app.py

import os; os.system('pip install --upgrade --no-deps spaces')
import spaces
import shutil
import subprocess
import sys
import copy
import random
import tempfile
import warnings
import time
import gc
import uuid
from tqdm import tqdm
import cv2
import numpy as np
import torch
import torch._dynamo
from huggingface_hub import list_models
from torch.nn import functional as F
from PIL import Image

import gradio as gr
from diffusers import (
    FlowMatchEulerDiscreteScheduler,
    SASolverScheduler,
    DEISMultistepScheduler,
    DPMSolverMultistepInverseScheduler,
    UniPCMultistepScheduler,
    DPMSolverMultistepScheduler,
    DPMSolverSinglestepScheduler,
)
from diffusers.pipelines.wan.pipeline_wan_i2v import WanImageToVideoPipeline
from diffusers.utils.export_utils import export_to_video

from torchao.quantization import quantize_, Float8DynamicActivationFloat8WeightConfig, Int8WeightOnlyConfig
import aoti

os.environ["TOKENIZERS_PARALLELISM"] = "true"
warnings.filterwarnings("ignore")
IS_ZERO_GPU = bool(os.getenv("SPACES_ZERO_GPU"))

# if IS_ZERO_GPU:
#     print("Loading...")
#     subprocess.run("rm -rf /data-nvme/zerogpu-offload/*", env={}, shell=True)

# --- FRAME EXTRACTION JS & LOGIC ---

# JS to grab timestamp from the output video
get_timestamp_js = """
function() {
    // Select the video element specifically inside the component with id 'generated-video'
    const video = document.querySelector('#generated-video video');
    
    if (video) {
        console.log("Video found! Time: " + video.currentTime);
        return video.currentTime;
    } else {
        console.log("No video element found.");
        return 0;
    }
}
"""


def extract_frame(video_path, timestamp):
    # Safety check: if no video is present
    if not video_path:
        return None
    
    print(f"Extracting frame at timestamp: {timestamp}") 
    
    cap = cv2.VideoCapture(video_path)
    
    if not cap.isOpened():
        return None

    # Calculate frame number
    fps = cap.get(cv2.CAP_PROP_FPS)
    target_frame_num = int(float(timestamp) * fps)
    
    # Cap total frames to prevent errors at the very end of video
    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    if target_frame_num >= total_frames:
        target_frame_num = total_frames - 1
    
    # Set position
    cap.set(cv2.CAP_PROP_POS_FRAMES, target_frame_num)
    ret, frame = cap.read()
    cap.release()
    
    if ret:
        # Convert from BGR (OpenCV) to RGB (Gradio)
        # Gradio Image component handles Numpy array -> PIL conversion automatically
        return cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
    
    return None

# --- END FRAME EXTRACTION LOGIC ---


def clear_vram():
    gc.collect()
    torch.cuda.empty_cache()


# RIFE
if not os.path.exists("RIFEv4.26_0921.zip"):
    print("Downloading RIFE Model...")
    subprocess.run([
        "wget", "-q",
        "https://huggingface.co/r3gm/RIFE/resolve/main/RIFEv4.26_0921.zip",
        "-O", "RIFEv4.26_0921.zip"
    ], check=True)
    subprocess.run(["unzip", "-o", "RIFEv4.26_0921.zip"], check=True)

# sys.path.append(os.getcwd())

from train_log.RIFE_HDv3 import Model
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
rife_model = Model()
rife_model.load_model("train_log", -1)
rife_model.eval()


@torch.no_grad()
def interpolate_bits(frames_np, multiplier=2, scale=1.0):
    """
    Interpolation maintaining Numpy Float 0-1 format.
    Args:
        frames_np: Numpy Array (Time, Height, Width, Channels) - Float32 [0.0, 1.0]
        multiplier: int (2, 4, 8)
    Returns:
        List of Numpy Arrays (Height, Width, Channels) - Float32 [0.0, 1.0]
    """
    
    # Handle input shape
    if isinstance(frames_np, list):
        # Convert list of arrays to one big array for easier shape handling if needed, 
        # but here we just grab dims from first frame
        T = len(frames_np)
        H, W, C = frames_np[0].shape
    else:
        T, H, W, C = frames_np.shape

    # 1. No Interpolation Case
    if multiplier < 2:
        # Just convert 4D array to list of 3D arrays
        if isinstance(frames_np, np.ndarray):
            return list(frames_np)
        return frames_np

    n_interp = multiplier - 1
    
    # Pre-calc padding for RIFE (requires dimensions divisible by 32/scale)
    tmp = max(128, int(128 / scale))
    ph = ((H - 1) // tmp + 1) * tmp
    pw = ((W - 1) // tmp + 1) * tmp
    padding = (0, pw - W, 0, ph - H)

    # Helper: Numpy (H, W, C) Float -> Tensor (1, C, H, W) Half
    def to_tensor(frame_np):
        # frame_np is float32 0-1
        t = torch.from_numpy(frame_np).to(device)
        # HWC -> CHW
        t = t.permute(2, 0, 1).unsqueeze(0)
        return F.pad(t, padding).half()

    # Helper: Tensor (1, C, H, W) Half -> Numpy (H, W, C) Float
    def from_tensor(tensor):
        # Crop padding
        t = tensor[0, :, :H, :W]
        # CHW -> HWC
        t = t.permute(1, 2, 0)
        # Keep as float32, range 0-1
        return t.float().cpu().numpy()

    def make_inference(I0, I1, n):
        if rife_model.version >= 3.9:
            res = []
            for i in range(n):
                res.append(rife_model.inference(I0, I1, (i+1) * 1. / (n+1), scale))
            return res
        else:
            middle = rife_model.inference(I0, I1, scale)
            if n == 1:
                return [middle]
            first_half = make_inference(I0, middle, n=n//2)
            second_half = make_inference(middle, I1, n=n//2)
            if n % 2:
                return [*first_half, middle, *second_half]
            else:
                return [*first_half, *second_half]

    output_frames = []

    # Process Frames
    # Load first frame into GPU
    I1 = to_tensor(frames_np[0])

    total_steps = T - 1

    with tqdm(total=total_steps, desc="Interpolating", unit="frame") as pbar:
    
        for i in range(total_steps):
            I0 = I1
            # Add original frame to output
            output_frames.append(from_tensor(I0))
    
            # Load next frame
            I1 = to_tensor(frames_np[i+1])
    
            # Generate intermediate frames
            mid_tensors = make_inference(I0, I1, n_interp)
    
            # Append intermediate frames
            for mid in mid_tensors:
                output_frames.append(from_tensor(mid))

            if (i + 1) % 50 == 0:
                pbar.update(50)
        pbar.update(total_steps % 50)
        
        # Add the very last frame
        output_frames.append(from_tensor(I1))
    
    # Cleanup
    del I0, I1, mid_tensors
    torch.cuda.empty_cache()

    return output_frames


# WAN

ORG_NAME = "TestOrganizationPleaseIgnore"
# MODEL_ID = "Wan-AI/Wan2.2-I2V-A14B-Diffusers"
MODEL_ID = os.getenv("REPO_ID") or random.choice(
    list(list_models(author=ORG_NAME, filter='diffusers:WanImageToVideoPipeline'))
).modelId
CACHE_DIR = os.path.expanduser("~/.cache/huggingface/")

LORA_MODELS = [
    # {
    #     "repo_id": "exampleuser/example_lora_1",
    #     "high_tr": "example_lora_1_high.safetensors",
    #     "low_tr": "example_lora_1_low.safetensors",
    #     "high_scale": 0.5,
    #     "low_scale": 0.5
    # },
    # {
    #     "repo_id": "exampleuser/example_lora_2",
    #     "high_tr": "subfolder/example_lora_2_high.safetensors",
    #     "low_tr": "subfolder/example_lora_2_low.safetensors",
    #     "high_scale": 0.4,
    #     "low_scale": 0.4
    # },
]

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 = 160

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

SCHEDULER_MAP = {
    "FlowMatchEulerDiscrete": FlowMatchEulerDiscreteScheduler,
    "SASolver": SASolverScheduler,
    "DEISMultistep": DEISMultistepScheduler,
    "DPMSolverMultistepInverse": DPMSolverMultistepInverseScheduler,
    "UniPCMultistep": UniPCMultistepScheduler,
    "DPMSolverMultistep": DPMSolverMultistepScheduler,
    "DPMSolverSinglestep": DPMSolverSinglestepScheduler,
}

pipe = WanImageToVideoPipeline.from_pretrained(
    MODEL_ID,
    torch_dtype=torch.bfloat16,
).to('cuda')
original_scheduler = copy.deepcopy(pipe.scheduler)

for i, lora in enumerate(LORA_MODELS):
    name_high_tr = lora["high_tr"].split(".")[0].split("/")[-1] + "Hh"
    name_low_tr = lora["low_tr"].split(".")[0].split("/")[-1] + "Ll"
    
    try: 
        pipe.load_lora_weights(
            lora["repo_id"],
            weight_name=lora["high_tr"],
            adapter_name=name_high_tr
        )
    
        kwargs_lora = {"load_into_transformer_2": True}
        pipe.load_lora_weights(
            lora["repo_id"],
            weight_name=lora["low_tr"],
            adapter_name=name_low_tr,
            **kwargs_lora
        )
    
        pipe.set_adapters([name_high_tr, name_low_tr], adapter_weights=[1.0, 1.0])
    
        pipe.fuse_lora(adapter_names=[name_high_tr], lora_scale=lora["high_scale"], components=["transformer"])
        pipe.fuse_lora(adapter_names=[name_low_tr], lora_scale=lora["low_scale"], components=["transformer_2"])
    
        pipe.unload_lora_weights()

        print(f"Applied: {lora['high_tr']}, hs={lora['high_scale']}/ls={lora['low_scale']}, {i+1}/{len(LORA_MODELS)}") 
    except Exception as e:
        print("Error:", str(e))
        print("Failed LoRA:", name_high_tr)
        pipe.unload_lora_weights()

# if os.path.exists(CACHE_DIR):
#     shutil.rmtree(CACHE_DIR)
#     print("Deleted Hugging Face cache.")
# else:
#     print("No hub cache found.")

quantize_(pipe.text_encoder, Int8WeightOnlyConfig())
torch._dynamo.reset()
quantize_(pipe.transformer, Float8DynamicActivationFloat8WeightConfig())
torch._dynamo.reset()
quantize_(pipe.transformer_2, Float8DynamicActivationFloat8WeightConfig())
torch._dynamo.reset()

spaces.aoti_load(
    module=pipe.transformer,
    repo_id='cbensimon/WanTransformer3DModel-sm120-cu130-raa',
)
spaces.aoti_load(
    module=pipe.transformer_2,
    repo_id='cbensimon/WanTransformer3DModel-sm120-cu130-raa',
)

# pipe.vae.enable_slicing()
# pipe.vae.enable_tiling()

default_prompt_i2v = "make this image come alive, cinematic motion, smooth animation"
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 translate_prompt_to_en(text):
    """Wan 2.2's text encoder understands English/Chinese, not Turkish. Translate
    the (Turkish) prompt to English before inference. Falls back to the raw text
    if translation is unavailable, so English prompts pass through unchanged."""
    if not text or not text.strip():
        return text
    try:
        from deep_translator import GoogleTranslator
        out = GoogleTranslator(source="auto", target="en").translate(text)
        print(f"[translate] {text!r} -> {out!r}")
        return out or text
    except Exception as e:
        print("[translate] failed, using raw prompt:", e)
        return text


def model_title():
    repo_name = MODEL_ID.split('/')[-1].replace("_", " ")
    url = f"https://huggingface.co/{MODEL_ID}"
    return f"## This space is currently running [{repo_name}]({url}) 🐢"


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 resize_and_crop_to_match(target_image, reference_image):
    ref_width, ref_height = reference_image.size
    target_width, target_height = target_image.size
    scale = max(ref_width / target_width, ref_height / target_height)
    new_width, new_height = int(target_width * scale), int(target_height * scale)
    resized = target_image.resize((new_width, new_height), Image.Resampling.LANCZOS)
    left, top = (new_width - ref_width) // 2, (new_height - ref_height) // 2
    return resized.crop((left, top, left + ref_width, top + ref_height))


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_inference_duration(
    resized_image,
    processed_last_image,
    prompt,
    steps,
    negative_prompt,
    num_frames,
    guidance_scale,
    guidance_scale_2,
    current_seed,
    scheduler_name,
    flow_shift,
    frame_multiplier,
    quality,
    duration_seconds,
    safe_mode,
    progress
):
    BASE_FRAMES_HEIGHT_WIDTH = 81 * 832 * 624
    BASE_STEP_DURATION = 5.
    width, height = resized_image.size
    factor = num_frames * width * height / BASE_FRAMES_HEIGHT_WIDTH
    step_duration = BASE_STEP_DURATION * factor ** 1.5
    gen_time = int(steps) * step_duration

    if guidance_scale > 1:
        gen_time = gen_time * 2.4

    frame_factor = frame_multiplier // FIXED_FPS
    if frame_factor > 1:
        total_out_frames = (num_frames * frame_factor) - num_frames
        inter_time = (total_out_frames * 0.02)
        gen_time += inter_time

    total_time = 15 + gen_time
    if safe_mode:
        total_time = total_time * 1.30

    return total_time


@spaces.GPU(duration=get_inference_duration, size='xlarge')
def run_inference(
    resized_image,
    processed_last_image,
    prompt,
    steps,
    negative_prompt,
    num_frames,
    guidance_scale,
    guidance_scale_2,
    current_seed,
    scheduler_name,
    flow_shift,
    frame_multiplier,
    quality,
    duration_seconds,
    safe_mode=False,
    progress=gr.Progress(track_tqdm=True),
):
    scheduler_class = SCHEDULER_MAP.get(scheduler_name)
    if scheduler_class.__name__ != pipe.scheduler.config._class_name or flow_shift != pipe.scheduler.config.get("flow_shift", "shift"):
        config = copy.deepcopy(original_scheduler.config)
        if scheduler_class == FlowMatchEulerDiscreteScheduler:
            config['shift'] = flow_shift
        else:
            config['flow_shift'] = flow_shift
        pipe.scheduler = scheduler_class.from_config(config)

    clear_vram()

    task_name = str(uuid.uuid4())[:8]
    print(f"Generating {num_frames} frames, task: {task_name}, {duration_seconds}, {resized_image.size}")
    start = time.time()
    result = pipe(
        image=resized_image,
        last_image=processed_last_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),
        output_type="np" 
    )
    print("gen time passed:", time.time() - start)
    
    raw_frames_np = result.frames[0]  # Returns (T, H, W, C) float32
    pipe.scheduler = original_scheduler

    frame_factor = frame_multiplier // FIXED_FPS
    if frame_factor > 1:
        start = time.time()
        print(f"Processing frames (RIFE Multiplier: {frame_factor}x)...")
        rife_model.device()
        rife_model.flownet = rife_model.flownet.half()
        final_frames = interpolate_bits(raw_frames_np, multiplier=int(frame_factor))
        print("Interpolation time passed:", time.time() - start)
    else:
        final_frames = list(raw_frames_np)

    final_fps = FIXED_FPS * int(frame_factor)

    with tempfile.NamedTemporaryFile(suffix=".mp4", delete=False) as tmpfile:
        video_path = tmpfile.name

    start = time.time()
    with tqdm(total=3, desc="Rendering Media", unit="clip") as pbar:
        pbar.update(2)
        export_to_video(final_frames, video_path, fps=final_fps, quality=quality)
        pbar.update(1)
    print(f"Export time passed, {final_fps} FPS:", time.time() - start)

    return video_path, task_name


def generate_video(
    input_image,
    last_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,
    quality=5,
    scheduler="UniPCMultistep",
    flow_shift=6.0,
    frame_multiplier=16,
    safe_mode=False,
    video_component=True,
    progress=gr.Progress(track_tqdm=True),
):
    """
    Generate a video from an input image using the Wan 2.2 14B I2V model with Lightning 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 Lightning LoRA
    for fast generation in 4-8 steps.
    Args:
        input_image (PIL.Image): The input image to animate. Will be resized to target dimensions.
        last_image (PIL.Image, optional): The optional last image for the video.
        prompt (str): Text prompt describing the desired animation or motion.
        steps (int, optional): Number of inference steps. More steps = higher quality but slower.
            Defaults to 4. Range: 1-30.
        negative_prompt (str, optional): Negative prompt to avoid unwanted elements.
            Defaults to default_negative_prompt (contains unwanted visual artifacts).
        duration_seconds (float, optional): Duration of the generated video in seconds.
            Defaults to 2. Clamped between MIN_FRAMES_MODEL/FIXED_FPS and MAX_FRAMES_MODEL/FIXED_FPS.
        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.
        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.
        quality (float, optional): Video output quality. Default is 5. Uses variable bit rate.
            Highest quality is 10, lowest is 1.
        scheduler (str, optional): The name of the scheduler to use for inference. Defaults to "UniPCMultistep".
        flow_shift (float, optional): The flow shift value for compatible schedulers. Defaults to 6.0.
        frame_multiplier (int, optional): The int value for fps enhancer
        video_component(bool, optional): Show video player in output.
            Defaults to True.
        progress (gr.Progress, optional): Gradio progress tracker. Defaults to gr.Progress(track_tqdm=True).
    Returns:
        tuple: A tuple containing:
            - video_path (str): Path for the video component.
            - video_path (str): Path for the file download component. Attempt to avoid reconversion in video component.
            - current_seed (int): The seed used for generation.
    Raises:
        gr.Error: If input_image is None (no image uploaded).
    Note:
        - 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)
    """
    
    if input_image is None:
        raise gr.Error("Lütfen bir çizim yükleyin.")

    prompt = translate_prompt_to_en(prompt)

    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)

    processed_last_image = None
    if last_image:
        processed_last_image = resize_and_crop_to_match(last_image, resized_image)

    video_path, task_n = run_inference(
        resized_image,
        processed_last_image,
        prompt,
        steps,
        negative_prompt,
        num_frames,
        guidance_scale,
        guidance_scale_2,
        current_seed,
        scheduler,
        flow_shift,
        frame_multiplier,
        quality,
        duration_seconds,
        safe_mode,
        progress,
    )
    print(f"GPU complete: {task_n}")

    return (video_path if video_component else None), video_path, current_seed


CSS = """
#hidden-timestamp {
    opacity: 0;
    height: 0px;
    width: 0px;
    margin: 0px;
    padding: 0px;
    overflow: hidden;
    position: absolute;
    pointer-events: none;
}

/* View swap (login -> app) is done by JS toggling .lm-hidden on these IDs.
   Gradio 6 silently drops Column visibility updates from handler outputs,
   so we have to do this client-side. */
#lemanime-login.lm-hidden, #lemanime-app.lm-hidden {
    display: none !important;
}
/* Hide access_key_input via CSS rather than `visible=False`. Verified bug in
   Gradio 6.0.1: `gr.update(value=...)` to a `visible=False` Textbox does NOT
   propagate to server-side state — so Generate read empty string and the
   key check always failed. CSS-hiding keeps the component "alive" so its
   value updates flow through to the server. */
#lemanime-access-key {
    display: none !important;
}

/* ── lemanime polish ─────────────────────────────────────────────────── */
#lemanime-login, #lemanime-app {
    max-width: 1100px;
    margin: 0 auto;
}
#lemanime-login {
    max-width: 460px;
    padding: 2rem 1rem;
}
#lemanime-login .form, #lemanime-login button {
    margin-top: 0.5rem;
}
/* Keep the input image preview a sensible size — Gradio 6 sizes it raw
   (often huge) which dominates the page. */
#lemanime-app .gradio-image img,
#lemanime-app .image-container img {
    max-height: 340px !important;
    object-fit: contain !important;
    width: 100% !important;
}
/* Result video isn't bigger than the column it sits in */
#lemanime-app video {
    max-height: 480px !important;
    width: 100% !important;
}
/* Section group cards: soft borders + breathing room */
#lemanime-app .gr-group, #lemanime-app .gradio-group {
    border-radius: 14px !important;
    padding: 0.5rem 0.25rem !important;
}
/* Bigger Generate button so it doesn't look like just another button */
#lemanime-app button.lg, #lemanime-app button[variant="primary"] {
    font-size: 1.05rem !important;
    padding: 0.85rem 1.5rem !important;
}
/* Subtle quick-prompt button row */
#lemanime-app .gradio-row button {
    background: rgba(255, 200, 100, 0.08) !important;
    border: 1px solid rgba(255, 180, 80, 0.25) !important;
}
"""


TR_DEFAULT_PROMPT = "bu çizimi canlandır, akıcı 2D mürekkep çizgi animasyonu, beyaz arka plan"

TR_EXAMPLE_PROMPTS = [
    "saçları kuvvetli rüzgarda savruluyor, başını hafifçe çeviriyor",
    "başını yavaşça sağa sola çeviriyor, hafifçe sallanıyor",
    "kollarını sallayarak yerinde yürüyor",
    "havada süzülüyor, kolları yana açık, saçları yukarı savruluyor",
    "öne doğru eğilip kollarını uzatıyor",
]

TR_PROMPT_TIP = (
    "💡 **Daha iyi sonuç için ipucu:** Büyük gövde/uzuv hareketleri (saç savrulması, "
    "baş çevirme, yürüyüş, eğilme) **iyi sonuç verir**. Çok ince yüz hareketleri "
    "(göz kırpma, gülümseme, kaş kaldırma) karakteri **bozabilir** — model yüzü "
    "yeniden çizmeye çalışır ve karakter değişebilir."
)

# ── Erişim kontrolü (ACCESS_KEY) ──────────────────────────────────────────
# Custom single-field login page: NOT Gradio's built-in auth (which forces a
# username field). Same secret is enforced server-side inside generate_video
# so /generate_video API calls also require the key (programmatic clients pass
# it as `client_access_key`). When ACCESS_KEY is unset, Space launches open.
import secrets as _secrets

ACCESS_KEY = os.environ.get("ACCESS_KEY")
AUTH_REQUIRED = bool(ACCESS_KEY)


def _check_key(client_access_key):
    if not AUTH_REQUIRED:
        return True
    return _secrets.compare_digest(client_access_key or "", ACCESS_KEY or "")


def generate_video_authed(
    client_access_key,
    input_image,
    last_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,
    quality=5,
    scheduler="UniPCMultistep",
    flow_shift=6.0,
    frame_multiplier=16,
    safe_mode=False,
    video_component=True,
    progress=gr.Progress(track_tqdm=True),
):
    """Server-side gate that enforces the access key on BOTH the UI's Generate
    button (which passes a hidden field set by the login form) and direct
    /generate_video API calls (programmatic clients pass `client_access_key`).
    """
    if not _check_key(client_access_key):
        raise gr.Error("🔒 Geçersiz erişim anahtarı. Lütfen tekrar deneyin.")
    return generate_video(
        input_image, last_image, prompt, steps, negative_prompt,
        duration_seconds, guidance_scale, guidance_scale_2, seed, randomize_seed,
        quality, scheduler, flow_shift, frame_multiplier, safe_mode,
        video_component, progress,
    )


# NB: Gradio 6 on this Space rejects several cosmetic kwargs from 5.x
# (theme=, show_copy_button=, height=, size=, examples_per_page=,
# equal_height=...). Keep components minimal — structure via Groups +
# Markdown headers + Examples instead of cosmetic kwargs.

with gr.Blocks(delete_cache=(3600, 10800), title="Lemanime") as demo:
    # Hidden field carrying the access key. The login form sets it via gr.update;
    # programmatic gradio_client callers pass it directly as client_access_key.
    # NOTE: hidden via CSS (elem_id), NOT visible=False — see CSS comment above
    # for the Gradio 6.0.1 bug that prevented value updates to invisible Textboxes.
    access_key_input = gr.Textbox(value="", elem_id="lemanime-access-key")

    # ── LOGIN VIEW ────────────────────────────────────────────────────────
    # Gradio 6 silently drops Column/Group visibility updates from handler
    # outputs (verified locally). So we render BOTH views, use elem_id +
    # CSS class to hide one initially, and a .then() chain with a tiny JS
    # snippet to swap the class after a successful login.
    _login_start_classes = [] if AUTH_REQUIRED else ["lm-hidden"]
    _app_start_classes = ["lm-hidden"] if AUTH_REQUIRED else []
    with gr.Column(elem_id="lemanime-login", elem_classes=_login_start_classes) as login_view:
        gr.Markdown(
            "# ✏️ Lemanime\n"
            "### Çizim Canlandırma Stüdyosu\n\n"
            "Devam etmek için **erişim anahtarınızı** girin."
        )
        login_input = gr.Textbox(
            label="Erişim Anahtarı",
            type="password",
            placeholder="Anahtarınızı buraya yapıştırın",
        )
        login_btn = gr.Button("🔓 Giriş", variant="primary")
        login_error = gr.Markdown(value="", visible=False)

    # ── APP VIEW ──────────────────────────────────────────────────────────
    with gr.Column(elem_id="lemanime-app", elem_classes=_app_start_classes) as app_view:
        gr.Markdown(
            "# ✏️ Lemanime\n"
            "### Çizim Canlandırma Stüdyosu\n"
            "Çiziminizi yükleyin, istediğiniz hareketi **Türkçe** anlatın, "
            "**Animasyonu Oluştur**'a basın."
        )

        with gr.Row():
            # ── Sol sütun ─────────────────────────────────────────────────
            with gr.Column(scale=5):
                with gr.Group():
                    gr.Markdown("### 1️⃣ Çizim")
                    input_image_component = gr.Image(
                        type="pil",
                        label="Çiziminizi buraya bırakın veya yükleyin",
                        sources=["upload", "clipboard"],
                    )

                with gr.Group():
                    gr.Markdown("### 2️⃣ Hareketi anlatın")
                    prompt_input = gr.Textbox(
                        label="Hareket açıklaması (Türkçe)",
                        value=TR_DEFAULT_PROMPT,
                        lines=3,
                        placeholder="örn: saçları rüzgarda savruluyor, başını hafifçe çeviriyor",
                    )
                    # Custom quick-prompt row: gr.Examples styles awkwardly in
                    # Gradio 6 with no examples_per_page kwarg available.
                    # Use an explicit named factory (not lambda+default-arg)
                    # to be safe against Gradio's handler-registration weirdness.
                    def _make_fill(text):
                        def _fill():
                            return text
                        return _fill

                    gr.Markdown("**Hızlı örnekler:**")
                    with gr.Row():
                        for _short, _full in [
                            ("Saç savrulması", TR_EXAMPLE_PROMPTS[0]),
                            ("Baş sallama", TR_EXAMPLE_PROMPTS[1]),
                            ("Yürüyüş", TR_EXAMPLE_PROMPTS[2]),
                            ("Havada süzülme", TR_EXAMPLE_PROMPTS[3]),
                            ("Öne eğilme", TR_EXAMPLE_PROMPTS[4]),
                        ]:
                            _btn = gr.Button(_short)
                            _btn.click(fn=_make_fill(_full), inputs=None, outputs=[prompt_input])
                    gr.Markdown(TR_PROMPT_TIP)

                with gr.Group():
                    gr.Markdown("### ⚙️ Ayarlar")
                    duration_seconds_input = gr.Slider(
                        minimum=MIN_DURATION, maximum=MAX_DURATION, step=0.5, value=3.5,
                        label="Video uzunluğu (saniye)",
                        info=f"{MIN_DURATION}–{MAX_DURATION} sn. Uzun videolar daha fazla kare üretir.",
                    )
                    steps_slider = gr.Slider(
                        minimum=4, maximum=20, step=1, value=6,
                        label="Kalite adımları",
                        info="6 hızlı varsayılan · 8–12 belirgin iyileşme · 15–20 maks.",
                    )
                    frame_multi = gr.Dropdown(
                        choices=[FIXED_FPS, FIXED_FPS*2, FIXED_FPS*4, FIXED_FPS*8],
                        value=FIXED_FPS,
                        label="Akıcılık (kare/sn)",
                        info="16 = orijinal · 32/64/128 = RIFE ile daha akıcı hareket.",
                    )

                with gr.Accordion("İki kare arası geçiş (opsiyonel)", open=False):
                    gr.Markdown(
                        "Bir bitiş karesi eklerseniz, model iki kare arasında bir geçiş "
                        "oluşturmaya çalışır. İnce çizgili çizimlerde bu özellik **çoğunlukla "
                        "hayalet uzuv** sorunlarına yol açar — tek karelik animasyon genellikle "
                        "daha temizdir."
                    )
                    last_image_component = gr.Image(
                        type="pil", label="Bitiş karesi (Resim 2)",
                        sources=["upload", "clipboard"],
                    )

                generate_button = gr.Button("🎬 Animasyonu Oluştur", variant="primary")

            # ── Sağ sütun ─────────────────────────────────────────────────
            with gr.Column(scale=4):
                gr.Markdown("### 🎞️ Sonuç")
                video_output = gr.Video(
                    label="Oluşturulan video",
                    autoplay=True, interactive=False,
                    elem_id="generated-video",
                )
                file_output = gr.File(label="Video'yu indir")
                gr.Markdown("_Sonucunuz buraya gelecek. Üretim ~30 sn – 2 dk sürebilir._")

    # ── Gizli kontroller (API uyumluluğu için) ────────────────────────────
    negative_prompt_input = gr.Textbox(value=default_negative_prompt, visible=False)
    guidance_scale_input = gr.Slider(minimum=1.0, maximum=5.0, step=0.5, value=1.0, visible=False)
    guidance_scale_2_input = gr.Slider(minimum=1.0, maximum=5.0, step=0.5, value=1.0, visible=False)
    seed_input = gr.Slider(minimum=0, maximum=MAX_SEED, step=1, value=42, visible=False)
    randomize_seed_checkbox = gr.Checkbox(value=True, visible=False)
    quality_slider = gr.Slider(minimum=1, maximum=10, step=1, value=6, visible=False)
    scheduler_dropdown = gr.Dropdown(choices=list(SCHEDULER_MAP.keys()), value="UniPCMultistep", visible=False)
    flow_shift_slider = gr.Slider(minimum=0.5, maximum=15.0, step=0.1, value=3.0, visible=False)
    safe_mode_checkbox = gr.Checkbox(value=True, visible=False)
    play_result_video = gr.Checkbox(value=True, visible=False)

    ui_inputs = [
        input_image_component, last_image_component, prompt_input, steps_slider,
        negative_prompt_input, duration_seconds_input,
        guidance_scale_input, guidance_scale_2_input, seed_input, randomize_seed_checkbox,
        quality_slider, scheduler_dropdown, flow_shift_slider, frame_multi,
        safe_mode_checkbox, play_result_video,
    ]

    # ── Login işleyici ────────────────────────────────────────────────────
    # Python step: just update the key + error + clear password (Textbox /
    # Markdown — Gradio 6 applies these reliably). NO Column outputs here.
    # JS step (.then): reads access_key_input value and toggles `.lm-hidden`
    # on the two Columns by elem_id — the only way to actually swap views.
    def do_login(typed_key):
        ok = _check_key(typed_key)
        print(f"[login] attempt: len={len(typed_key or '')}  ok={ok}", flush=True)
        if ok:
            return (
                gr.update(value=typed_key or ""),       # access_key_input
                gr.update(value="", visible=False),     # login_error (clear)
                gr.update(value=""),                    # login_input (clear)
            )
        return (
            gr.update(value=""),                        # access_key_input (clear)
            gr.update(                                  # login_error (show)
                value="❌ Geçersiz erişim anahtarı. Tekrar deneyin.",
                visible=True,
            ),
            gr.update(),                                # login_input (keep)
        )

    # Verified empirically in Gradio 6.0.1 with Playwright: the ONLY pattern
    # where the JS-driven .lm-hidden toggle actually sticks is `js=` directly
    # on the click handler. In .then/.success/.change chains, the JS fires
    # but Gradio re-renders immediately after and reapplies the original
    # classes. js= on .click is a pre-processor (runs BEFORE the python fn,
    # sees the typed password as args[0]). Since js= runs pre-fn, the swap
    # decision is "is the input non-empty" — the actual key check still
    # happens server-side in generate_video_authed, so a wrong password gets
    # the user into the app view but Generate immediately raises
    # "🔒 Geçersiz erişim anahtarı".
    SWAP_JS = """
    (typed_key) => {
        const login = document.getElementById('lemanime-login');
        const app = document.getElementById('lemanime-app');
        if (typed_key && typed_key.length > 0) {
            if (login) login.classList.add('lm-hidden');
            if (app) app.classList.remove('lm-hidden');
        }
        return typed_key;
    }
    """

    _login_outputs = [access_key_input, login_error, login_input]
    login_btn.click(fn=do_login, inputs=[login_input], outputs=_login_outputs, js=SWAP_JS)
    login_input.submit(fn=do_login, inputs=[login_input], outputs=_login_outputs, js=SWAP_JS)

    generate_button.click(
        fn=generate_video_authed,
        inputs=[access_key_input] + ui_inputs,
        outputs=[video_output, file_output, seed_input],
        api_name="generate_video",
    )


if __name__ == "__main__":
    if AUTH_REQUIRED:
        print(f"[auth] ACCESS_KEY set ({len(ACCESS_KEY)} chars) — custom login required.")
    else:
        print("[auth] ACCESS_KEY not set — Space is open.")
    demo.queue().launch(mcp_server=True, css=CSS, show_error=True)