app.py

import os
import time
from dataclasses import dataclass
from functools import lru_cache

import gradio as gr
import numpy as np
from huggingface_hub import hf_hub_download
from openvino import Core
from PIL import Image, ImageOps

MODEL_REPO_ID = os.getenv("MODEL_REPO_ID", "ibrhr/BiRefNet-lite-openvino-xeon-w2145")
DEVICE = os.getenv("OPENVINO_DEVICE", "CPU")
DEFAULT_MODEL_VARIANT_KEY = os.getenv("MODEL_VARIANT", "fp32_1024x1024")

IMAGENET_MEAN = np.array([0.485, 0.456, 0.406], dtype=np.float32)
IMAGENET_STD = np.array([0.229, 0.224, 0.225], dtype=np.float32)


@dataclass(frozen=True)
class ModelVariant:
    key: str
    label: str
    xml: str
    precision: str
    input_size: int
    benchmark_ms: float
    benchmark_fps: float
    notes: str

    @property
    def bin(self) -> str:
        return self.xml.replace(".xml", ".bin")


MODEL_VARIANTS = (
    ModelVariant(
        key="int8_1024x1024",
        label="INT8 NNCF - 1024x1024 - 1272 ms / 0.79 FPS",
        xml="openvino_int8/birefnet_lite_1024x1024_int8.xml",
        precision="INT8 NNCF",
        input_size=1024,
        benchmark_ms=1272.2,
        benchmark_fps=0.79,
        notes="Best benchmarked full-quality option on the target CPU.",
    ),
    ModelVariant(
        key="int8_512x512",
        label="INT8 NNCF - 512x512 - 332 ms / 3.01 FPS",
        xml="openvino_int8/birefnet_lite_512x512_int8.xml",
        precision="INT8 NNCF",
        input_size=512,
        benchmark_ms=332.32,
        benchmark_fps=3.01,
        notes="Fastest benchmarked option, with lower input resolution.",
    ),
    ModelVariant(
        key="fp16_1024x1024",
        label="FP16 - 1024x1024 - 1419 ms / 0.70 FPS",
        xml="openvino_fp16/birefnet_lite_1024x1024_fp16.xml",
        precision="FP16",
        input_size=1024,
        benchmark_ms=1419.0,
        benchmark_fps=0.70,
        notes="Smaller weights than FP32 at full input resolution.",
    ),
    ModelVariant(
        key="fp16_512x512",
        label="FP16 - 512x512 - 366 ms / 2.73 FPS",
        xml="openvino_fp16/birefnet_lite_512x512_fp16.xml",
        precision="FP16",
        input_size=512,
        benchmark_ms=365.97,
        benchmark_fps=2.73,
        notes="Smaller weights than FP32 at lower input resolution.",
    ),
    ModelVariant(
        key="fp32_1024x1024",
        label="FP32 - 1024x1024 - 1441 ms / 0.69 FPS",
        xml="openvino_fp32/birefnet_lite_1024x1024.xml",
        precision="FP32",
        input_size=1024,
        benchmark_ms=1440.9,
        benchmark_fps=0.69,
        notes="Original default and reference OpenVINO precision.",
    ),
    ModelVariant(
        key="fp32_512x512",
        label="FP32 - 512x512 - 366 ms / 2.73 FPS",
        xml="openvino_fp32/birefnet_lite_512x512.xml",
        precision="FP32",
        input_size=512,
        benchmark_ms=366.46,
        benchmark_fps=2.73,
        notes="Reference OpenVINO precision at lower input resolution.",
    ),
    ModelVariant(
        key="int8wo_1024x1024",
        label="INT8 weight-only - 1024x1024 - 1440 ms / 0.69 FPS",
        xml="openvino_int8wo/birefnet_lite_1024x1024_int8wo.xml",
        precision="INT8 weight-only",
        input_size=1024,
        benchmark_ms=1439.53,
        benchmark_fps=0.69,
        notes="Alternative weight-only quantized full-resolution model.",
    ),
    ModelVariant(
        key="int8wo_512x512",
        label="INT8 weight-only - 512x512 - 366 ms / 2.73 FPS",
        xml="openvino_int8wo/birefnet_lite_512x512_int8wo.xml",
        precision="INT8 weight-only",
        input_size=512,
        benchmark_ms=365.75,
        benchmark_fps=2.73,
        notes="Alternative weight-only quantized lower-resolution model.",
    ),
)
MODEL_VARIANTS_BY_KEY = {variant.key: variant for variant in MODEL_VARIANTS}


@dataclass(frozen=True)
class Runtime:
    compiled_model: object
    input_node: object
    output_node: object
    variant: ModelVariant
    model_path: str
    load_seconds: float
    device: str


def get_model_variant(variant_key: str | None) -> ModelVariant:
    key = variant_key or DEFAULT_MODEL_VARIANT_KEY
    if key not in MODEL_VARIANTS_BY_KEY:
        valid_keys = ", ".join(MODEL_VARIANTS_BY_KEY)
        raise gr.Error(f"Unknown model variant '{key}'. Valid variants: {valid_keys}")
    return MODEL_VARIANTS_BY_KEY[key]


def _resampling(name: str) -> int:
    return getattr(Image.Resampling, name)


@lru_cache(maxsize=len(MODEL_VARIANTS))
def get_runtime(variant_key: str) -> Runtime:
    variant = get_model_variant(variant_key)
    started = time.perf_counter()
    model_path = hf_hub_download(repo_id=MODEL_REPO_ID, filename=variant.xml)
    weights_path = hf_hub_download(repo_id=MODEL_REPO_ID, filename=variant.bin)

    core = Core()
    model = core.read_model(model=model_path, weights=weights_path)
    model.reshape({model.input(0): [1, 3, variant.input_size, variant.input_size]})
    compiled_model = core.compile_model(model, DEVICE)

    return Runtime(
        compiled_model=compiled_model,
        input_node=compiled_model.input(0),
        output_node=compiled_model.output(0),
        variant=variant,
        model_path=model_path,
        load_seconds=time.perf_counter() - started,
        device=DEVICE,
    )


def preprocess(image: Image.Image, model_size: int) -> np.ndarray:
    rgb_image = ImageOps.exif_transpose(image).convert("RGB")
    resized = rgb_image.resize((model_size, model_size), _resampling("BICUBIC"))
    array = np.asarray(resized, dtype=np.float32) / 255.0
    array = (array - IMAGENET_MEAN) / IMAGENET_STD
    array = np.transpose(array, (2, 0, 1))[None, ...]
    return np.ascontiguousarray(array, dtype=np.float32)


def sigmoid(array: np.ndarray) -> np.ndarray:
    clipped = np.clip(array, -50.0, 50.0)
    return 1.0 / (1.0 + np.exp(-clipped))


def postprocess_mask(output: np.ndarray, size: tuple[int, int]) -> Image.Image:
    mask = np.asarray(output, dtype=np.float32)
    while mask.ndim > 2:
        mask = mask[0]

    mask = sigmoid(mask)
    mask = np.clip(mask * 255.0, 0, 255).astype(np.uint8)
    mask_image = Image.fromarray(mask, mode="L")
    return mask_image.resize(size, _resampling("LANCZOS"))


def remove_background(image: Image.Image, model_variant_key: str):
    if image is None:
        raise gr.Error("Upload an image first.")

    total_started = time.perf_counter()
    variant = get_model_variant(model_variant_key)
    runtime = get_runtime(variant.key)
    original = ImageOps.exif_transpose(image).convert("RGB")

    preprocess_started = time.perf_counter()
    tensor = preprocess(original, variant.input_size)
    preprocess_seconds = time.perf_counter() - preprocess_started

    inference_started = time.perf_counter()
    output = runtime.compiled_model({runtime.input_node: tensor})[runtime.output_node]
    inference_seconds = time.perf_counter() - inference_started

    postprocess_started = time.perf_counter()
    mask_image = postprocess_mask(output, original.size)
    cutout = original.convert("RGBA")
    cutout.putalpha(mask_image)
    postprocess_seconds = time.perf_counter() - postprocess_started

    total_seconds = time.perf_counter() - total_started
    timing = (
        f"Total: {total_seconds:.3f} s\n"
        f"Preprocess: {preprocess_seconds * 1000:.1f} ms\n"
        f"Inference: {inference_seconds * 1000:.1f} ms\n"
        f"Postprocess: {postprocess_seconds * 1000:.1f} ms"
    )

    specs = {
        "model": MODEL_REPO_ID,
        "variant": variant.key,
        "variant_label": variant.label,
        "model_xml": variant.xml,
        "device": runtime.device,
        "precision": variant.precision,
        "model_input_size": f"{variant.input_size}x{variant.input_size}",
        "benchmark_ms": variant.benchmark_ms,
        "benchmark_fps": variant.benchmark_fps,
        "variant_notes": variant.notes,
        "uploaded_image_size": f"{original.width}x{original.height}",
        "input_tensor_shape": list(tensor.shape),
        "output_tensor_shape": list(np.asarray(output).shape),
        "model_load_seconds": round(runtime.load_seconds, 3),
        "total_seconds": round(total_seconds, 3),
        "preprocess_ms": round(preprocess_seconds * 1000, 1),
        "inference_ms": round(inference_seconds * 1000, 1),
        "postprocess_ms": round(postprocess_seconds * 1000, 1),
    }
    return mask_image, cutout, timing, specs


with gr.Blocks(title="BiRefNet OpenVINO") as demo:
    gr.Markdown("# BiRefNet OpenVINO")
    with gr.Row():
        input_image = gr.Image(label="Image", type="pil")
        model_dropdown = gr.Dropdown(
            label="Model variant",
            choices=[(variant.label, variant.key) for variant in MODEL_VARIANTS],
            value=get_model_variant(DEFAULT_MODEL_VARIANT_KEY).key,
            interactive=True,
        )
    run_button = gr.Button("Run", variant="primary")
    with gr.Row():
        mask_output = gr.Image(label="Mask", type="pil")
        cutout_output = gr.Image(label="Background removed", type="pil", format="png")
    with gr.Row():
        timing_output = gr.Textbox(label="Processing time", lines=4)
        specs_output = gr.JSON(label="Specs")

    run_button.click(
        fn=remove_background,
        inputs=[input_image, model_dropdown],
        outputs=[mask_output, cutout_output, timing_output, specs_output],
    )
    input_image.upload(
        fn=remove_background,
        inputs=[input_image, model_dropdown],
        outputs=[mask_output, cutout_output, timing_output, specs_output],
    )


if __name__ == "__main__":
    demo.queue(max_size=8).launch()