app_2511.py

# app_2511.py
import random
from typing import Optional, Tuple

import gradio as gr
import numpy as np
import torch
from PIL import Image

from diffusers import QwenImageEditPlusPipeline

# -----------------------------------------------------------------------------
# Optional: Hugging Face Spaces helpers (chỉ có trên Spaces runtime)
# -----------------------------------------------------------------------------
try:
    import spaces  # type: ignore
except Exception:
    spaces = None

# -----------------------------------------------------------------------------
# Global config
# -----------------------------------------------------------------------------
dtype = torch.bfloat16
device = "cuda" if torch.cuda.is_available() else "cpu"
MAX_SEED = np.iinfo(np.int32).max

_pipe: Optional[QwenImageEditPlusPipeline] = None


def get_pipe() -> QwenImageEditPlusPipeline:
    """
    Load Qwen-Image-Edit-2511 + fal Multiple Angles LoRA as a global singleton.
    """
    global _pipe
    if _pipe is not None:
        return _pipe

    pipe = QwenImageEditPlusPipeline.from_pretrained(
        "Qwen/Qwen-Image-Edit-2511",
        torch_dtype=dtype,
    )

    if device == "cuda":
        pipe = pipe.to("cuda")

    # Load LoRA: fal/Qwen-Image-Edit-2511-Multiple-Angles-LoRA
    pipe.load_lora_weights(
        "fal/Qwen-Image-Edit-2511-Multiple-Angles-LoRA",
        weight_name="qwen-image-edit-2511-multiple-angles-lora.safetensors",
        adapter_name="angles",
    )
    pipe.set_adapters(["angles"], adapter_weights=[1.0])

    # Fuse LoRA for speed. Strength recommended around 0.8–1.0, default 0.9.
    pipe.fuse_lora(adapter_names=["angles"], lora_scale=0.9)
    pipe.unload_lora_weights()

    # Optional AOTI (Spaces only) - keep same pattern; adjust repo if you have a 2511-specific AOTI artifact.
    if spaces is not None and hasattr(spaces, "aoti_blocks_load"):
        try:
            spaces.aoti_blocks_load(pipe.transformer, "zerogpu-aoti/Qwen-Image", variant="fa3")
            print("[INFO] spaces.aoti_blocks_load enabled")
        except Exception as e:
            print(f"[WARN] spaces.aoti_blocks_load failed: {e}")

    _pipe = pipe
    return _pipe


# -----------------------------------------------------------------------------
# Angle prompt (fal schema)
# -----------------------------------------------------------------------------
AZIMUTH_CHOICES = [
    "front view",
    "front-right quarter view",
    "right side view",
    "back-right quarter view",
    "back view",
    "back-left quarter view",
    "left side view",
    "front-left quarter view",
]

ELEVATION_CHOICES = [
    "low-angle shot",
    "eye-level shot",
    "elevated shot",
    "high-angle shot",
]

DISTANCE_CHOICES = [
    "close-up",
    "medium shot",
    "wide shot",
]


def build_fal_angles_prompt(azimuth: str, elevation: str, distance: str) -> str:
    # fal requires: <sks> [azimuth] [elevation] [distance]
    return f"<sks> {azimuth} {elevation} {distance}"


# -----------------------------------------------------------------------------
# Inference
# -----------------------------------------------------------------------------
def infer_angles_edit(
    image: Optional[Image.Image],
    azimuth: str,
    elevation: str,
    distance: str,
    seed: int,
    randomize_seed: bool,
    true_guidance_scale: float,
    num_inference_steps: int,
    height: int,
    width: int,
) -> Tuple[Image.Image, int, str]:
    if image is None:
        raise gr.Error("Please upload an image.")

    if randomize_seed:
        seed = random.randint(0, MAX_SEED)

    prompt = build_fal_angles_prompt(azimuth, elevation, distance)
    print("Prompt:", prompt)

    pipe = get_pipe()
    generator = torch.Generator(device=device).manual_seed(int(seed))

    out = pipe(
        image=[image.convert("RGB")],
        prompt=prompt,
        height=height if height and height > 0 else None,
        width=width if width and width > 0 else None,
        num_inference_steps=int(num_inference_steps),
        true_cfg_scale=float(true_guidance_scale),
        guidance_scale=1.0,
        negative_prompt=" ",
        generator=generator,
        num_images_per_prompt=1,
    ).images[0]

    return out, seed, prompt


# -----------------------------------------------------------------------------
# UI (Gradio 6.3.0 compatible)
# -----------------------------------------------------------------------------
CAMERA_3D_HTML_TEMPLATE = """
<div style="width:100%;height:400px;background:#1a1a1a;border-radius:12px;
display:flex;align-items:center;justify-content:center;color:#aaa;">
  <div style="text-align:center;">
    <div style="font-size:18px;color:#00ff88;font-weight:700;">3D Camera Preview</div>
    <div style="margin-top:8px;">(Gradio 6) JS ↔ Python sync removed</div>
    <div style="margin-top:6px;">Use dropdowns to control camera</div>
  </div>
</div>
"""

css = """
#camera-3d-control { min-height: 400px; }
"""

with gr.Blocks(css=css) as demo:
    gr.Markdown("## 🎬 Qwen Image Edit — 2511 + Multiple Angles LoRA (UI only)")

    with gr.Row():
        with gr.Column(scale=1):
            image = gr.Image(label="Input Image", type="pil", height=280)

            with gr.Tab("🎮 3D Camera Preview"):
                _ = gr.HTML(value=CAMERA_3D_HTML_TEMPLATE, elem_id="camera-3d-control")

            with gr.Tab("🎚️ Angles (fal schema)"):
                azimuth = gr.Dropdown(
                    choices=AZIMUTH_CHOICES,
                    value="front view",
                    label="Azimuth (8 directions)",
                )
                elevation = gr.Dropdown(
                    choices=ELEVATION_CHOICES,
                    value="eye-level shot",
                    label="Elevation (4 levels)",
                )
                distance = gr.Dropdown(
                    choices=DISTANCE_CHOICES,
                    value="medium shot",
                    label="Distance (3 levels)",
                )

            run_btn = gr.Button("🚀 Generate", variant="primary")

        with gr.Column(scale=1):
            result = gr.Image(label="Output Image", height=350)
            prompt_preview = gr.Textbox(label="Generated Prompt", interactive=False)

            with gr.Accordion("⚙️ Advanced", open=False):
                seed = gr.Slider(0, MAX_SEED, step=1, label="Seed", value=0)
                randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)

                # Thường 2511 edit chạy ổn với steps 8–16, true_cfg_scale khoảng 3–6 (tùy ảnh).
                true_guidance_scale = gr.Slider(1.0, 10.0, step=0.1, label="True CFG Scale", value=4.0)
                num_inference_steps = gr.Slider(1, 40, step=1, label="Steps", value=12)

                height = gr.Slider(256, 2048, step=8, value=1024, label="Height")
                width = gr.Slider(256, 2048, step=8, value=1024, label="Width")

            # (Optional) Quick tips
            gr.Markdown(
                "- Prompt format used: `<sks> azimuth elevation distance`\n"
                "- Nếu ảnh bị méo: giảm **True CFG Scale** hoặc tăng **Steps** nhẹ.\n"
                "- Nếu đổi góc chưa mạnh: tăng **True CFG Scale** một chút (không cần tăng LoRA strength)."
            )

    run_btn.click(
        fn=infer_angles_edit,
        inputs=[
            image,
            azimuth,
            elevation,
            distance,
            seed,
            randomize_seed,
            true_guidance_scale,
            num_inference_steps,
            height,
            width,
        ],
        outputs=[result, seed, prompt_preview],
        # chưa cần api -> không set api_name
    )

if __name__ == "__main__":
    demo.queue(max_size=16)
    demo.launch(
        server_name="127.0.0.1",
        server_port=7861,
        share=False,
        debug=True,
        show_error=True,
        prevent_thread_lock=True,
    )

    # giữ process sống nếu chạy bằng python app_2511.py
    import time as _time
    while True:
        _time.sleep(3600)