant="primary", size="lg") | |
| with gr.Column(scale=1): | |
| video_output = gr.Video(label="์์ฑ๋ ์์", autoplay=True, interactive=False) | |
| ui_inputs = [ | |
| input_image_component, prompt_input, steps_slider, | |
| negative_prompt_input, duration_seconds_input, | |
| guidance_scale_input, guidance_scale_2_input, seed_input, randomize_seed_checkbox | |
| ] | |
| generate_button.click(fn=generate_video, inputs=ui_inputs, outputs=[video_output, seed_input]) | |
| gr.Examples( | |
| examples=[ | |
| [ | |
| "wan_i2v_input.JPG", | |
| "POV ์ ์นด ์์, ์ ๊ธ๋ผ์ค ๋ ํฐ ๊ณ ์์ด๊ฐ ์ํ๋ณด๋์ ์์ ํธ์ํ ๋ฏธ์. ๋ฐฐ๊ฒฝ์ ์ด๋ ํด๋ณ(๋ง์ ๋ฌผ, ๋ น์ ์ธ๋, ๊ตฌ๋ฆ ๋ ํธ๋ฅธ ํ๋). ์ํ๋ณด๋๊ฐ ๊ธฐ์ธ์ด์ง๊ณ ๊ณ ์์ด๊ฐ ๋ฐ๋ค๋ก ๋จ์ด์ง๋ฉฐ ์นด๋ฉ๋ผ๊ฐ ๊ฑฐํ๊ณผ ํ๋น๊ณผ ํจ๊ป ๋ฌผ์์ผ๋ก ๋น ์ง. ์ ๊น ๋ฌผ์์์ ๊ณ ์์ด ์ผ๊ตด ๋ณด์ด๋ค๊ฐ ๋ค์ ์๋ฉด ์๋ก ์ฌ๋ผ์ ์ ์นด ์ดฌ์ ๊ณ์, ์ฆ๊ฑฐ์ด ์ฌ๋ฆ ํด๊ฐ ๋ถ์๊ธฐ.", | |
| 4, | |
| ], | |
| [ | |
| "wan22_input_2.jpg", | |
| "์ธ๋ จ๋ ๋ฌ ํ์ฌ ์ฐจ๋์ด ์ผ์ชฝ์์ ์ค๋ฅธ์ชฝ์ผ๋ก ๋ฏธ๋๋ฌ์ง๋ฏ ์ด๋ํ๋ฉฐ ๋ฌ ๋จผ์ง๋ฅผ ์ผ์ผํด. ํฐ ์ฐ์ฃผ๋ณต์ ์ ์ ์ฐ์ฃผ์ธ๋ค์ด ๋ฌ ํน์ ์ ๋ฐ๋ ๋์์ผ๋ก ํ์น. ๋จผ ๋ฐฐ๊ฒฝ์์ VTOL ๋นํ์ฒด๊ฐ ์์ง์ผ๋ก ํ๊ฐํ์ฌ ํ๋ฉด์ ์กฐ์ฉํ ์ฐฉ๋ฅ. ์ฅ๋ฉด ์ ์ฒด์ ๊ฑธ์ณ ์ดํ์ค์ ์ธ ์ค๋ก๋ผ๊ฐ ๋ณ์ด ๊ฐ๋ํ ํ๋์ ๊ฐ๋ก์ง๋ฅด๋ฉฐ ์ถค์ถ๊ณ , ๋ น์, ํ๋์, ๋ณด๋ผ์ ๋น์ ์ปคํผ์ด ๋ฌ ํ๊ฒฝ์ ์ ๋น๋กญ๊ณ ๋ง๋ฒ ๊ฐ์ ๋น์ผ๋ก ๊ฐ์.", | |
| 4, | |
| ], | |
| [ | |
| "kill_bill.jpeg", | |
| "์ฐ๋ง ์๋จผ์ ์บ๋ฆญํฐ ๋ฒ ์ํธ๋ฆญ์ค ํค๋๊ฐ ์ํ ๊ฐ์ ์กฐ๋ช ์์์ ๋ ์นด๋ก์ด ์นดํ๋ ๊ฒ์ ์์ ์ ์ผ๋ก ๋ค๊ณ ์์. ๊ฐ์๊ธฐ ๊ดํ ๋๋ ๊ฐ์ฒ ์ด ๋ถ๋๋ฌ์์ง๊ณ ์๊ณก๋๊ธฐ ์์ํ๋ฉฐ ๊ฐ์ด๋ ๊ธ์์ฒ๋ผ ๊ตฌ์กฐ์ ์์ ์ฑ์ ์๊ธฐ ์์. ๊ฒ๋ ์ ์๋ฒฝํ ๋์ด ์ฒ์ฒํ ํ์ด์ง๊ณ ๋์ด์ง๋ฉฐ, ๋ น์ ๊ฐ์ฒ ์ด ์๋น ๋ฌผ์ค๊ธฐ๋ก ์๋๋ก ํ๋ฌ๋ด๋ฆผ. ๋ณํ์ ์ฒ์์๋ ๋ฏธ๋ฌํ๊ฒ ์์๋๋ค๊ฐ ๊ธ์์ด ์ ์ ๋ ์ ๋์ ์ด ๋๋ฉด์ ๊ฐ์ํ. ์นด๋ฉ๋ผ๋ ๊ทธ๋ ์ ์ผ๊ตด์ ๊ณ ์ ํ๊ณ ๋ ์นด๋ก์ด ๋๋น์ด ์ ์ฐจ ์ข์์ง๋๋ฐ, ์น๋ช ์ ์ธ ์ง์ค์ด ์๋๋ผ ๋ฌด๊ธฐ๊ฐ ๋์์์ ๋ น๋ ๊ฒ์ ๋ณด๋ฉฐ ํผ๋๊ณผ ๊ฒฝ์ . ํธํก์ด ์ฝ๊ฐ ๋นจ๋ผ์ง๋ฉฐ ์ด ๋ถ๊ฐ๋ฅํ ๋ณํ์ ๋ชฉ๊ฒฉ. ๋ น๋ ํ์์ด ๊ฐํ๋๊ณ ์นดํ๋์ ์๋ฒฝํ ํํ๊ฐ ์ ์ ์ถ์์ ์ด ๋๋ฉฐ ์์์ ์์์ฒ๋ผ ๋จ์ด์ง. ๋ น์ ๋ฐฉ์ธ์ด ๋ถ๋๋ฌ์ด ๊ธ์ ์ถฉ๊ฒฉ์๊ณผ ํจ๊ป ๋ฐ๋ฅ์ ๋จ์ด์ง. ํ์ ์ด ์ฐจ๋ถํ ์ค๋น์์ ๋นํน๊ฐ๊ณผ ์ฐ๋ ค๋ก ๋ฐ๋๋ฉฐ ์ ์ค์ ์ธ ๋ณต์์ ๋๊ตฌ๊ฐ ์์์ ๋ฌธ์ ๊ทธ๋๋ก ์กํ๋์ด ๋ฌด๋ฐฉ๋น ์ํ๊ฐ ๋จ.", | |
| 6, | |
| ], | |
| ], | |
| inputs=[input_image_component, prompt_input, steps_slider], | |
| outputs=[video_output, seed_input], | |
| fn=generate_video, | |
| cache_examples="lazy" | |
| ) | |
| if __name__ == "__main__": | |
| demo.queue().launch(mcp_server=True) |