app.py

# PyTorch 2.8 (temporary hack)
import os
os.system('pip install --upgrade --pre --extra-index-url https://download.pytorch.org/whl/nightly/cu126 "torch<2.9" spaces')

# Actual demo code
import spaces
import torch
from diffusers import WanPipeline, AutoencoderKLWan
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 optimization import optimize_pipeline_


MODEL_ID = "Wan-AI/Wan2.2-T2V-A14B-Diffusers"

LANDSCAPE_WIDTH = 1024
LANDSCAPE_HEIGHT = 1024
MAX_SEED = np.iinfo(np.int32).max

FIXED_FPS = 16
MIN_FRAMES_MODEL = 8
MAX_FRAMES_MODEL = 81

MIN_DURATION = round(MIN_FRAMES_MODEL/FIXED_FPS,1)
MAX_DURATION = round(MAX_FRAMES_MODEL/FIXED_FPS,1)

vae = AutoencoderKLWan.from_pretrained("Wan-AI/Wan2.2-T2V-A14B-Diffusers", subfolder="vae", torch_dtype=torch.float32)
pipe = WanPipeline.from_pretrained(MODEL_ID,
    transformer=WanTransformer3DModel.from_pretrained('linoyts/Wan2.2-T2V-A14B-Diffusers-BF16',
        subfolder='transformer',
        torch_dtype=torch.bfloat16,
        device_map='cuda',
    ),
    transformer_2=WanTransformer3DModel.from_pretrained('linoyts/Wan2.2-T2V-A14B-Diffusers-BF16',
        subfolder='transformer_2',
        torch_dtype=torch.bfloat16,
        device_map='cuda',
    ),
    vae=vae,
    torch_dtype=torch.bfloat16,
).to('cuda')


for i in range(3): 
    gc.collect()
    torch.cuda.synchronize() 
    torch.cuda.empty_cache()

optimize_pipeline_(pipe,
    prompt='prompt',
    height=LANDSCAPE_HEIGHT,
    width=LANDSCAPE_WIDTH,
    num_frames=MAX_FRAMES_MODEL,
)


default_prompt_i2v = "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage."
default_negative_prompt = "色调艳丽, 过曝, 静态, 细节模糊不清, 字幕, 风格, 作品, 画作, 画面, 静止, 整体发灰, 最差质量, 低质量, JPEG压缩残留, 丑陋的, 残缺的, 多余的手指, 画得不好的手部, 画得不好的脸部, 畸形的, 毁容的, 形态畸形的肢体, 手指融合, 静止不动的画面, 杂乱的背景, 三条腿, 背景人很多, 倒着走"


def get_duration(
    prompt,
    negative_prompt,
    guidance_scale,
    guidance_scale_2,
    steps,
    seed,
    randomize_seed,
    progress,
):
    return steps * 15

@spaces.GPU(duration=get_duration)
def generate_image(
    prompt,
    negative_prompt=default_negative_prompt,
    guidance_scale = 3.5,
    guidance_scale_2 = 4,
    steps = 27,
    seed = 42,
    randomize_seed = False,
    progress=gr.Progress(track_tqdm=True),
):
    """
    Generate a video from an input image using the Wan 2.2 14B I2V model with Phantom LoRA.
    
    This function takes an input image and generates a video animation based on the provided
    prompt and parameters. It uses an FP8 qunatized Wan 2.2 14B Image-to-Video model in with Phantom LoRA
    for fast generation in 6-8 steps.
    
    Args:
        prompt (str): Text prompt describing the desired animation or motion.
        negative_prompt (str, optional): Negative prompt to avoid unwanted elements. 
            Defaults to default_negative_prompt (contains unwanted visual artifacts).
        guidance_scale (float, optional): Controls adherence to the prompt. Higher values = more adherence.
            Defaults to 1.0. Range: 0.0-20.0.
        guidance_scale_2 (float, optional): Controls adherence to the prompt. Higher values = more adherence.
            Defaults to 1.0. Range: 0.0-20.0.
        steps (int, optional): Number of inference steps. More steps = higher quality but slower.
            Defaults to 4. Range: 1-30.
        seed (int, optional): Random seed for reproducible results. Defaults to 42.
            Range: 0 to MAX_SEED (2147483647).
        randomize_seed (bool, optional): Whether to use a random seed instead of the provided seed.
            Defaults to False.
        progress (gr.Progress, optional): Gradio progress tracker. Defaults to gr.Progress(track_tqdm=True).
    
    Returns:
        tuple: A tuple containing:
            - video_path (str): Path to the generated video file (.mp4)
            - current_seed (int): The seed used for generation (useful when randomize_seed=True)
    
    Raises:
        gr.Error: If input_image is None (no image uploaded).
    
    Note:
        - The function automatically resizes the input image to the target dimensions
        - Frame count is calculated as duration_seconds * FIXED_FPS (24)
        - Output dimensions are adjusted to be multiples of MOD_VALUE (32)
        - The function uses GPU acceleration via the @spaces.GPU decorator
        - Generation time varies based on steps and duration (see get_duration function)
    """
    
   
    current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)

    out_img = pipe(
        prompt=prompt,
        negative_prompt=negative_prompt,
        height=1024,
        width=1024,
        num_frames=1,
        guidance_scale=float(guidance_scale),
        guidance_scale_2=float(guidance_scale_2),
        num_inference_steps=int(steps),
        output_type="pil",
        generator=torch.Generator(device="cuda").manual_seed(current_seed),
    ).frames[0][0]

    return out_img, current_seed

with gr.Blocks() as demo:
    gr.Markdown("# Wan 2.2 T2I (14B)")
    #gr.Markdown("run Wan 2.2 in just 6-8 steps, with [FusionX Phantom LoRA by DeeJayT](https://huggingface.co/vrgamedevgirl84/Wan14BT2VFusioniX/tree/main/FusionX_LoRa), compatible with 🧨 diffusers")
    with gr.Row():
        with gr.Column():
            prompt_input = gr.Textbox(label="Prompt", value=default_prompt_i2v)
            #duration_seconds_input = gr.Slider(minimum=MIN_DURATION, maximum=MAX_DURATION, step=0.1, value=MAX_DURATION, label="Duration (seconds)", info=f"Clamped to model's {MIN_FRAMES_MODEL}-{MAX_FRAMES_MODEL} frames at {FIXED_FPS}fps.")
            
            with gr.Accordion("Advanced Settings", open=False):
                negative_prompt_input = gr.Textbox(label="Negative Prompt", value=default_negative_prompt, lines=3)
                seed_input = gr.Slider(label="Seed", minimum=0, maximum=MAX_SEED, step=1, value=42, interactive=True)
                randomize_seed_checkbox = gr.Checkbox(label="Randomize seed", value=True, interactive=True)
                steps_slider = gr.Slider(minimum=1, maximum=30, step=1, value=27, label="Inference Steps") 
                guidance_scale_input = gr.Slider(minimum=0.0, maximum=10.0, step=0.5, value=3.5, label="Guidance Scale - high noise stage")
                guidance_scale_2_input = gr.Slider(minimum=0.0, maximum=10.0, step=0.5, value=4, label="Guidance Scale 2 - low noise stage")

            generate_button = gr.Button("Generate Image", variant="primary")
        with gr.Column():
            img_output = gr.Image(label="Generated Image", interactive=False)
    
    ui_inputs = [ 
        prompt_input,
        negative_prompt_input,
        guidance_scale_input, guidance_scale_2_input, steps_slider, seed_input, randomize_seed_checkbox
    ]
    generate_button.click(fn=generate_image, inputs=ui_inputs, outputs=[img_output, seed_input])

    gr.Examples(
        examples=[ 
            [
                "Two anthropomorphic cats in comfy boxing gear and bright gloves fight intensely on a spotlighted stage."
            ],
        ],
        inputs=[prompt_input], outputs=[img_output, seed_input], fn=generate_image, cache_examples="lazy"
    )

if __name__ == "__main__":
    demo.queue().launch(mcp_server=True)