app.py

# =============================================================
# 0️⃣  FORCE ALL CACHES TO EPHEMERAL /tmp  (DO NOT COUNT TO 150 GB)
# =============================================================
import os, shutil, pathlib
# -----------------------------------------------------------------
#  Clean any leftover cache that may already be on the persistent volume.
#  This runs **once** at container start, before any import that touches HF.
# -----------------------------------------------------------------
for p in [
    pathlib.Path.home() / ".cache",
    pathlib.Path("/workspace") / ".cache",
    pathlib.Path("/tmp") / "hf_cache",
    pathlib.Path("/tmp") / "torch_home",
]:
    if p.exists():
        shutil.rmtree(p, ignore_errors=True)

# -----------------------------------------------------------------
#  Point every HF / torch cache to /tmp (which is a RAM‑disk and is
#  NOT counted against the Space’s disk quota).
# -----------------------------------------------------------------
os.environ["HF_HUB_CACHE"]       = "/tmp/hf_cache"
os.environ["TRANSFORMERS_CACHE"] = "/tmp/hf_cache"
os.environ["HF_DATASETS_CACHE"]  = "/tmp/hf_cache"
os.environ["TORCH_HOME"]         = "/tmp/torch_home"

# =============================================================
# 1️⃣  IMPORTS
# =============================================================
import spaces
import torch
import numpy as np
import random
import gc
import tempfile
import requests
import logging
from PIL import Image

import gradio as gr
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

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

logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)

# =============================================================
# 2️⃣  CONFIG
# =============================================================
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

default_prompt_i2v = "make this image come alive, cinematic motion, smooth animation"
default_negative_prompt = (
    "colorful tones, overexposed, static, unclear details, subtitles, style, artwork, painting, screen, "
    "still, overall gray, worst quality, low quality, JPEG compression artifacts, ugly, deformed, "
    "extra fingers, poorly drawn hands, poorly drawn face, deformed, mutated, deformed limbs, "
    "fused fingers, still screen, messy background, three legs, many people in background, walking backwards"
)

# ------------------------------------------------------------
# 3️⃣  TRANSLATOR (Albanian → English) – unchanged
# ------------------------------------------------------------
def translate_albanian_to_english(text: str) -> str:
    if not text.strip():
        return text
    for attempt in range(2):
        try:
            response = requests.post(
                "https://hal1993-mdftranslation1234567890abcdef1234567890-fc073a6.hf.space/v1/translate",
                json={"from_language": "sq", "to_language": "en", "input_text": text},
                headers={"accept": "application/json", "Content-Type": "application/json"},
                timeout=8,
            )
            response.raise_for_status()
            translated = response.json().get("translate", text)
            logger.info(f"Translated: {text[:50]}... → {translated[:50]}...")
            return translated.strip() or text
        except Exception as e:
            logger.warning(f"Translation failed (attempt {attempt + 1}): {e}")
            if attempt == 1:
                return text
    return text

# ------------------------------------------------------------
# 4️⃣  MODEL LOADING  (all caches forced to /tmp)
# ------------------------------------------------------------
pipe = WanImageToVideoPipeline.from_pretrained(
    "Wan-AI/Wan2.2-I2V-A14B-Diffusers",
    torch_dtype=torch.bfloat16,
    cache_dir="/tmp/hf_cache",
).to("cuda")

pipe.transformer = WanTransformer3DModel.from_pretrained(
    "cbensimon/Wan2.2-I2V-A14B-bf16-Diffusers",
    subfolder="transformer",
    torch_dtype=torch.bfloat16,
    device_map="cuda",
    cache_dir="/tmp/hf_cache",
)

pipe.transformer_2 = WanTransformer3DModel.from_pretrained(
    "cbensimon/Wan2.2-I2V-A14B-bf16-Diffusers",
    subfolder="transformer_2",
    torch_dtype=torch.bfloat16,
    device_map="cuda",
    cache_dir="/tmp/hf_cache",
)

# ---- LoRA -------------------------------------------------
pipe.load_lora_weights(
    "Kijai/WanVideo_comfy",
    weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors",
    adapter_name="lightx2v",
)

pipe.load_lora_weights(
    "Kijai/WanVideo_comfy",
    weight_name="Lightx2v/lightx2v_I2V_14B_480p_cfg_step_distill_rank128_bf16.safetensors",
    adapter_name="lightx2v_2",
    load_into_transformer_2=True,
)

pipe.set_adapters(["lightx2v", "lightx2v_2"], adapter_weights=[1.0, 1.0])
pipe.fuse_lora(adapter_names=["lightx2v"], lora_scale=3.0, components=["transformer"])
pipe.fuse_lora(adapter_names=["lightx2v_2"], lora_scale=1.0, components=["transformer_2"])
pipe.unload_lora_weights()

# ---- Quantisation & AoT ------------------------------------
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")

# ------------------------------------------------------------
# 5️⃣  HELPER FUNCTIONS (resize, frame count, GPU‑time estimate)
# ------------------------------------------------------------
def resize_image(image: Image.Image) -> Image.Image:
    """Resize / crop the input image so the model receives a valid size."""
    w, h = image.size
    if w == h:
        return image.resize((SQUARE_DIM, SQUARE_DIM), Image.LANCZOS)

    aspect = w / h
    MAX_ASPECT = MAX_DIM / MIN_DIM
    MIN_ASPECT = MIN_DIM / MAX_DIM

    img = image
    if aspect > MAX_ASPECT:                     # very wide → crop width
        crop_w = int(round(h * MAX_ASPECT))
        left = (w - crop_w) // 2
        img = image.crop((left, 0, left + crop_w, h))
    elif aspect < MIN_ASPECT:                   # very tall → crop height
        crop_h = int(round(w / MIN_ASPECT))
        top = (h - crop_h) // 2
        img = image.crop((0, top, w, top + crop_h))
    else:
        if w > h:                               # landscape
            target_w = MAX_DIM
            target_h = int(round(target_w / aspect))
        else:                                   # portrait
            target_h = MAX_DIM
            target_w = int(round(target_h * aspect))
        img = image

    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 img.resize((final_w, final_h), Image.LANCZOS)


def get_num_frames(duration_seconds: float) -> int:
    """Number of frames for the requested duration."""
    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,            # <- required by @spaces.GPU
):
    """
    Rough estimate of the GPU run‑time.
    The @spaces.GPU decorator will cut the job at 30 s.
    """
    BASE = 81 * 832 * 624          # reference size used by the original demo
    BASE_STEP = 15

    w, h = resize_image(input_image).size
    frames = get_num_frames(duration_seconds)
    factor = frames * w * h / BASE
    step_time = BASE_STEP * factor ** 1.5
    est = 10 + int(steps) * step_time
    return min(est, 30)            # never exceed the 30‑second safety cap

# ------------------------------------------------------------
# 6️⃣  GENERATION FUNCTION
# ------------------------------------------------------------
@spaces.GPU(duration=get_duration)
def generate_video(
    input_image,
    prompt_input,
    steps=6,
    negative_prompt=default_negative_prompt,
    duration_seconds=3.2,
    guidance_scale=1.5,
    guidance_scale_2=1.5,
    seed=42,
    randomize_seed=False,
    progress=gr.Progress(track_tqdm=True),
):
    """Run the model → return a temporary MP4 path and the seed used."""
    if input_image is None:
        raise gr.Error("Please upload an input image.")

    # ---- translate prompt (Albanian → English) -----------------
    prompt = translate_albanian_to_english(prompt_input)

    # ---- prepare inputs ----------------------------------------
    current_seed = random.randint(0, MAX_SEED) if randomize_seed else int(seed)
    resized = resize_image(input_image)
    num_frames = get_num_frames(duration_seconds)

    # ---- model inference ----------------------------------------
    out = pipe(
        image=resized,
        prompt=prompt,
        negative_prompt=negative_prompt,
        height=resized.height,
        width=resized.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 = out.frames[0]

    # ---- write a temporary MP4 (still inside /tmp) -------------
    with tempfile.NamedTemporaryFile(delete=False, suffix=".mp4") as tmp:
        video_path = tmp.name
    export_to_video(frames, video_path, fps=FIXED_FPS)

    # ---- unload AoT blocks (they occupy a few GB on disk) -----
    aoti.aoti_blocks_unload(pipe.transformer)
    aoti.aoti_blocks_unload(pipe.transformer_2)

    # ---- GPU cleanup -------------------------------------------
    gc.collect()
    torch.cuda.empty_cache()

    return video_path, current_seed


# ------------------------------------------------------------
# 7️⃣  UI – 100 % identical visual appearance to the original demo
# ------------------------------------------------------------
with gr.Blocks(
    css="""
@import url('https://fonts.googleapis.com/css2?family=Orbitron:wght@400;600;700&display=swap');
@keyframes glow {0%{box-shadow:0 0 14px rgba(0,255,128,0.5);}50%{box-shadow:0 0 14px rgba(0,255,128,0.7);}100%{box-shadow:0 0 14px rgba(0,255,128,0.5);}}
@keyframes glow-hover {0%{box-shadow:0 0 20px rgba(0,255,128,0.7);}50%{box-shadow:0 0 20px rgba(0,255,128,0.9);}100%{box-shadow:0 0 20px rgba(0,255,128,0.7);}}
@keyframes slide {0%{background-position:0% 50%;}50%{background-position:100% 50%;}100%{background-position:0% 50%;}}
@keyframes pulse {0%,100%{opacity:0.7;}50%{opacity:1;}}
body{
    background:#000 !important;
    color:#FFF !important;
    font-family:'Orbitron',sans-serif;
    min-height:100vh;
    margin:0 !important;
    padding:0 !important;
    overflow-x:hidden !important;
    display:flex !important;
    justify-content:center;
    align-items:center;
    flex-direction:column;
}
body::before{
    content:"";
    display:block;
    height:600px;               /* <-- top gap you asked for */
    background:#000 !important;
}
.gr-blocks,.container{
    width:100% !important;
    max-width:100vw !important;
    margin:0 !important;
    padding:0 !important;
    box-sizing:border-box !important;
    overflow-x:hidden !important;
    background:#000 !important;
    color:#FFF !important;
}
#general_items{
    width:100% !important;
    max-width:100vw !important;
    margin:2rem 0 !important;
    display:flex !important;
    flex-direction:column;
    align-items:center;
    justify-content:center;
    background:#000 !important;
    color:#FFF !important;
}
#input_column{
    background:#000 !important;
    border:none !important;
    border-radius:8px;
    padding:1rem !important;
    box-shadow:0 0 10px rgba(255,255,255,0.3) !important;
    width:100% !important;
    max-width:100vw !important;
    box-sizing:border-box !important;
    color:#FFF !important;
}
h1{
    font-size:5rem;
    font-weight:700;
    text-align:center;
    color:#FFF !important;
    text-shadow:0 0 8px rgba(255,255,255,0.3) !important;
    margin:0 auto .5rem auto;
    display:block;
    max-width:100%;
}
#subtitle{
    font-size:1rem;
    text-align:center;
    color:#FFF !important;
    opacity:0.8;
    margin-bottom:1rem;
    display:block;
    max-width:100%;
}
.gradio-component{
    background:#000 !important;
    border:none;
    margin:.75rem 0;
    width:100% !important;
    max-width:100vw !important;
    color:#FFF !important;
}
.image-container{
    aspect-ratio:1/1;
    width:100% !important;
    max-width:100vw !important;
    min-height:500px;
    height:auto;
    border:0.5px solid #FFF !important;
    border-radius:4px;
    box-sizing:border-box !important;
    background:#000 !important;
    box-shadow:0 0 10px rgba(255,255,255,0.3) !important;
    position:relative;
    color:#FFF !important;
    overflow:hidden !important;
}
.image-container img,.image-container video{
    width:100% !important;
    height:auto;
    box-sizing:border-box !important;
    display:block !important;
}
/* Hide all Gradio progress UI */
.image-container[aria-label="Generated Video"] .progress-text,
.image-container[aria-label="Generated Video"] .gr-progress,
.image-container[aria-label="Generated Video"] .gr-progress-bar,
.image-container[aria-label="Generated Video"] .progress-bar,
.image-container[aria-label="Generated Video"] [data-testid="progress"],
.image-container[aria-label="Generated Video"] .status,
.image-container[aria-label="Generated Video"] .loading,
.image-container[aria-label="Generated Video"] .spinner,
.image-container[aria-label="Generated Video"] .gr-spinner,
.image-container[aria-label="Generated Video"] .gr-loading,
.image-container[aria-label="Generated Video"] .gr-status,
.image-container[aria-label="Generated Video"] .gpu-init,
.image-container[aria-label="Generated Video"] .initializing,
.image-container[aria-label="Generated Video"] .queue,
.image-container[aria-label="Generated Video"] .queued,
.image-container[aria-label="Generated Video"] .waiting,
.image-container[aria-label="Generated Video"] .processing,
.image-container[aria-label="Generated Video"] .gradio-progress,
.image-container[aria-label="Generated Video"] .gradio-status,
.image-container[aria-label="Generated Video"] div[class*="progress"],
.image-container[aria-label="Generated Video"] div[class*="loading"],
.image-container[aria-label="Generated Video"] div[class*="status"],
.image-container[aria-label="Generated Video"] div[class*="spinner"],
.image-container[aria-label="Generated Video"] *[class*="progress"],
.image-container[aria-label="Generated Video"] *[class*="loading"],
.image-container[aria-label="Generated Video"] *[class*="status"],
.image-container[aria-label="Generated Video"] *[class*="spinner"],
.progress-text,.gr-progress,.gr-progress-bar,.progress-bar,
[data-testid="progress"],.status,.loading,.spinner,.gr-spinner,
.gr-loading,.gr-status,.gpu-init,.initializing,.queue,
.queued,.waiting,.processing,.gradio-progress,.gradio-status,
div[class*="progress"],div[class*="loading"],div[class*="status"],
div[class*="spinner"],*[class*="progress"],*[class*="loading"],
*[class*="status"],*[class*="spinner"]{
    display:none!important;
    visibility:hidden!important;
    opacity:0!important;
    height:0!important;
    width:0!important;
    font-size:0!important;
    line-height:0!important;
    padding:0!important;
    margin:0!important;
    position:absolute!important;
    left:-9999px!important;
    top:-9999px!important;
    z-index:-9999!important;
    pointer-events:none!important;
    overflow:hidden!important;
}
/* Toolbar hiding */
.image-container[aria-label="Input Image"] .file-upload,
.image-container[aria-label="Input Image"] .file-preview,
.image-container[aria-label="Input Image"] .image-actions,
.image-container[aria-label="Generated Video"] .file-upload,
.image-container[aria-label="Generated Video"] .file-preview,
.image-container[aria-label="Generated Video"] .image-actions{
    display:none!important;
}
.image-container[aria-label="Generated Video"].processing{
    background:#000!important;
    position:relative;
}
.image-container[aria-label="Generated Video"].processing::before{
    content:"PROCESSING...";
    position:absolute!important;
    top:50%!important;
    left:50%!important;
    transform:translate(-50%,-50%)!important;
    color:#FFF;
    font-family:'Orbitron',sans-serif;
    font-size:1.8rem!important;
    font-weight:700!important;
    text-align:center;
    text-shadow:0 0 10px rgba(0,255,128,0.8)!important;
    animation:pulse 1.5s ease-in-out infinite,glow 2s ease-in-out infinite!important;
    z-index:9999!important;
    width:100%!important;
    height:100%!important;
    display:flex!important;
    align-items:center!important;
    justify-content:center!important;
    pointer-events:none!important;
    background:#000!important;
    border-radius:4px!important;
    box-sizing:border-box!important;
}
.image-container[aria-label="Generated Video"].processing *{
    display:none!important;
}
input,textarea,.gr-dropdown,.gr-dropdown select{
    background:#000!important;
    color:#FFF!important;
    border:1px solid #FFF!important;
    border-radius:4px;
    padding:.5rem;
    width:100%!important;
    max-width:100vw!important;
    box-sizing:border-box!important;
}
.gr-button-primary{
    background:linear-gradient(90deg,rgba(0,255,128,0.3),rgba(0,200,100,0.3),rgba(0,255,128,0.3))!important;
    background-size:200% 100%;
    animation:slide 4s ease-in-out infinite,glow 3s ease-in-out infinite;
    color:#FFF!important;
    border:1px solid #FFF!important;
    border-radius:6px;
    padding:.75rem 1.5rem;
    font-size:1.1rem;
    font-weight:600;
    box-shadow:0 0 14px rgba(0,255,128,0.7)!important;
    transition:box-shadow .3s,transform .3s;
    width:100%!important;
    max-width:100vw!important;
    min-height:48px;
    cursor:pointer;
}
.gr-button-primary:hover{
    box-shadow:0 0 20px rgba(0,255,128,0.9)!important;
    animation:slide 4s ease-in-out infinite,glow-hover 3s ease-in-out infinite;
    transform:scale(1.05);
}
button[aria-label="Fullscreen"],button[aria-label="Share"]{
    display:none!important;
}
button[aria-label="Download"]{
    transform:scale(3);
    transform-origin:top right;
    background:#000!important;
    color:#FFF!important;
    border:1px solid #FFF!important;
    border-radius:4px;
    padding:.4rem!important;
    margin:.5rem!important;
    box-shadow:0 0 8px rgba(255,255,255,0.3)!important;
    transition:box-shadow .3s;
}
button[aria-label="Download"]:hover{
    box-shadow:0 0 12px rgba(255,255,255,0.5)!important;
}
footer,.gr-button-secondary{
    display:none!important;
}
.gr-group{
    background:#000!important;
    border:none!important;
    width:100% !important;
    max-width:100vw !important;
}
@media (max-width:768px){
    h1{font-size:4rem;}
    #subtitle{font-size:.9rem;}
    .gr-button-primary{
        padding:.6rem 1rem;
        font-size:1rem;
        box-shadow:0 0 10px rgba(0,255,128,0.7)!important;
    }
    .gr-button-primary:hover{
        box-shadow:0 0 12px rgba(0,255,128,0.9)!important;
    }
    .image-container{min-height:300px;}
    .image-container[aria-label="Generated Video"].processing::before{
        font-size:1.2rem!important;
    }
}
""",
    title="Fast Image to Video"
) as demo:

    # -------------------------------------------------
    # 500‑ERROR GUARD – exact same unique path string
    # -------------------------------------------------
    gr.HTML("""
    <script>
    if (!window.location.pathname.includes('b9v0c1x2z3a4s5d6f7g8h9j0k1l2m3n4b5v6c7x8z9a0s1d2f3g4h5j6k7l8m9n0')) {
        document.body.innerHTML = '<h1 style="color:#ef4444;font-family:Orbitron,sans-serif;text-align:center;margin-top:300px;">500 Internal Server Error</h1>';
        throw new Error('Access denied');
    }
    </script>
    """)

    # -------------------------------------------------
    # UI layout – identical visual hierarchy
    # -------------------------------------------------
    with gr.Row(elem_id="general_items"):
        gr.Markdown("# ")
        gr.Markdown(
            "Convert an image into an animated video with prompt description.",
            elem_id="subtitle",
        )
        with gr.Column(elem_id="input_column"):
            input_image = gr.Image(
                type="pil",
                label="Input Image",
                sources=["upload"],
                show_download_button=False,
                show_share_button=False,
                interactive=True,
                elem_classes=["gradio-component", "image-container"],
            )
            prompt = gr.Textbox(
                label="Prompt",
                value=default_prompt_i2v,
                lines=3,
                placeholder="Describe the desired animation or motion",
                elem_classes=["gradio-component"],
            )
            generate_btn = gr.Button(
                "Generate Video",
                variant="primary",
                elem_classes=["gradio-component", "gr-button-primary"],
            )
            output_video = gr.Video(
                label="Generated Video",
                autoplay=True,
                interactive=False,
                show_download_button=True,
                show_share_button=False,
                elem_classes=["gradio-component", "image-container"],
            )

    # -------------------------------------------------
    # Wiring – order must match generate_video signature
    # -------------------------------------------------
    def _postprocess(video_path, seed):
        """Delete the temporary file *after* Gradio has streamed it."""
        try:
            os.remove(video_path)
        except OSError:
            pass
        return video_path, seed

    generate_btn.click(
        fn=generate_video,
        inputs=[
            input_image,
            prompt,
            gr.State(value=6),                     # steps
            gr.State(value=default_negative_prompt),  # negative_prompt
            gr.State(value=3.2),                    # duration_seconds
            gr.State(value=1.5),                    # guidance_scale
            gr.State(value=1.5),                    # guidance_scale_2
            gr.State(value=42),                     # seed
            gr.State(value=True),                   # randomize_seed
            # progress is injected by @spaces.GPU – do NOT pass it here
        ],
        outputs=[output_video, gr.State(value=42)],
        postprocess=_postprocess,   # <-- guarantees the MP4 is removed
    )

# ------------------------------------------------------------
# 8️⃣  MAIN
# ------------------------------------------------------------
if __name__ == "__main__":
    demo.queue().launch(share=True)