import argparse
import gradio as gr
import json
import os
import tempfile
from typing import Any

import cv2
import numpy as np
import torch
from PIL import Image, ImageOps
from sklearn.decomposition import PCA
from stl import mesh
from transformers import AutoModelForDepthEstimation, AutoProcessor

try:
    import pillow_heif

    pillow_heif.register_heif_opener()
    HEIF_SUPPORT = True
except Exception as heif_error:  # pragma: no cover - surfaced in UI if HEIC open fails
    pillow_heif = None
    HEIF_SUPPORT = False
    HEIF_IMPORT_ERROR = heif_error
else:
    HEIF_IMPORT_ERROR = None

MIN_RESOLUTION_LIMIT = 10_000
DEFAULT_MAX_RESOLUTION = int(os.getenv("DEFAULT_MAX_RESOLUTION", "1500000"))
MAX_RESOLUTION_LIMIT = int(os.getenv("MAX_RESOLUTION_LIMIT", "20000000"))


def parse_launch_args() -> argparse.Namespace:
    parser = argparse.ArgumentParser(description="Image to 3D Relief Gradio app")
    parser.add_argument(
        "--max-resolution-pixels",
        type=int,
        default=None,
        help=(
            "Maximum number of pixels processed for depth estimation. "
            "Default comes from DEFAULT_MAX_RESOLUTION or 1,500,000."
        ),
    )
    parser.add_argument(
        "--server-name",
        default=os.getenv("GRADIO_SERVER_NAME", "0.0.0.0"),
        help="Host/interface for Gradio to bind.",
    )
    parser.add_argument(
        "--server-port",
        type=int,
        default=int(os.getenv("PORT", os.getenv("GRADIO_SERVER_PORT", "7860"))),
        help="Port for Gradio to bind.",
    )
    args, _ = parser.parse_known_args()
    return args


LAUNCH_ARGS = parse_launch_args()
REQUESTED_MAX_RESOLUTION = LAUNCH_ARGS.max_resolution_pixels or DEFAULT_MAX_RESOLUTION
CONFIGURED_MAX_RESOLUTION = int(max(
    MIN_RESOLUTION_LIMIT,
    min(MAX_RESOLUTION_LIMIT, REQUESTED_MAX_RESOLUTION),
))
MODEL_ID = "depth-anything/Depth-Anything-V2-Large-hf"

DEFAULT_TEXTURE_STRENGTH = 0.1
DEFAULT_TEXTURE_SMOOTHING = 5
DEFAULT_MIN_Z = 0.5
DEFAULT_MAX_Z = 5.0
DEFAULT_CURVE_POINTS = [
    {"x": 0.0, "y": 0.0},
    {"x": 1.0, "y": 1.0},
]
ACCEPTED_IMAGE_TYPES = [
    ".png",
    ".jpg",
    ".jpeg",
    ".webp",
    ".bmp",
    ".tif",
    ".tiff",
    ".heic",
    ".heif",
]


device = "cuda" if torch.cuda.is_available() else "cpu"
processor = AutoProcessor.from_pretrained(MODEL_ID, use_fast=True)
model = AutoModelForDepthEstimation.from_pretrained(MODEL_ID).to(device)
model.eval()
print(f"Model loaded successfully on {device}.")


APP_CSS = """
#curve-points-json, #depth-histogram-json, #auto-preview-depth-btn {
    display: none !important;
}
.depth-editor-card {
    border: 1px solid var(--border-color-primary);
    border-radius: 12px;
    padding: 14px;
    background: var(--background-fill-secondary);
}
.depth-editor-toolbar {
    display: flex;
    flex-wrap: wrap;
    gap: 8px;
    margin: 8px 0 10px;
}
.depth-editor-toolbar button {
    border: 1px solid var(--border-color-primary);
    border-radius: 8px;
    padding: 6px 10px;
    background: var(--button-secondary-background-fill);
    color: var(--body-text-color);
    cursor: pointer;
}
.depth-editor-toolbar button:hover {
    background: var(--button-secondary-background-fill-hover);
}
#depth-curve-canvas {
    width: 100%;
    max-width: 720px;
    height: 360px;
    display: block;
    touch-action: none;
    border: 1px solid var(--border-color-primary);
    border-radius: 8px;
    background: #111827;
}
.depth-editor-help {
    margin-top: 8px;
    color: var(--body-text-color-subdued);
    font-size: 0.9rem;
    line-height: 1.35;
}
.depth-editor-readout {
    min-height: 1.2rem;
    margin-top: 6px;
    color: var(--body-text-color-subdued);
    font-family: var(--font-mono);
    font-size: 0.85rem;
}
"""


CURVE_EDITOR_HEAD = """
<script>
(() => {
  const DEFAULT_POINTS = [{x: 0, y: 0}, {x: 1, y: 1}];
  const PRESETS = {
    linear: DEFAULT_POINTS,
    subject: [{x: 0, y: 0}, {x: 0.22, y: 0.13}, {x: 0.50, y: 0.50}, {x: 0.78, y: 0.88}, {x: 1, y: 1}],
    compressHigh: [{x: 0, y: 0}, {x: 0.45, y: 0.52}, {x: 0.75, y: 0.74}, {x: 1, y: 0.86}],
    compressLow: [{x: 0, y: 0.14}, {x: 0.25, y: 0.26}, {x: 0.55, y: 0.50}, {x: 1, y: 1}],
    foregroundPop: [{x: 0, y: 0}, {x: 0.35, y: 0.22}, {x: 0.62, y: 0.70}, {x: 1, y: 1}],
    invert: [{x: 0, y: 1}, {x: 1, y: 0}]
  };

  const clamp = (v, lo = 0, hi = 1) => Math.max(lo, Math.min(hi, v));

  function getTextarea(elemId) {
    const root = document.getElementById(elemId);
    if (!root) return null;
    return root.querySelector('textarea, input');
  }

  function setTextareaValue(elemId, value) {
    const el = getTextarea(elemId);
    if (!el) return false;
    const proto = el.tagName === 'TEXTAREA' ? window.HTMLTextAreaElement.prototype : window.HTMLInputElement.prototype;
    const setter = Object.getOwnPropertyDescriptor(proto, 'value')?.set;
    if (setter) setter.call(el, value);
    else el.value = value;
    el.dispatchEvent(new Event('input', { bubbles: true }));
    el.dispatchEvent(new Event('change', { bubbles: true }));
    return true;
  }

  function readTextareaValue(elemId) {
    const el = getTextarea(elemId);
    return el ? el.value : '';
  }

  function clickGradioButton(elemId) {
    const root = document.getElementById(elemId);
    const button = root?.querySelector('button');
    if (button) button.click();
  }

  function normalizePointMode(rawMode) {
    return String(rawMode || 'smooth').toLowerCase() === 'corner' ? 'corner' : 'smooth';
  }

  function normalizePoints(points) {
    if (!Array.isArray(points)) return normalizePoints(DEFAULT_POINTS);
    const normalized = [];
    for (const point of points) {
      const x = typeof point?.x === 'number' ? point.x : Array.isArray(point) ? point[0] : NaN;
      const y = typeof point?.y === 'number' ? point.y : Array.isArray(point) ? point[1] : NaN;
      const mode = Array.isArray(point) ? 'smooth' : normalizePointMode(point?.mode);
      if (Number.isFinite(x) && Number.isFinite(y)) normalized.push({ x: clamp(x), y: clamp(y), mode });
    }
    if (!normalized.length) normalized.push(...DEFAULT_POINTS.map(point => ({ ...point, mode: 'smooth' })));
    normalized.sort((a, b) => a.x - b.x);
    normalized[0].x = 0;
    normalized[0].mode = 'smooth';
    normalized[normalized.length - 1].x = 1;
    normalized[normalized.length - 1].mode = 'smooth';
    return normalized;
  }

  function parseJson(value, fallback) {
    try { return JSON.parse(value); } catch (_) { return fallback; }
  }

  function initCurveEditor() {
    const canvas = document.getElementById('depth-curve-canvas');
    if (!canvas || canvas.dataset.depthCurveReady === '1') return;
    canvas.dataset.depthCurveReady = '1';
    const ctx = canvas.getContext('2d');
    const readout = document.getElementById('depth-curve-readout');
    const state = {
      points: normalizePoints(DEFAULT_POINTS),
      histogram: [],
      dragging: -1,
      selected: -1,
      pressedSelected: -1,
      pointerMoved: false,
      pointerStart: null,
      lastCurveJson: '',
      lastHistogramJson: '',
      hasDepth: false,
      refreshTimer: null
    };

    function resizeCanvas() {
      const rect = canvas.getBoundingClientRect();
      const dpr = window.devicePixelRatio || 1;
      const width = Math.max(320, Math.round(rect.width * dpr));
      const height = Math.max(240, Math.round(rect.height * dpr));
      if (canvas.width !== width || canvas.height !== height) {
        canvas.width = width;
        canvas.height = height;
        return true;
      }
      return false;
    }

    function graphRect() {
      const pad = 34 * (window.devicePixelRatio || 1);
      return { x: pad, y: pad * 0.65, w: canvas.width - pad * 1.45, h: canvas.height - pad * 1.5 };
    }

    function toCanvas(point) {
      const r = graphRect();
      return { x: r.x + point.x * r.w, y: r.y + (1 - point.y) * r.h };
    }

    function fromCanvas(x, y) {
      const r = graphRect();
      return { x: clamp((x - r.x) / r.w), y: clamp(1 - ((y - r.y) / r.h)) };
    }

    function drawGrid(r) {
      ctx.save();
      ctx.strokeStyle = 'rgba(255,255,255,0.13)';
      ctx.lineWidth = 1;
      for (let i = 0; i <= 4; i++) {
        const x = r.x + (r.w * i / 4);
        const y = r.y + (r.h * i / 4);
        ctx.beginPath(); ctx.moveTo(x, r.y); ctx.lineTo(x, r.y + r.h); ctx.stroke();
        ctx.beginPath(); ctx.moveTo(r.x, y); ctx.lineTo(r.x + r.w, y); ctx.stroke();
      }
      ctx.strokeStyle = 'rgba(255,255,255,0.5)';
      ctx.strokeRect(r.x, r.y, r.w, r.h);
      ctx.fillStyle = 'rgba(255,255,255,0.72)';
      ctx.font = `${12 * (window.devicePixelRatio || 1)}px ui-monospace, SFMono-Regular, Menlo, monospace`;
      ctx.fillText('input height value →', r.x + r.w - 145 * (window.devicePixelRatio || 1), r.y + r.h + 24 * (window.devicePixelRatio || 1));
      ctx.save();
      ctx.translate(r.x - 24 * (window.devicePixelRatio || 1), r.y + 95 * (window.devicePixelRatio || 1));
      ctx.rotate(-Math.PI / 2);
      ctx.fillText('output height', 0, 0);
      ctx.restore();
      ctx.restore();
    }

    function drawHistogram(r) {
      if (!state.histogram?.length) return;
      const maxBin = Math.max(...state.histogram, 1);
      const binW = r.w / state.histogram.length;
      ctx.save();
      ctx.fillStyle = 'rgba(96, 165, 250, 0.35)';
      state.histogram.forEach((count, i) => {
        const h = (count / maxBin) * r.h;
        ctx.fillRect(r.x + i * binW, r.y + r.h - h, Math.max(1, binW), h);
      });
      ctx.restore();
    }

    function getPchipEndpointSlope(h0, h1, d0, d1) {
      const slope = ((2 * h0 + h1) * d0 - h0 * d1) / Math.max(h0 + h1, 1e-6);
      if (slope === 0 || d0 === 0 || Math.sign(slope) !== Math.sign(d0)) return 0;
      if (Math.sign(d0) !== Math.sign(d1) && Math.abs(slope) > Math.abs(3 * d0)) return 3 * d0;
      return slope;
    }

    function getBezierCurveData(points) {
      if (points.length < 2) return null;
      const xs = points.map(point => point.x);
      const ys = points.map(point => point.y);
      const intervalCount = points.length - 1;
      const h = new Array(intervalCount);
      const secants = new Array(intervalCount);
      for (let i = 0; i < intervalCount; i++) {
        h[i] = Math.max(xs[i + 1] - xs[i], 1e-6);
        secants[i] = (ys[i + 1] - ys[i]) / h[i];
      }
      const shared = new Array(points.length).fill(0);
      if (points.length === 2) {
        shared[0] = secants[0];
        shared[1] = secants[0];
      } else {
        shared[0] = getPchipEndpointSlope(h[0], h[1], secants[0], secants[1]);
        shared[points.length - 1] = getPchipEndpointSlope(h[h.length - 1], h[h.length - 2], secants[secants.length - 1], secants[secants.length - 2]);
        for (let i = 1; i < points.length - 1; i++) {
          const leftDelta = secants[i - 1];
          const rightDelta = secants[i];
          if (leftDelta === 0 || rightDelta === 0 || Math.sign(leftDelta) !== Math.sign(rightDelta)) {
            shared[i] = 0;
          } else {
            const leftH = h[i - 1];
            const rightH = h[i];
            const w1 = 2 * rightH + leftH;
            const w2 = rightH + 2 * leftH;
            shared[i] = (w1 + w2) / ((w1 / leftDelta) + (w2 / rightDelta));
          }
        }
      }
      const leftSlopes = [...shared];
      const rightSlopes = [...shared];
      for (let i = 1; i < points.length - 1; i++) {
        if (points[i].mode === 'corner') {
          leftSlopes[i] = secants[i - 1];
          rightSlopes[i] = secants[i];
        }
      }
      return { xs, ys, h, leftSlopes, rightSlopes };
    }

    function getSegmentControls(segmentIndex, curveData) {
      const x0 = curveData.xs[segmentIndex];
      const y0 = curveData.ys[segmentIndex];
      const x1 = curveData.xs[segmentIndex + 1];
      const y1 = curveData.ys[segmentIndex + 1];
      const width = curveData.h[segmentIndex];
      return {
        cp1: { x: x0 + width / 3, y: y0 + curveData.rightSlopes[segmentIndex] * width / 3 },
        cp2: { x: x1 - width / 3, y: y1 - curveData.leftSlopes[segmentIndex + 1] * width / 3 },
        p3: { x: x1, y: y1 }
      };
    }

    function drawPointMarker(point, index) {
      const c = toCanvas(point);
      const radius = 7 * (window.devicePixelRatio || 1);
      const isSelected = index === state.selected;
      const isEndpoint = index === 0 || index === state.points.length - 1;
      const isCorner = !isEndpoint && point.mode === 'corner';
      ctx.save();
      ctx.strokeStyle = isSelected ? '#fb923c' : '#38bdf8';
      ctx.lineWidth = 2.5 * (window.devicePixelRatio || 1);
      if (isCorner) {
        ctx.beginPath();
        ctx.moveTo(c.x, c.y - radius);
        ctx.lineTo(c.x + radius, c.y);
        ctx.lineTo(c.x, c.y + radius);
        ctx.lineTo(c.x - radius, c.y);
        ctx.closePath();
        ctx.fillStyle = '#111827';
        ctx.fill();
        ctx.stroke();
      } else {
        ctx.beginPath();
        ctx.arc(c.x, c.y, radius, 0, Math.PI * 2);
        ctx.fillStyle = isSelected ? '#fb923c' : '#38bdf8';
        ctx.fill();
        ctx.strokeStyle = '#0f172a';
        ctx.lineWidth = 2 * (window.devicePixelRatio || 1);
        ctx.stroke();
      }
      ctx.restore();
    }

    function drawCurve() {
      const curveData = getBezierCurveData(state.points);
      if (!curveData) return;
      ctx.save();
      ctx.strokeStyle = '#f8fafc';
      ctx.lineWidth = 3 * (window.devicePixelRatio || 1);
      ctx.beginPath();
      const start = toCanvas(state.points[0]);
      ctx.moveTo(start.x, start.y);
      for (let i = 0; i < state.points.length - 1; i++) {
        const controls = getSegmentControls(i, curveData);
        const cp1 = toCanvas(controls.cp1);
        const cp2 = toCanvas(controls.cp2);
        const p3 = toCanvas(controls.p3);
        ctx.bezierCurveTo(cp1.x, cp1.y, cp2.x, cp2.y, p3.x, p3.y);
      }
      ctx.stroke();
      state.points.forEach((point, index) => drawPointMarker(point, index));
      ctx.restore();
    }

    function draw() {
      resizeCanvas();
      const r = graphRect();
      ctx.clearRect(0, 0, canvas.width, canvas.height);
      ctx.fillStyle = '#111827';
      ctx.fillRect(0, 0, canvas.width, canvas.height);
      drawHistogram(r);
      drawGrid(r);
      drawCurve();
      if (readout) {
        const selected = state.selected >= 0 ? state.points[state.selected] : null;
        readout.textContent = selected
          ? `Selected ${state.selected === 0 || state.selected === state.points.length - 1 ? 'endpoint' : selected.mode} point: input ${selected.x.toFixed(3)} → output ${selected.y.toFixed(3)}`
          : `${state.points.length} control points · cubic Bézier`;
      }
    }

    function pushCurve(autoPreview = true) {
      state.points = normalizePoints(state.points);
      const payload = JSON.stringify(state.points.map(p => ({ x: +p.x.toFixed(5), y: +p.y.toFixed(5), mode: p.mode })));
      state.lastCurveJson = payload;
      setTextareaValue('curve-points-json', payload);
      draw();
      if (autoPreview && state.hasDepth) {
        window.clearTimeout(state.refreshTimer);
        state.refreshTimer = window.setTimeout(() => clickGradioButton('auto-preview-depth-btn'), 350);
      }
    }

    function nearestPoint(x, y) {
      const threshold = 15 * (window.devicePixelRatio || 1);
      let best = -1;
      let bestDistance = Infinity;
      state.points.forEach((point, index) => {
        const c = toCanvas(point);
        const distance = Math.hypot(c.x - x, c.y - y);
        if (distance < threshold && distance < bestDistance) {
          best = index;
          bestDistance = distance;
        }
      });
      return best;
    }

    function pointerPosition(event) {
      const rect = canvas.getBoundingClientRect();
      const dpr = window.devicePixelRatio || 1;
      return { x: (event.clientX - rect.left) * dpr, y: (event.clientY - rect.top) * dpr };
    }

    function toggleSelectedPointMode() {
      if (state.selected <= 0 || state.selected >= state.points.length - 1) return;
      state.points[state.selected].mode = state.points[state.selected].mode === 'corner' ? 'smooth' : 'corner';
      pushCurve(true);
    }

    canvas.addEventListener('pointerdown', (event) => {
      event.preventDefault();
      canvas.setPointerCapture(event.pointerId);
      const pos = pointerPosition(event);
      state.pointerStart = pos;
      state.pointerMoved = false;
      state.pressedSelected = -1;
      const nearest = nearestPoint(pos.x, pos.y);
      if (nearest >= 0) {
        if (state.selected === nearest) state.pressedSelected = nearest;
        state.dragging = nearest;
        state.selected = nearest;
      } else {
        const point = { ...fromCanvas(pos.x, pos.y), mode: 'smooth' };
        state.points.push(point);
        state.points.sort((a, b) => a.x - b.x);
        state.selected = state.points.findIndex(p => p === point);
        state.dragging = state.selected;
        pushCurve(false);
      }
      draw();
    });

    canvas.addEventListener('pointermove', (event) => {
      if (state.dragging < 0) return;
      event.preventDefault();
      const pos = pointerPosition(event);
      if (state.pointerStart && Math.hypot(pos.x - state.pointerStart.x, pos.y - state.pointerStart.y) > 3 * (window.devicePixelRatio || 1)) {
        state.pointerMoved = true;
      }
      const point = fromCanvas(pos.x, pos.y);
      const i = state.dragging;
      const left = i === 0 ? 0 : state.points[i - 1].x + 0.002;
      const right = i === state.points.length - 1 ? 1 : state.points[i + 1].x - 0.002;
      state.points[i].x = i === 0 ? 0 : i === state.points.length - 1 ? 1 : clamp(point.x, left, right);
      state.points[i].y = point.y;
      pushCurve(true);
    });

    canvas.addEventListener('pointerup', (event) => {
      const shouldToggle = state.dragging === state.pressedSelected && !state.pointerMoved;
      if (shouldToggle) toggleSelectedPointMode();
      state.dragging = -1;
      state.pressedSelected = -1;
      state.pointerMoved = false;
      state.pointerStart = null;
      try { canvas.releasePointerCapture(event.pointerId); } catch (_) {}
      draw();
    });

    canvas.addEventListener('dblclick', (event) => {
      event.preventDefault();
      const pos = pointerPosition(event);
      const nearest = nearestPoint(pos.x, pos.y);
      if (nearest > 0 && nearest < state.points.length - 1) {
        state.points.splice(nearest, 1);
        state.selected = -1;
        pushCurve(true);
      }
    });

    window.addEventListener('keydown', (event) => {
      if ((event.key === 'Delete' || event.key === 'Backspace') && state.selected > 0 && state.selected < state.points.length - 1) {
        state.points.splice(state.selected, 1);
        state.selected = -1;
        pushCurve(true);
      }
    });

    document.addEventListener('click', (event) => {
      const button = event.target.closest('[data-depth-preset]');
      if (!button) return;
      const preset = button.dataset.depthPreset;
      state.points = normalizePoints(PRESETS[preset] || DEFAULT_POINTS);
      state.selected = -1;
      pushCurve(true);
    });

    window.addEventListener('resize', draw);

    window.setInterval(() => {
      const curveJson = readTextareaValue('curve-points-json');
      if (curveJson && curveJson !== state.lastCurveJson) {
        state.points = normalizePoints(parseJson(curveJson, DEFAULT_POINTS));
        state.lastCurveJson = curveJson;
        draw();
      }
      const histogramJson = readTextareaValue('depth-histogram-json');
      if (histogramJson && histogramJson !== state.lastHistogramJson) {
        const parsed = parseJson(histogramJson, []);
        state.histogram = Array.isArray(parsed) ? parsed : [];
        state.hasDepth = state.histogram.some(count => Number(count) > 0);
        state.lastHistogramJson = histogramJson;
        draw();
      }
    }, 500);

    pushCurve(false);
    draw();
  }

  function startCurveEditorWatcher() {
    initCurveEditor();
    window.setInterval(initCurveEditor, 500);
  }

  if (document.readyState === 'loading') {
    document.addEventListener('DOMContentLoaded', startCurveEditorWatcher);
  } else {
    startCurveEditorWatcher();
  }
  window.addEventListener('load', initCurveEditor);
})();
</script>
"""


CURVE_EDITOR_HTML = """
<div class="depth-editor-card">
  <strong>Depth histogram + curves editor</strong>
  <div class="depth-editor-toolbar">
    <button type="button" data-depth-preset="linear">Linear</button>
    <button type="button" data-depth-preset="subject">Midtone contrast</button>
    <button type="button" data-depth-preset="compressHigh">Compress high values</button>
    <button type="button" data-depth-preset="compressLow">Compress low values</button>
    <button type="button" data-depth-preset="foregroundPop">Foreground pop</button>
    <button type="button" data-depth-preset="invert">Invert curve</button>
  </div>
  <canvas id="depth-curve-canvas" width="720" height="360"></canvas>
  <div id="depth-curve-readout" class="depth-editor-readout"></div>
  <div class="depth-editor-help">
    Drag control points to remap the normalized height interpretation before STL export. The curve uses piecewise cubic Bézier segments.
    Click the curve to add a point. Click an already selected point to toggle smooth/corner continuity. Corner points draw as diamonds. Double-click a non-endpoint, or select it and press Delete, to remove it.
    The blue histogram shows the current smoothed base-depth distribution.
  </div>
</div>
"""


def clamp(value: float, minimum: float, maximum: float) -> float:
    return max(minimum, min(maximum, value))


def normalize_array(values: np.ndarray) -> np.ndarray:
    values = values.astype(np.float32, copy=False)
    v_min = float(np.nanmin(values))
    v_max = float(np.nanmax(values))
    if not np.isfinite(v_min) or not np.isfinite(v_max) or v_max <= v_min:
        return np.zeros_like(values, dtype=np.float32)
    return ((values - v_min) / (v_max - v_min)).astype(np.float32)


def sanitize_max_pixels(max_pixels: Any) -> int:
    try:
        requested = int(float(max_pixels))
    except (TypeError, ValueError):
        requested = DEFAULT_MAX_RESOLUTION
    return int(clamp(requested, MIN_RESOLUTION_LIMIT, MAX_RESOLUTION_LIMIT))

def coerce_float(value: Any, default: float) -> float:
    try:
        if value is None:
            return float(default)
        return float(value)
    except (TypeError, ValueError):
        return float(default)


def coerce_z_bounds(min_z: Any, max_z: Any) -> tuple[float, float]:
    min_z_value = coerce_float(min_z, DEFAULT_MIN_Z)
    max_z_value = coerce_float(max_z, DEFAULT_MAX_Z)
    return min_z_value, max_z_value

def get_lanczos_resample_filter() -> int:
    return getattr(getattr(Image, "Resampling", Image), "LANCZOS")


def load_and_resize_image(input_filepath: str, max_pixels: int) -> tuple[np.ndarray, str]:
    if input_filepath is None:
        raise gr.Error("Please upload an image.")

    filepath = str(input_filepath)
    filepath_lower = filepath.lower()

    try:
        image_pil = Image.open(filepath)
        image_pil = ImageOps.exif_transpose(image_pil).convert("RGB")
    except Exception as exc:
        heif_hint = ""
        if filepath_lower.endswith((".heic", ".heif")):
            if HEIF_SUPPORT:
                heif_hint = " The file looks like HEIC/HEIF; pillow-heif is installed but could not decode this file."
            else:
                heif_hint = f" HEIC/HEIF support is unavailable because pillow-heif failed to load: {HEIF_IMPORT_ERROR}"
        raise gr.Error(f"Could not open image file.{heif_hint} Error: {exc}") from exc

    original_w, original_h = image_pil.size
    original_pixels = original_w * original_h
    max_pixels = sanitize_max_pixels(max_pixels)
    resized = False

    if original_pixels > max_pixels:
        ratio = (max_pixels / original_pixels) ** 0.5
        new_w = max(1, int(original_w * ratio))
        new_h = max(1, int(original_h * ratio))
        image_pil = image_pil.resize((new_w, new_h), get_lanczos_resample_filter())
        resized = True

    input_image = np.array(image_pil, dtype=np.uint8)
    current_h, current_w = input_image.shape[:2]
    status = (
        f"Loaded {original_w}×{original_h} image. "
        f"Processing at {current_w}×{current_h} ({current_w * current_h:,} pixels)."
    )
    if resized:
        status += f" Downsampled to stay under the {max_pixels:,}-pixel limit."
    if filepath_lower.endswith((".heic", ".heif")):
        status += " HEIC/HEIF decoded server-side."
    return input_image, status


def estimate_depth_map(input_image: np.ndarray) -> np.ndarray:
    image = Image.fromarray(input_image).convert("RGB")
    with torch.no_grad():
        inputs = processor(images=image, return_tensors="pt").to(device)
        outputs = model(**inputs)
        predicted_depth = outputs.predicted_depth
        depth = torch.nn.functional.interpolate(
            predicted_depth.unsqueeze(1),
            size=(image.height, image.width),
            mode="bilinear",
            align_corners=False,
        ).squeeze().cpu().numpy()
    return normalize_depth_for_emboss(depth)


def apply_smoothing(values: np.ndarray, smoothing: int) -> np.ndarray:
    try:
        ksize = int(smoothing)
    except (TypeError, ValueError):
        ksize = 0
    if ksize <= 1:
        return values.astype(np.float32, copy=True)
    if ksize % 2 == 0:
        ksize += 1
    return cv2.GaussianBlur(values.astype(np.float32), (ksize, ksize), 0).astype(np.float32)


def sanitize_curve_points(curve_points_json: str | None) -> list[dict[str, Any]]:
    try:
        raw_points = json.loads(curve_points_json or "")
    except (TypeError, json.JSONDecodeError):
        raw_points = DEFAULT_CURVE_POINTS

    points: list[dict[str, Any]] = []
    if isinstance(raw_points, list):
        for point in raw_points:
            mode = "smooth"
            if isinstance(point, dict):
                x, y = point.get("x"), point.get("y")
                raw_mode = str(point.get("mode", "smooth")).lower()
                if raw_mode in {"corner", "broken", "linear"}:
                    mode = "corner"
            elif isinstance(point, (list, tuple)) and len(point) >= 2:
                x, y = point[0], point[1]
            else:
                continue
            try:
                x_f = clamp(float(x), 0.0, 1.0)
                y_f = clamp(float(y), 0.0, 1.0)
            except (TypeError, ValueError):
                continue
            points.append({"x": x_f, "y": y_f, "mode": mode})

    if len(points) < 2:
        points = [{**point, "mode": "smooth"} for point in DEFAULT_CURVE_POINTS]

    points.sort(key=lambda point: float(point["x"]))
    deduped: list[dict[str, Any]] = []
    for point in points:
        if deduped and abs(float(point["x"]) - float(deduped[-1]["x"])) < 1e-5:
            deduped[-1] = point
        else:
            deduped.append(point)

    if len(deduped) < 2:
        deduped = [{**point, "mode": "smooth"} for point in DEFAULT_CURVE_POINTS]

    deduped[0]["x"] = 0.0
    deduped[-1]["x"] = 1.0
    deduped[0]["mode"] = "smooth"
    deduped[-1]["mode"] = "smooth"
    return deduped



def curve_points_to_json(points: list[dict[str, Any]]) -> str:
    return json.dumps([
        {
            "x": round(float(point["x"]), 5),
            "y": round(float(point["y"]), 5),
            "mode": "corner" if str(point.get("mode", "smooth")) == "corner" else "smooth",
        }
        for point in points
    ])



def pchip_endpoint_slope(h0: float, h1: float, delta0: float, delta1: float) -> float:
    slope = ((2.0 * h0 + h1) * delta0 - h0 * delta1) / max(h0 + h1, 1e-6)
    if slope == 0.0 or delta0 == 0.0 or np.sign(slope) != np.sign(delta0):
        return 0.0
    if np.sign(delta0) != np.sign(delta1) and abs(slope) > abs(3.0 * delta0):
        return float(3.0 * delta0)
    return float(slope)



def cubic_bezier_curve_data(points: list[dict[str, Any]]) -> tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray]:
    x_values = np.array([point["x"] for point in points], dtype=np.float32)
    y_values = np.array([point["y"] for point in points], dtype=np.float32)
    interval_count = len(points) - 1
    if interval_count <= 0:
        return x_values, y_values, np.zeros_like(x_values), np.zeros_like(x_values)

    h_values = np.maximum(np.diff(x_values), 1e-6)
    secant_slopes = np.diff(y_values) / h_values
    shared_slopes = np.zeros(len(points), dtype=np.float32)

    if len(points) == 2:
        shared_slopes[:] = secant_slopes[0]
    else:
        shared_slopes[0] = pchip_endpoint_slope(float(h_values[0]), float(h_values[1]), float(secant_slopes[0]), float(secant_slopes[1]))
        shared_slopes[-1] = pchip_endpoint_slope(float(h_values[-1]), float(h_values[-2]), float(secant_slopes[-1]), float(secant_slopes[-2]))
        for i in range(1, len(points) - 1):
            left_delta = float(secant_slopes[i - 1])
            right_delta = float(secant_slopes[i])
            if left_delta == 0.0 or right_delta == 0.0 or np.sign(left_delta) != np.sign(right_delta):
                shared_slopes[i] = 0.0
            else:
                left_h = float(h_values[i - 1])
                right_h = float(h_values[i])
                w1 = 2.0 * right_h + left_h
                w2 = right_h + 2.0 * left_h
                shared_slopes[i] = float((w1 + w2) / ((w1 / left_delta) + (w2 / right_delta)))

    left_slopes = shared_slopes.copy()
    right_slopes = shared_slopes.copy()
    for i in range(1, len(points) - 1):
        if str(points[i].get("mode", "smooth")) == "corner":
            left_slopes[i] = secant_slopes[i - 1]
            right_slopes[i] = secant_slopes[i]

    return x_values, y_values, left_slopes.astype(np.float32), right_slopes.astype(np.float32)



def evaluate_piecewise_cubic_bezier(values: np.ndarray, points: list[dict[str, Any]]) -> np.ndarray:
    x_values, y_values, left_slopes, right_slopes = cubic_bezier_curve_data(points)
    if len(points) <= 1:
        return np.full_like(values, y_values[0] if len(y_values) else 0.0, dtype=np.float32)

    flat_values = values.astype(np.float32, copy=False).reshape(-1)
    interval_indices = np.searchsorted(x_values, flat_values, side="right") - 1
    interval_indices = np.clip(interval_indices, 0, len(x_values) - 2)
    interval_widths = np.maximum(x_values[interval_indices + 1] - x_values[interval_indices], 1e-6)
    t_values = np.clip((flat_values - x_values[interval_indices]) / interval_widths, 0.0, 1.0)
    omt_values = 1.0 - t_values

    p0 = y_values[interval_indices]
    p1 = y_values[interval_indices] + right_slopes[interval_indices] * (interval_widths / 3.0)
    p2 = y_values[interval_indices + 1] - left_slopes[interval_indices + 1] * (interval_widths / 3.0)
    p3 = y_values[interval_indices + 1]

    adjusted = (
        (omt_values ** 3) * p0
        + 3.0 * (omt_values ** 2) * t_values * p1
        + 3.0 * omt_values * (t_values ** 2) * p2
        + (t_values ** 3) * p3
    )
    return adjusted.reshape(values.shape).astype(np.float32)



def apply_depth_curve(depth_map: np.ndarray, curve_points_json: str | None, invert_depth: bool) -> np.ndarray:
    working = depth_map.astype(np.float32, copy=True)
    if invert_depth:
        working = 1.0 - working
    points = sanitize_curve_points(curve_points_json)
    adjusted = evaluate_piecewise_cubic_bezier(working, points)
    return np.clip(adjusted, 0.0, 1.0)


def depth_to_preview_image(depth_map: np.ndarray) -> Image.Image:
    preview = np.clip(depth_map, 0.0, 1.0)
    preview_uint8 = np.rint(np.clip(preview, 0.0, 1.0) * 255.0).astype(np.uint8)
    return Image.fromarray(preview_uint8)


def histogram_json(depth_map: np.ndarray, bins: int = 96) -> str:
    counts, _ = np.histogram(np.clip(depth_map, 0.0, 1.0), bins=bins, range=(0.0, 1.0))
    return json.dumps(counts.astype(int).tolist())


def prepare_depth(input_filepath: str):
    input_image, load_status = load_and_resize_image(input_filepath, CONFIGURED_MAX_RESOLUTION)
    depth_normalized = estimate_depth_map(input_image)
    curve_json = curve_points_to_json(DEFAULT_CURVE_POINTS)
    adjusted_depth = apply_depth_curve(depth_normalized, curve_json, invert_depth=False)
    height_map = combine_depth_and_texture(
        input_image,
        adjusted_depth,
        DEFAULT_TEXTURE_STRENGTH,
        DEFAULT_TEXTURE_SMOOTHING,
    )
    state = {
        "image_rgb": input_image,
        "raw_depth": depth_normalized,
        "adjusted_depth": adjusted_depth,
        "height_map": height_map,
        "curve_points_json": curve_json,
        "depth_map_smoothing": 0,
        "invert_depth": False,
        "texture_strength": DEFAULT_TEXTURE_STRENGTH,
        "texture_smoothing": DEFAULT_TEXTURE_SMOOTHING,
        "load_status": load_status,
    }
    status = (
        f"{load_status}\nDepth estimation complete. Use the curve and luminance texture controls to tune the 2D height map before generating the STL."
    )
    return (
        state,
        Image.fromarray(input_image),
        depth_to_preview_image(depth_normalized),
        depth_to_preview_image(height_map),
        histogram_json(depth_normalized),
        curve_json,
        status,
        "Preview is using the linear curve with the default luminance texture settings.",
    )


def clear_depth_outputs():
    return (
        None,
        None,
        None,
        None,
        "[]",
        curve_points_to_json(DEFAULT_CURVE_POINTS),
        "Upload an image to estimate its depth map.",
        "",
    )


def clear_preview_outputs(message: str = "Upload an image to estimate its depth map before tuning preview controls."):
    return (
        None,
        None,
        "[]",
        curve_points_to_json(DEFAULT_CURVE_POINTS),
        message,
    )

def auto_prepare_depth(input_filepath: str):
    if input_filepath is None:
        return clear_depth_outputs()
    return prepare_depth(input_filepath)

def get_smoothed_depth_from_state(depth_state: dict[str, Any] | None, depth_map_smoothing: int) -> np.ndarray:
    if not depth_state or "raw_depth" not in depth_state:
        raise gr.Error("Please upload an image and estimate its depth first.")
    return normalize_array(apply_smoothing(depth_state["raw_depth"], depth_map_smoothing))


def update_depth_preview(
    depth_state: dict[str, Any] | None,
    curve_points_json: str | None,
    depth_map_smoothing: int,
    invert_depth: bool,
    texture_strength: float,
    texture_smoothing: int,
):
    if not depth_state or "raw_depth" not in depth_state:
        return clear_preview_outputs()

    texture_strength = coerce_float(texture_strength, DEFAULT_TEXTURE_STRENGTH)
    texture_smoothing = int(coerce_float(texture_smoothing, DEFAULT_TEXTURE_SMOOTHING))
    smoothed_depth = get_smoothed_depth_from_state(depth_state, depth_map_smoothing)
    adjusted_depth = apply_depth_curve(smoothed_depth, curve_points_json, invert_depth)
    points = sanitize_curve_points(curve_points_json)
    normalized_curve_json = curve_points_to_json(points)
    height_map = combine_depth_and_texture(
        depth_state["image_rgb"],
        adjusted_depth,
        texture_strength,
        texture_smoothing,
    )
    if depth_state is not None:
        depth_state["adjusted_depth"] = adjusted_depth
        depth_state["height_map"] = height_map
        depth_state["curve_points_json"] = normalized_curve_json
        depth_state["depth_map_smoothing"] = int(depth_map_smoothing)
        depth_state["invert_depth"] = bool(invert_depth)
        depth_state["texture_strength"] = float(texture_strength)
        depth_state["texture_smoothing"] = int(texture_smoothing)
    status = (
        f"Preview updated with {len(points)} curve points, texture strength {float(texture_strength):.3f}, "
        f"texture smoothing {int(texture_smoothing)}."
    )
    return (
        depth_state,
        depth_to_preview_image(height_map),
        histogram_json(smoothed_depth),
        normalized_curve_json,
        status,
    )


def get_adjusted_depth_for_settings(
    depth_state: dict[str, Any],
    curve_points_json: str | None,
    depth_map_smoothing: int,
    invert_depth: bool,
) -> np.ndarray:
    normalized_curve_json = curve_points_to_json(sanitize_curve_points(curve_points_json))
    cached_depth = depth_state.get("adjusted_depth")
    cache_matches = (
        cached_depth is not None
        and depth_state.get("curve_points_json") == normalized_curve_json
        and int(depth_state.get("depth_map_smoothing", -1)) == int(depth_map_smoothing)
        and bool(depth_state.get("invert_depth", False)) == bool(invert_depth)
    )
    if cache_matches:
        return cached_depth

    smoothed_depth = get_smoothed_depth_from_state(depth_state, depth_map_smoothing)
    adjusted_depth = apply_depth_curve(smoothed_depth, normalized_curve_json, invert_depth)
    depth_state["adjusted_depth"] = adjusted_depth
    depth_state["curve_points_json"] = normalized_curve_json
    depth_state["depth_map_smoothing"] = int(depth_map_smoothing)
    depth_state["invert_depth"] = bool(invert_depth)
    return adjusted_depth


def get_height_map_for_settings(
    depth_state: dict[str, Any],
    curve_points_json: str | None,
    depth_map_smoothing: int,
    invert_depth: bool,
    texture_strength: float,
    texture_smoothing: int,
) -> np.ndarray:
    texture_strength = coerce_float(texture_strength, DEFAULT_TEXTURE_STRENGTH)
    texture_smoothing = int(coerce_float(texture_smoothing, DEFAULT_TEXTURE_SMOOTHING))
    normalized_curve_json = curve_points_to_json(sanitize_curve_points(curve_points_json))
    cached_height_map = depth_state.get("height_map")
    cache_matches = (
        cached_height_map is not None
        and depth_state.get("curve_points_json") == normalized_curve_json
        and int(depth_state.get("depth_map_smoothing", -1)) == int(depth_map_smoothing)
        and bool(depth_state.get("invert_depth", False)) == bool(invert_depth)
        and float(depth_state.get("texture_strength", -1.0)) == float(texture_strength)
        and int(depth_state.get("texture_smoothing", -1)) == int(texture_smoothing)
    )
    if cache_matches:
        return cached_height_map

    adjusted_depth = get_adjusted_depth_for_settings(
        depth_state,
        normalized_curve_json,
        depth_map_smoothing,
        invert_depth,
    )
    height_map = combine_depth_and_texture(
        depth_state["image_rgb"],
        adjusted_depth,
        texture_strength,
        texture_smoothing,
    )
    depth_state["height_map"] = height_map
    depth_state["texture_strength"] = float(texture_strength)
    depth_state["texture_smoothing"] = int(texture_smoothing)
    return height_map


def combine_depth_and_texture(
    image_rgb: np.ndarray,
    adjusted_depth: np.ndarray,
    texture_strength: float,
    texture_smoothing: int,
) -> np.ndarray:
    texture_strength = clamp(coerce_float(texture_strength, DEFAULT_TEXTURE_STRENGTH), 0.0, 0.5)
    if texture_strength <= 0.0:
        return np.clip(adjusted_depth, 0.0, 1.0).astype(np.float32)

    gray_image = cv2.cvtColor(image_rgb, cv2.COLOR_RGB2GRAY).astype(np.float32) / 255.0
    low_frequency = apply_smoothing(gray_image, 5)
    detail = gray_image - low_frequency
    texture_smoothing = int(coerce_float(texture_smoothing, DEFAULT_TEXTURE_SMOOTHING))
    if texture_smoothing > 1:
        detail = apply_smoothing(detail, texture_smoothing)

    if detail.shape != adjusted_depth.shape:
        detail = cv2.resize(
            detail,
            (adjusted_depth.shape[1], adjusted_depth.shape[0]),
            interpolation=cv2.INTER_LINEAR,
        )

    scale = float(np.percentile(np.abs(detail), 95.0))
    if not np.isfinite(scale) or scale <= 1e-6:
        detail = np.zeros_like(adjusted_depth, dtype=np.float32)
    else:
        detail = np.clip(detail / scale, -1.0, 1.0).astype(np.float32)
    combined_map = adjusted_depth + (detail * texture_strength)
    return np.clip(combined_map, 0.0, 1.0).astype(np.float32)


def apply_pca_correction_to_z(z_data: np.ndarray, x_length: float, min_z: float, max_z: float) -> np.ndarray:
    height, width = z_data.shape
    y_length = x_length * (height / width)
    x_coords_1d = np.linspace(0, x_length, width)
    y_coords_1d = np.linspace(y_length, 0, height)
    x_grid, y_grid = np.meshgrid(x_coords_1d, y_coords_1d)
    points = np.stack([x_grid.flatten(), y_grid.flatten(), z_data.flatten()], axis=1)
    n_points = points.shape[0]
    n_samples = min(n_points, 50_000)
    sample_indices = np.random.choice(n_points, n_samples, replace=False)
    pca = PCA(n_components=3)
    pca.fit(points[sample_indices])
    normal = pca.components_[2]
    if normal[2] < 0:
        normal *= -1
    p0 = pca.mean_
    z_plane = p0[2] - (normal[0] * (x_grid - p0[0]) + normal[1] * (y_grid - p0[1])) / normal[2]
    corrected_z = z_data - z_plane
    corrected_normalized = normalize_array(corrected_z)
    return min_z + corrected_normalized * (max_z - min_z)


def build_stl_faces(z_data: np.ndarray, x_length: float, close_body: bool) -> tuple[np.ndarray, float]:
    height, width = z_data.shape
    y_length = float(x_length * (height / width))
    z_data = z_data.astype(np.float32, copy=False)

    x_coords = np.linspace(0, x_length, width, dtype=np.float32)
    y_coords = np.linspace(y_length, 0, height, dtype=np.float32)
    x_grid, y_grid = np.meshgrid(x_coords, y_coords)
    vertices = np.stack([x_grid, y_grid, z_data], axis=-1)

    tl = vertices[:-1, :-1]
    bl = vertices[1:, :-1]
    br = vertices[1:, 1:]
    tr = vertices[:-1, 1:]

    quad_count = (height - 1) * (width - 1)
    top_faces = np.empty((quad_count * 2, 3, 3), dtype=np.float32)
    top_faces[0::2] = np.stack([tl, bl, br], axis=2).reshape(-1, 3, 3)
    top_faces[1::2] = np.stack([tl, br, tr], axis=2).reshape(-1, 3, 3)

    if not close_body:
        return top_faces, y_length

    v_tl = vertices[0, 0]
    v_tr = vertices[0, width - 1]
    v_bl = vertices[height - 1, 0]
    v_br = vertices[height - 1, width - 1]
    b_tl = np.array([v_tl[0], v_tl[1], 0], dtype=np.float32)
    b_tr = np.array([v_tr[0], v_tr[1], 0], dtype=np.float32)
    b_bl = np.array([v_bl[0], v_bl[1], 0], dtype=np.float32)
    b_br = np.array([v_br[0], v_br[1], 0], dtype=np.float32)

    side_faces = np.array([
        [v_tl, b_tl, b_tr], [v_tl, b_tr, v_tr],
        [v_br, b_br, b_bl], [v_br, b_bl, v_bl],
        [v_bl, b_bl, b_tl], [v_bl, b_tl, v_tl],
        [v_tr, b_tr, b_br], [v_tr, b_br, v_br],
    ], dtype=np.float32)
    base_faces = np.array([
        [b_tl, b_br, b_bl],
        [b_tl, b_tr, b_br],
    ], dtype=np.float32)
    return np.concatenate([top_faces, side_faces, base_faces], axis=0), y_length



def flip_faces_for_preview(faces: np.ndarray, y_length: float) -> np.ndarray:
    preview_faces = faces.copy()
    preview_faces[..., 1] = np.float32(y_length) - preview_faces[..., 1]
    preview_faces = preview_faces[:, [0, 2, 1], :]
    return preview_faces



def save_stl_faces(faces: np.ndarray) -> str:
    surface = mesh.Mesh(np.zeros(faces.shape[0], dtype=mesh.Mesh.dtype))
    surface.vectors = faces

    with tempfile.NamedTemporaryFile(delete=False, suffix=".stl") as tmpfile:
        surface.save(tmpfile.name)
        return tmpfile.name



def build_stl_mesh(z_data: np.ndarray, x_length: float, close_body: bool, flip_for_preview: bool = False) -> str:
    faces, y_length = build_stl_faces(z_data, x_length, close_body)
    if flip_for_preview:
        faces = flip_faces_for_preview(faces, y_length)
    return save_stl_faces(faces)


def generate_3d_model_from_adjusted_depth(
    depth_state: dict[str, Any] | None,
    curve_points_json: str | None,
    depth_map_smoothing: int,
    invert_depth: bool,
    texture_strength: float,
    texture_smoothing: int,
    max_z: float,
    min_z: float,
    x_length: float,
    do_pca_correction: bool,
    close_body: bool,
):
    min_z, max_z = coerce_z_bounds(min_z, max_z)
    x_length = coerce_float(x_length, 100.0)
    if max_z <= min_z:
        raise gr.Error("Max Z-height must be greater than Min Z-height.")
    if x_length <= 0:
        raise gr.Error("X Length must be positive.")
    if not depth_state or "image_rgb" not in depth_state:
        raise gr.Error("Please estimate a depth map before generating an STL.")

    height_map = get_height_map_for_settings(
        depth_state,
        curve_points_json,
        depth_map_smoothing,
        invert_depth,
        texture_strength,
        texture_smoothing,
    )
    z_data = min_z + height_map * (max_z - min_z)

    if do_pca_correction:
        z_data = apply_pca_correction_to_z(z_data, x_length, min_z, max_z)

    export_faces, y_length = build_stl_faces(z_data, x_length, close_body)
    export_stl_path = save_stl_faces(export_faces)
    preview_stl_path = save_stl_faces(flip_faces_for_preview(export_faces, y_length))
    point_count = len(sanitize_curve_points(curve_points_json))
    status = (
        f"Generated STL from previewed height map using {point_count} curve points and texture strength "
        f"{float(texture_strength):.3f}."
    )
    return preview_stl_path, export_stl_path, status


# CONTROLLED_EMBOSS_V1
DEFAULT_HEIGHT_GAMMA = 1.0
DEFAULT_BACKGROUND_FLATTEN = 0.08


def normalize_depth_for_emboss(values: np.ndarray) -> np.ndarray:
    values = values.astype(np.float32, copy=False)
    finite_values = values[np.isfinite(values)]
    if finite_values.size == 0:
        return np.zeros_like(values, dtype=np.float32)
    low, high = np.nanpercentile(finite_values, [2.0, 98.0])
    if not np.isfinite(low) or not np.isfinite(high) or high <= low:
        return normalize_array(values)
    return np.clip((values - low) / (high - low), 0.0, 1.0).astype(np.float32)


def apply_emboss_controls(height_map: np.ndarray, height_gamma: float, background_flatten: float) -> np.ndarray:
    gamma = clamp(coerce_float(height_gamma, DEFAULT_HEIGHT_GAMMA), 0.25, 4.0)
    flatten = clamp(coerce_float(background_flatten, DEFAULT_BACKGROUND_FLATTEN), 0.0, 0.8)
    controlled = np.power(np.clip(height_map, 0.0, 1.0), gamma).astype(np.float32)
    if flatten > 0.0:
        controlled = np.clip((controlled - flatten) / max(1.0 - flatten, 1e-6), 0.0, 1.0)
    return controlled.astype(np.float32)


def height_to_hillshade_image(height_map: np.ndarray) -> Image.Image:
    heights = np.clip(height_map, 0.0, 1.0).astype(np.float32)
    gradient_y, gradient_x = np.gradient(heights)
    normals = np.dstack((-gradient_x * 4.0, -gradient_y * 4.0, np.ones_like(heights)))
    normals /= np.maximum(np.linalg.norm(normals, axis=2, keepdims=True), 1e-6)
    light = np.array([-0.45, -0.35, 0.82], dtype=np.float32)
    light /= np.linalg.norm(light)
    shade = 0.18 + 0.82 * np.tensordot(normals, light, axes=([2], [0]))
    return Image.fromarray(np.rint(np.clip(shade, 0.0, 1.0) * 255.0).astype(np.uint8))


def save_height_map_png(height_map: np.ndarray) -> str:
    height_uint16 = np.rint(np.clip(height_map, 0.0, 1.0) * 65535.0).astype(np.uint16)
    with tempfile.NamedTemporaryFile(delete=False, suffix=".png") as tmpfile:
        Image.fromarray(height_uint16, mode="I;16").save(tmpfile.name, format="PNG")
        return tmpfile.name


def build_controlled_height_map(
    image_rgb: np.ndarray,
    adjusted_depth: np.ndarray,
    texture_strength: float,
    texture_smoothing: int,
    height_gamma: float,
    background_flatten: float,
) -> np.ndarray:
    controlled_depth = apply_emboss_controls(adjusted_depth, height_gamma, background_flatten)
    return combine_depth_and_texture(image_rgb, controlled_depth, texture_strength, texture_smoothing)


def estimate_depth_map(input_image: np.ndarray) -> np.ndarray:
    image = Image.fromarray(input_image).convert("RGB")
    with torch.no_grad():
        inputs = processor(images=image, return_tensors="pt").to(device)
        outputs = model(**inputs)
        predicted_depth = outputs.predicted_depth
        depth = torch.nn.functional.interpolate(
            predicted_depth.unsqueeze(1),
            size=(image.height, image.width),
            mode="bilinear",
            align_corners=False,
        ).squeeze().cpu().numpy()
    return normalize_depth_for_emboss(depth)


def prepare_depth(input_filepath: str):
    input_image, load_status = load_and_resize_image(input_filepath, CONFIGURED_MAX_RESOLUTION)
    depth_normalized = estimate_depth_map(input_image)
    curve_json = curve_points_to_json(DEFAULT_CURVE_POINTS)
    adjusted_depth = apply_depth_curve(depth_normalized, curve_json, invert_depth=False)
    height_map = build_controlled_height_map(
        input_image,
        adjusted_depth,
        DEFAULT_TEXTURE_STRENGTH,
        DEFAULT_TEXTURE_SMOOTHING,
        DEFAULT_HEIGHT_GAMMA,
        DEFAULT_BACKGROUND_FLATTEN,
    )
    state = {
        "image_rgb": input_image,
        "raw_depth": depth_normalized,
        "adjusted_depth": adjusted_depth,
        "height_map": height_map,
        "curve_points_json": curve_json,
        "depth_map_smoothing": 0,
        "invert_depth": False,
        "texture_strength": DEFAULT_TEXTURE_STRENGTH,
        "texture_smoothing": DEFAULT_TEXTURE_SMOOTHING,
        "height_gamma": DEFAULT_HEIGHT_GAMMA,
        "background_flatten": DEFAULT_BACKGROUND_FLATTEN,
        "load_status": load_status,
    }
    status = (
        f"{load_status}\nDepth estimation complete. Tune gamma and background flatten on the true grayscale height map before adding fine detail."
    )
    return (
        state,
        Image.fromarray(input_image),
        depth_to_preview_image(depth_normalized),
        depth_to_preview_image(height_map),
        height_to_hillshade_image(height_map),
        save_height_map_png(height_map),
        histogram_json(depth_normalized),
        curve_json,
        status,
        "Preview uses robust depth normalization, linear gamma, and the default background flatten threshold.",
    )


def clear_depth_outputs():
    return (
        None,
        None,
        None,
        None,
        None,
        None,
        "[]",
        curve_points_to_json(DEFAULT_CURVE_POINTS),
        "Upload an image to estimate its depth map.",
        "",
    )


def clear_preview_outputs(message: str = "Upload an image to estimate its depth map before tuning preview controls."):
    return (
        None,
        None,
        None,
        None,
        "[]",
        curve_points_to_json(DEFAULT_CURVE_POINTS),
        message,
    )


def get_smoothed_depth_from_state(depth_state: dict[str, Any] | None, depth_map_smoothing: int) -> np.ndarray:
    if not depth_state or "raw_depth" not in depth_state:
        raise gr.Error("Please upload an image and estimate its depth first.")
    return normalize_depth_for_emboss(apply_smoothing(depth_state["raw_depth"], depth_map_smoothing))


def update_depth_preview(
    depth_state: dict[str, Any] | None,
    curve_points_json: str | None,
    depth_map_smoothing: int,
    invert_depth: bool,
    texture_strength: float,
    texture_smoothing: int,
    height_gamma: float,
    background_flatten: float,
):
    if not depth_state or "raw_depth" not in depth_state:
        return clear_preview_outputs()

    texture_strength = clamp(coerce_float(texture_strength, DEFAULT_TEXTURE_STRENGTH), 0.0, 0.5)
    texture_smoothing = int(coerce_float(texture_smoothing, DEFAULT_TEXTURE_SMOOTHING))
    height_gamma = clamp(coerce_float(height_gamma, DEFAULT_HEIGHT_GAMMA), 0.25, 4.0)
    background_flatten = clamp(coerce_float(background_flatten, DEFAULT_BACKGROUND_FLATTEN), 0.0, 0.8)
    smoothed_depth = get_smoothed_depth_from_state(depth_state, depth_map_smoothing)
    adjusted_depth = apply_depth_curve(smoothed_depth, curve_points_json, invert_depth)
    points = sanitize_curve_points(curve_points_json)
    normalized_curve_json = curve_points_to_json(points)
    height_map = build_controlled_height_map(
        depth_state["image_rgb"],
        adjusted_depth,
        texture_strength,
        texture_smoothing,
        height_gamma,
        background_flatten,
    )
    depth_state["adjusted_depth"] = adjusted_depth
    depth_state["height_map"] = height_map
    depth_state["curve_points_json"] = normalized_curve_json
    depth_state["depth_map_smoothing"] = int(depth_map_smoothing)
    depth_state["invert_depth"] = bool(invert_depth)
    depth_state["texture_strength"] = float(texture_strength)
    depth_state["texture_smoothing"] = int(texture_smoothing)
    depth_state["height_gamma"] = float(height_gamma)
    depth_state["background_flatten"] = float(background_flatten)
    status = (
        f"Controlled preview updated with gamma {height_gamma:.2f}, background flatten {background_flatten:.2f}, "
        f"and fine detail strength {texture_strength:.3f}."
    )
    return (
        depth_state,
        depth_to_preview_image(height_map),
        height_to_hillshade_image(height_map),
        save_height_map_png(height_map),
        histogram_json(smoothed_depth),
        normalized_curve_json,
        status,
    )


def get_adjusted_depth_for_settings(
    depth_state: dict[str, Any],
    curve_points_json: str | None,
    depth_map_smoothing: int,
    invert_depth: bool,
) -> np.ndarray:
    normalized_curve_json = curve_points_to_json(sanitize_curve_points(curve_points_json))
    cached_depth = depth_state.get("adjusted_depth")
    cache_matches = (
        cached_depth is not None
        and depth_state.get("curve_points_json") == normalized_curve_json
        and int(depth_state.get("depth_map_smoothing", -1)) == int(depth_map_smoothing)
        and bool(depth_state.get("invert_depth", False)) == bool(invert_depth)
    )
    if cache_matches:
        return cached_depth

    smoothed_depth = get_smoothed_depth_from_state(depth_state, depth_map_smoothing)
    adjusted_depth = apply_depth_curve(smoothed_depth, normalized_curve_json, invert_depth)
    depth_state["adjusted_depth"] = adjusted_depth
    depth_state["curve_points_json"] = normalized_curve_json
    depth_state["depth_map_smoothing"] = int(depth_map_smoothing)
    depth_state["invert_depth"] = bool(invert_depth)
    return adjusted_depth


def get_height_map_for_settings(
    depth_state: dict[str, Any],
    curve_points_json: str | None,
    depth_map_smoothing: int,
    invert_depth: bool,
    texture_strength: float,
    texture_smoothing: int,
    height_gamma: float,
    background_flatten: float,
) -> np.ndarray:
    texture_strength = clamp(coerce_float(texture_strength, DEFAULT_TEXTURE_STRENGTH), 0.0, 0.5)
    texture_smoothing = int(coerce_float(texture_smoothing, DEFAULT_TEXTURE_SMOOTHING))
    height_gamma = clamp(coerce_float(height_gamma, DEFAULT_HEIGHT_GAMMA), 0.25, 4.0)
    background_flatten = clamp(coerce_float(background_flatten, DEFAULT_BACKGROUND_FLATTEN), 0.0, 0.8)
    normalized_curve_json = curve_points_to_json(sanitize_curve_points(curve_points_json))
    cached_height_map = depth_state.get("height_map")
    cache_matches = (
        cached_height_map is not None
        and depth_state.get("curve_points_json") == normalized_curve_json
        and int(depth_state.get("depth_map_smoothing", -1)) == int(depth_map_smoothing)
        and bool(depth_state.get("invert_depth", False)) == bool(invert_depth)
        and float(depth_state.get("texture_strength", -1.0)) == float(texture_strength)
        and int(depth_state.get("texture_smoothing", -1)) == int(texture_smoothing)
        and float(depth_state.get("height_gamma", -1.0)) == float(height_gamma)
        and float(depth_state.get("background_flatten", -1.0)) == float(background_flatten)
    )
    if cache_matches:
        return cached_height_map

    adjusted_depth = get_adjusted_depth_for_settings(
        depth_state,
        normalized_curve_json,
        depth_map_smoothing,
        invert_depth,
    )
    height_map = build_controlled_height_map(
        depth_state["image_rgb"],
        adjusted_depth,
        texture_strength,
        texture_smoothing,
        height_gamma,
        background_flatten,
    )
    depth_state["height_map"] = height_map
    depth_state["texture_strength"] = float(texture_strength)
    depth_state["texture_smoothing"] = int(texture_smoothing)
    depth_state["height_gamma"] = float(height_gamma)
    depth_state["background_flatten"] = float(background_flatten)
    return height_map


def generate_3d_model_from_adjusted_depth(
    depth_state: dict[str, Any] | None,
    curve_points_json: str | None,
    depth_map_smoothing: int,
    invert_depth: bool,
    texture_strength: float,
    texture_smoothing: int,
    height_gamma: float,
    background_flatten: float,
    max_z: float,
    min_z: float,
    x_length: float,
    do_pca_correction: bool,
    close_body: bool,
):
    min_z, max_z = coerce_z_bounds(min_z, max_z)
    x_length = coerce_float(x_length, 100.0)
    if max_z <= min_z:
        raise gr.Error("Max Z-height must be greater than Min Z-height.")
    if x_length <= 0:
        raise gr.Error("X Length must be positive.")
    if not depth_state or "image_rgb" not in depth_state:
        raise gr.Error("Please estimate a depth map before generating an STL.")

    height_map = get_height_map_for_settings(
        depth_state,
        curve_points_json,
        depth_map_smoothing,
        invert_depth,
        texture_strength,
        texture_smoothing,
        height_gamma,
        background_flatten,
    )
    z_data = min_z + height_map * (max_z - min_z)
    if do_pca_correction:
        z_data = apply_pca_correction_to_z(z_data, x_length, min_z, max_z)

    export_faces, y_length = build_stl_faces(z_data, x_length, close_body)
    export_stl_path = save_stl_faces(export_faces)
    preview_stl_path = save_stl_faces(flip_faces_for_preview(export_faces, y_length))
    point_count = len(sanitize_curve_points(curve_points_json))
    status = (
        f"Generated STL from controlled height map using {point_count} curve points, gamma {float(height_gamma):.2f}, "
        f"and background flatten {float(background_flatten):.2f}."
    )
    return preview_stl_path, export_stl_path, status


# PUBLIC_TEST_EMBOSS_V1
TEST_LINEAR_CURVE = [
    {"x": 0.0, "y": 0.0},
    {"x": 1.0, "y": 1.0},
]
TEST_COMPRESS_HIGH_CURVE = [
    {"x": 0.0, "y": 0.0},
    {"x": 0.45, "y": 0.52},
    {"x": 0.75, "y": 0.74},
    {"x": 1.0, "y": 0.86},
]


def prepare_public_test_outputs(
    input_filepath: str | None,
    version: str,
    texture_strength: float,
) -> tuple[Image.Image | None, Image.Image | None, str]:
    if not input_filepath:
        return None, None, "Upload an image first."

    input_image, _ = load_and_resize_image(input_filepath, CONFIGURED_MAX_RESOLUTION)
    depth_map = estimate_depth_map(input_image)
    if version == "Version 2 — Compress high values":
        curve_points = TEST_COMPRESS_HIGH_CURVE
    else:
        curve_points = TEST_LINEAR_CURVE

    adjusted_depth = apply_depth_curve(
        depth_map,
        curve_points_to_json(curve_points),
        invert_depth=False,
    )
    height_map = build_controlled_height_map(
        input_image,
        adjusted_depth,
        texture_strength,
        DEFAULT_TEXTURE_SMOOTHING,
        DEFAULT_HEIGHT_GAMMA,
        DEFAULT_BACKGROUND_FLATTEN,
    )
    return (
        depth_to_preview_image(height_map),
        height_to_hillshade_image(height_map),
        f"Completed {version} with Fine Detail Strength {float(texture_strength):.3f}.",
    )


with gr.Blocks(title="Test Embossed Effect") as demo:
    gr.Markdown("# Test Embossed Effect")
    gr.Markdown("Upload one image and compare the true grayscale height map with its embossed surface preview.")

    input_image = gr.Image(
        type="filepath",
        label="Upload one image",
    )
    version = gr.Radio(
        choices=["Version 1 — Linear", "Version 2 — Compress high values"],
        value="Version 1 — Linear",
        label="Height interpretation version",
    )
    texture_strength = gr.Slider(
        minimum=0.0,
        maximum=0.5,
        value=DEFAULT_TEXTURE_STRENGTH,
        step=0.001,
        label="Fine Detail Strength",
        info="Adds zero-mean fine source detail. Default uses strength 0.1 and smoothing 5.",
    )

    rerun_button = gr.Button("Run / Rerun Preview", variant="primary")
    run_status = gr.Textbox(
        value="Upload an image, choose a version, then click Run / Rerun Preview.",
        label="Run status",
        interactive=False,
    )

    with gr.Row():
        height_map_output = gr.Image(
            format="png",
            image_mode="L",
            label="True grayscale height map",
            interactive=False,
        )
        hillshade_output = gr.Image(
            format="png",
            image_mode="L",
            label="Embossed surface preview",
            interactive=False,
        )

    test_inputs = [input_image, version, texture_strength]
    test_outputs = [height_map_output, hillshade_output, run_status]
    input_image.change(
        fn=prepare_public_test_outputs,
        inputs=test_inputs,
        outputs=test_outputs,
        show_progress="full",
    )
    rerun_button.click(
        fn=prepare_public_test_outputs,
        inputs=test_inputs,
        outputs=test_outputs,
        show_progress="full",
    )


if __name__ == "__main__":
    demo.queue().launch(
        server_name=LAUNCH_ARGS.server_name,
        server_port=LAUNCH_ARGS.server_port,
        theme="base",
    )