"""Gradio front end for the Sking image -> Minecraft skin pipeline.

Startup order is load-bearing:
  0. import spaces          installs ZeroGPU's CUDA emulation shim -- MUST be
                            before anything touches torch, or `.to("cuda")`
                            dies with "No CUDA GPUs are available".
  1. build + launch Gradio  immediately, so the container binds port 7860 well
                            inside HF's health-check window.
  2. bootstrap in a thread  downloads ~19 GB and verifies SHA-256 in the
                            background. Blocking startup on this is what caused
                            RUNTIME_ERROR with no traceback: the platform killed
                            the container before it ever became healthy.
"""

# --- 0. `spaces` MUST be imported before torch ------------------------------
try:
    import spaces  # type: ignore

    def gpu(duration: int):
        return spaces.GPU(duration=duration)

    ZERO_GPU = True
except ImportError:  # dedicated-GPU Docker Space, or a local run
    def gpu(duration: int):
        def identity(fn):
            return fn
        return identity

    ZERO_GPU = False

import base64  # noqa: E402
import os  # noqa: E402
import html  # noqa: E402
import tempfile  # noqa: E402
import threading  # noqa: E402
import traceback  # noqa: E402
from pathlib import Path  # noqa: E402

import gradio as gr  # noqa: E402
from PIL import Image  # noqa: E402

# `weights` first: it owns HF_HOME, which huggingface_hub.constants freezes at
# import time. sking_pipeline/__init__.py now guarantees this regardless, but
# keep the order explicit so the intent survives a future "tidy up imports".
from sking_pipeline import weights  # noqa: E402,F401  (isort: skip)
from sking_pipeline import stage1_flux, stage2_uv  # noqa: E402

# ZeroGPU platform caps, measured live + documented:
#   * a call RESERVES 1.5x the declared duration against the caller's daily
#     quota before anything runs (225s requested for duration=150, etc.);
#   * a FREE account may not declare more than 120s per call -- duration=300
#     raises "ZeroGPU illegal duration ... larger than the maximum allowed";
#     only PRO (1500s/day) or dedicated hardware may exceed it;
#   * daily quota: ~60-180s anonymous, 300s signed-in free, 1500s PRO.
# Production parity (entropydrop.com sends n_step=100, guidance=4, 1024px)
# costs ~160s of GPU on the measured free path (~1.48 s/step @1024 + stage2),
# so the single-call PRO path (300s) still owns true 100-step parity. Free gets
# closer by SPLITTING Flux and stage 2 into two 120s GPU calls so Flux keeps
# the whole free envelope for itself (~70 steps @1024).
GPU_SECONDS = int(os.environ.get("SKING_GPU_SECONDS", "120"))            # free Flux call
GPU_SECONDS_PROD = int(os.environ.get("SKING_GPU_SECONDS_PROD", "300"))  # PRO one-shot
STAGE2_SECONDS = int(os.environ.get("SKING_STAGE2_SECONDS", "60"))       # free stage-2 call

# Live zero-a10g measurement 2026-09-14 (free button log):
#   33 steps @1024 → 48.5 s wall  ⇒  ~1.48 s/step  (fixed overhead absorbed)
# Prior model of 1.25 s/step @768 scaled by (1024/768)^2 = 2.22 s/step was
# ~50% pessimistic and clamped free far below what the GPU actually delivers.
# Recalibrate to measured 1024 rate; scale by area for other sizes. Keep a
# small fixed cost + stage-2 estimate so the clamp still has headroom.
_FLUX_FIXED_S = float(os.environ.get("SKING_FLUX_FIXED_S", "4"))
_FLUX_PER_STEP_S_1024 = float(os.environ.get("SKING_FLUX_PER_STEP_S", "1.50"))
_STAGE2_EST_S = float(os.environ.get("SKING_STAGE2_EST_S", "25"))
# Free path: Flux alone owns the 120s call (stage 2 is a separate call).
# Use 0.90 of the envelope so a slow first-of-session step doesn't OOM the
# reservation; the kill still burns quota.
_BUDGET_FRACTION = float(os.environ.get("SKING_BUDGET_FRACTION", "0.90"))


def _per_step_s(size: int) -> float:
    """Seconds per Flux step at `size`×`size`, calibrated at 1024."""
    return _FLUX_PER_STEP_S_1024 * (size / 1024.0) ** 2


def _fit_gpu_budget(
    steps: int,
    size: int,
    budget_s: int,
    clamp: bool,
    *,
    include_stage2: bool = True,
):
    """Optionally clamp `steps` so the run fits the declared GPU duration.

    `clamp=False` is the production-parity path: runs exactly what was asked
    under the (PRO/paid-only) 300s declaration.

    `include_stage2=False` is the free split path: Flux owns the whole 120s
    envelope and stage 2 runs as a second @spaces.GPU call, so we do not
    subtract stage-2 time from the Flux budget. That is what pushes free from
    ~33 → ~70 steps @1024.
    """
    if not clamp:
        return steps, size, ""
    reserved = _STAGE2_EST_S if include_stage2 else 0.0
    budget = budget_s * _BUDGET_FRACTION - _FLUX_FIXED_S - reserved
    per_step = _per_step_s(size)
    if per_step <= 0:
        return steps, size, ""
    max_steps = int(budget / per_step)
    if steps <= max_steps:
        return steps, size, ""
    clamped = max(8, max_steps)
    path = "single-call" if include_stage2 else "split (Flux-only envelope)"
    note = (
        f" · ⚠ steps clamped {steps}→{clamped} for free ZeroGPU {path} "
        f"(budget {budget_s}s, ~{per_step:.2f}s/step @ {size}px). "
        f"True production parity is 100 steps @1024 on the PRO button."
    )
    return clamped, size, note
# ---------------------------------------------------------------------------
# background bootstrap
# ---------------------------------------------------------------------------

STATE = {
    "stage": "starting",
    "log": [],
    "ready": False,
    "stage1": False,
    "boot": None,
    "error": None,
    "last_error": None,
}
_LOCK = threading.Lock()


def _progress(msg: str) -> None:
    if STATE.get("halt_logging"):
        return
    with _LOCK:
        STATE["log"].append(msg)
        if len(STATE["log"]) > 200:
            del STATE["log"][:100]
    print(msg, flush=True)


def _bootstrap() -> None:
    """Download, verify and load the models -- at startup, on the main thread.

    Both constraints are hard and they pull in opposite directions:

    * ZeroGPU's CUDA emulation only intercepts `.to("cuda")` during *startup*,
      before the app begins serving. Loading later -- from a worker thread, or
      after `launch()` -- falls through to the real `torch._C._cuda_init`, which
      the `spaces` patch refuses: "Low-level CUDA init reached ... emulation mode
      did not intercept a CUDA operation".
    * So the weights must be resident before `launch()` is called.

    Downloads are fast on HF's Xet cache (~900 MB/s), so ~19 GB costs seconds;
    the model materialisation dominates. Failure here is recorded, not fatal:
    stage 2 still serves.
    """
    try:
        boot = weights.bootstrap(progress=_progress)
    except BaseException:  # noqa: BLE001
        with _LOCK:
            STATE["error"] = traceback.format_exc()
            STATE["stage"] = "error"
        print(STATE["error"], flush=True)
        return

    with _LOCK:
        STATE["boot"] = boot
        # Stage 2 needs nothing but the verified bundle, so publish readiness
        # here. A stage-1 failure must not take the working half of the app
        # down with it -- that bug left the Render tab reporting "still
        # loading" even though all 127 files had verified.
        STATE["ready"] = True

    if boot.get("stage1_viable") and boot.get("flux"):
        try:
            _progress("loading FLUX.2 [klein] 4B Base during startup")
            stage1_flux.load(boot["flux"], boot["lora"], progress=_progress)
            with _LOCK:
                STATE["stage1"] = True
            _progress("stage 1 ready")
        except BaseException:  # noqa: BLE001 - stage 2 keeps working
            with _LOCK:
                STATE["error"] = traceback.format_exc()
            print(STATE["error"], flush=True)
    else:
        _progress(f"stage 1 skipped: {boot.get('stage1_reason')}")

    with _LOCK:
        STATE["stage"] = "ready"
    _progress("READY")


def _bundle():
    boot = STATE.get("boot")
    if not boot:
        raise gr.Error("Models are still downloading — see the status panel.")
    return boot


def _require_stage1():
    if not STATE["ready"]:
        raise gr.Error("Models are still downloading — see the status panel.")
    if not STATE["stage1"]:
        raise gr.Error(
            "Stage 1 (FLUX.2 [klein] 4B) is not available on this hardware. "
            "Use the 'Render → Skin' tab, which runs the identical production "
            "UV parser on a render you supply."
        )


def _record_endpoint_error(where: str) -> None:
    """Capture an endpoint traceback and surface it to the *caller*.

    Two channels are needed because both passive ones fail here:
      * HF's run-log endpoint is a rolling SSE window saturated by Xet progress
        bars, and errorMessage is only a tail slice -- both truncated the real
        exception mid-frame, repeatedly.
      * ZeroGPU executes @spaces.GPU tasks in a forked worker, so writes to our
        module-level STATE never reach the parent process that serves
        /status_md. Recording there looked like "nothing was captured".
    So: print for the log, store for the panel, and re-raise as gr.Error whose
    message is the traceback -- that one always reaches the client.
    """
    tb = f"[{where}]\n" + traceback.format_exc()
    with _LOCK:
        STATE["last_error"] = tb
    print(tb, flush=True)
    raise gr.Error(tb[-1800:])


def status_md() -> str:
    with _LOCK:
        stage, err, log = STATE["stage"], STATE["error"], list(STATE["log"])
        s1, ready = STATE["stage1"], STATE["ready"]
        last = STATE.get("last_error")
    icon = {"starting": "⏳", "ready": "✅", "error": "❌"}.get(stage, "•")
    tail = "\n".join(f"    {line}" for line in log[-14:])
    head = (
        f"### {icon} Status: `{stage}`\n\n"
        f"- stage 1 (Flux + LoRA): **{'available' if s1 else 'not loaded'}**\n"
        f"- stage 2 (SKING_DDJ_v104): **{'available' if ready else 'loading'}**\n"
        f"- ZeroGPU: **{'yes' if ZERO_GPU else 'no (dedicated/local)'}**\n"
    )
    if err:
        head += f"\n**Startup error**\n\n```\n{err.strip()[-2000:]}\n```\n"
    if last:
        head += f"\n**Last request error**\n\n```\n{last.strip()[-2500:]}\n```\n"
    return head + f"\n<details><summary>startup log</summary>\n\n```\n{tail}\n```\n</details>"


# ---------------------------------------------------------------------------
# helpers
# ---------------------------------------------------------------------------

_OUT_DIR = Path(tempfile.gettempdir()) / "sking-out"
_OUT_DIR.mkdir(parents=True, exist_ok=True)
SKIN_OUT = _OUT_DIR / "skin.png"
RENDER_OUT = _OUT_DIR / "render.png"


def _upscale(img: Image.Image, factor: int = 8) -> Image.Image:
    w, h = img.size
    return img.resize((w * factor, h * factor), Image.NEAREST)


def _empty_viewer() -> str:
    return (
        '<div style="height:380px;background:#141414;border-radius:8px;display:flex;'
        'align-items:center;justify-content:center;color:#777;font:13px/1.5 '
        'ui-sans-serif,system-ui,sans-serif;">3D preview appears here after you '
        'generate a skin</div>'
    )


def _viewer_iframe(skin_b64: str) -> str:
    """3D preview via skinview3d inside an iframe srcdoc.

    Gradio updates gr.HTML through innerHTML, and browsers do not execute
    <script> inserted that way; a srcdoc document is parsed normally, so its
    scripts do run. Degrades to a blank box if the CDN is unreachable.
    """
    if not skin_b64:
        return _empty_viewer()

    doc = f"""<!doctype html>
<html><head><meta charset="utf-8"><style>
  html,body{{margin:0;padding:0;background:#141414;height:100%;overflow:hidden}}
  #c{{width:100vw;height:100vh}}
  #e{{color:#f88;font:12px/1.5 ui-monospace,monospace;padding:12px}}
</style></head>
<body><div id="c"></div><div id="e"></div>
<script type="module">
try {{
  const s = await import("https://cdn.jsdelivr.net/npm/skinview3d@3.4.2/+esm");
  const host = document.getElementById("c");
  const v = new s.SkinViewer({{width: host.clientWidth, height: host.clientHeight}});
  v.animation = new s.IdleAnimation();
  await v.loadSkin("data:image/png;base64,{skin_b64}");
}} catch (err) {{
  document.getElementById("e").textContent = "3D preview unavailable: " + err;
}}
</script></body></html>"""

    return (
        '<iframe srcdoc="' + html.escape(doc, quote=True) + '" '
        'style="width:100%;height:380px;border:0;border-radius:8px;background:#141414;" '
        'sandbox="allow-scripts"></iframe>'
    )


def _as_pil(x):
    """Coerce whatever Gradio hands us into a PIL Image.

    `gr.Image(type="pil")` normally preprocesses to a PIL Image, but API clients
    (gradio_client / curl) can surface a path, a dict, or a file-like instead --
    and `img.size` on a str is an AttributeError with no useful context.
    """
    if x is None:
        return None
    if isinstance(x, Image.Image):
        return x
    if isinstance(x, (str, Path)):
        return Image.open(x)
    if isinstance(x, dict):
        ref = x.get("path") or x.get("url") or x.get("name")
        if ref:
            return Image.open(ref)
    name = getattr(x, "name", None)
    if isinstance(name, (str, Path)):
        return Image.open(name)
    raise gr.Error(f"Unsupported image input: {type(x).__name__}")


def _validate_square(img: Image.Image) -> Image.Image:
    """Coerce an upload into the v104 contract: even square, 256..4096 px."""
    w, h = img.size
    side = min(w, h)
    left, top = (w - side) // 2, (h - side) // 2
    sq = img.crop((left, top, left + side, top + side))
    if side % 2:
        side -= 1
        sq = sq.crop((0, 0, side, side))
    side = max(256, min(side, 4096))
    if sq.size[0] != side:
        sq = sq.resize((side, side), Image.LANCZOS)
    return sq.convert("RGB")


# ---------------------------------------------------------------------------
# pipeline
# ---------------------------------------------------------------------------


def _pack_outputs(render: Image.Image, steps: int, guidance: float,
                  seed: int, clamp_note: str, path_tag: str):
    """Shared post-stage-2 packaging: skin file + 8x preview + 3D viewer."""
    skin = Image.open(SKIN_OUT).convert("RGBA")
    skin.load()
    with open(SKIN_OUT, "rb") as fh:
        skin_b64 = base64.b64encode(fh.read()).decode()
    info = (
        f"64×64 RGBA · {SKIN_OUT.stat().st_size} bytes · "
        f"seed {int(seed)} · {steps} steps · guidance {float(guidance):.1f} · "
        f"SKING_DDJ_v104 · {path_tag}{clamp_note}"
    )
    return render, _upscale(skin, 8), str(SKIN_OUT), info, _viewer_iframe(skin_b64)


@gpu(GPU_SECONDS)
def _stage1_free(image, seed, steps, guidance, size):
    """Free-tier Flux only. Owns the full 120s envelope (no stage 2).

    Returns (render, steps_used, guidance, seed, clamp_note) so the chained
    stage-2 call — which may run in a *different* ZeroGPU worker with no
    shared /tmp or STATE — gets everything it needs through Gradio's
    serialised outputs rather than the filesystem.
    """
    try:
        image = _as_pil(image)
        _require_stage1()
        if image is None:
            raise gr.Error("Upload a reference image first.")
        steps, size, clamp_note = _fit_gpu_budget(
            int(steps), int(size), GPU_SECONDS, True, include_stage2=False,
        )
        _progress(
            f"stage1_free type={type(image).__name__} steps={steps} "
            f"size={size} budget={GPU_SECONDS}s"
        )
        render = stage1_flux.generate(
            image, seed=int(seed), steps=steps,
            guidance=float(guidance), size=size,
        )
        # Best-effort local cache (useful on dedicated GPU / same-worker runs).
        try:
            render.save(RENDER_OUT, format="PNG")
        except Exception:  # noqa: BLE001
            pass
        return render, int(steps), float(guidance), int(seed), clamp_note
    except gr.Error:
        raise
    except Exception:
        _record_endpoint_error("stage1_free")


@gpu(STAGE2_SECONDS)
def _stage2_free(render, steps, guidance, seed, clamp_note):
    """Free-tier stage 2. Takes the render + knobs from `_stage1_free` via Gradio."""
    try:
        if not STATE["ready"]:
            raise gr.Error(
                "Stage 2 is still loading its 2.58 GB bundle — check the status "
                "panel and refresh."
            )
        render = _as_pil(render)
        if render is None:
            raise gr.Error(
                "Stage 1 returned no render — try Generate again, or use the "
                "Render → Skin tab with a saved front/back sheet."
            )
        boot = _bundle()
        # Write the Gradio-delivered render to a worker-local path; do NOT rely
        # on RENDER_OUT from stage 1 (ZeroGPU workers do not share /tmp).
        local_render = _OUT_DIR / f"free_render_{os.getpid()}.png"
        _validate_square(render).save(local_render, format="PNG")
        _progress(
            f"stage2_free pid={os.getpid()} render={local_render} "
            f"size={local_render.stat().st_size} steps={steps}"
        )
        stage2_uv.render_to_uv(
            local_render, SKIN_OUT, boot["bundle"], release=boot.get("release")
        )
        try:
            local_render.unlink(missing_ok=True)
        except Exception:  # noqa: BLE001
            pass
        steps_i = int(steps) if steps is not None else 0
        guidance_f = float(guidance) if guidance is not None else 4.0
        seed_i = int(seed) if seed is not None else 0
        note = clamp_note if isinstance(clamp_note, str) else ""
        return _pack_outputs(
            render.convert("RGB") if render.mode != "RGB" else render,
            steps_i, guidance_f, seed_i, note,
            path_tag=f"free-split · {steps_i}st",
        )
    except gr.Error:
        raise
    except Exception:
        _record_endpoint_error("stage2_free")


@gpu(GPU_SECONDS)
def image_to_skin(image, seed, steps, guidance, size):
    """Free-tier single-call fallback (API clients / old bookmarks).

    UI free button uses the split path (`_stage1_free` → `_stage2_free`) which
    keeps ~70 steps @1024. This oneshot still works but clamps harder (~45
    @1024) because Flux and stage 2 share one 120s envelope.
    """
    try:
        return _image_to_skin_oneshot(
            image, seed, steps, guidance, size,
            budget_s=GPU_SECONDS, clamp=True, path_tag="free-oneshot",
        )
    except gr.Error:
        raise
    except Exception:
        _record_endpoint_error("image_to_skin")


@gpu(GPU_SECONDS_PROD)
def image_to_skin_production(image, seed, steps, guidance, size):
    """Production parity: exactly what entropydrop.com sends (100 steps @1024).

    Measured cost ~160s GPU (100 × 1.48 s/step @1024 + stage 2), reserved 450s
    under the 1.5x rule -- beyond BOTH the free per-call cap (120s) and the free
    daily quota (300s). Free accounts are refused by the platform with
    "ZeroGPU illegal duration"; PRO (1500s/day) and dedicated-GPU Spaces run it
    unclamped.
    """
    try:
        return _image_to_skin_oneshot(
            image, seed, steps, guidance, size,
            budget_s=GPU_SECONDS_PROD, clamp=False, path_tag="production-parity",
        )
    except gr.Error:
        raise
    except Exception:
        _record_endpoint_error("image_to_skin_production")


def _image_to_skin_oneshot(image, seed, steps, guidance, size, budget_s, clamp,
                           path_tag: str):
    """Body of a single @spaces.GPU call doing Flux + stage 2.

    Called FROM an already-decorated endpoint (image_to_skin /
    image_to_skin_production). Do not decorate this helper itself -- nesting
    @spaces.GPU is undefined.
    """
    _progress(f"image_to_skin_oneshot type={type(image).__name__} "
              f"budget={budget_s}s clamp={clamp} tag={path_tag}")
    image = _as_pil(image)
    _require_stage1()
    if image is None:
        raise gr.Error("Upload a reference image first.")
    boot = _bundle()

    steps, size, clamp_note = _fit_gpu_budget(
        int(steps), int(size), budget_s, clamp, include_stage2=True,
    )
    render = stage1_flux.generate(
        image, seed=int(seed), steps=steps,
        guidance=float(guidance), size=size,
    )
    render.save(RENDER_OUT, format="PNG")
    stage2_uv.render_to_uv(
        RENDER_OUT, SKIN_OUT, boot["bundle"], release=boot.get("release")
    )
    return _pack_outputs(render, steps, guidance, seed, clamp_note, path_tag)

@gpu(STAGE2_SECONDS)
def render_to_skin_only(render):
    """Feed a combined front/back render straight into v104."""
    try:
        return _render_to_skin_impl(render)
    except gr.Error:
        raise
    except Exception:
        _record_endpoint_error("render_to_skin_only")


def _render_to_skin_impl(render):
    _progress(f"render_to_skin input type={type(render).__name__}")
    render = _as_pil(render)
    if not STATE["ready"]:
        raise gr.Error(
            "Stage 2 is still loading its 2.58 GB bundle — check the status "
            "panel and refresh. This usually takes 1-3 minutes."
        )
    if render is None:
        raise gr.Error("Upload a combined front/back render (even square, 256-4096 px).")
    boot = _bundle()

    src = _OUT_DIR / "debug_render.png"
    _validate_square(render).save(src, format="PNG")
    out = _OUT_DIR / "debug_skin.png"
    stage2_uv.render_to_uv(src, out, boot["bundle"], release=boot.get("release"))

    skin = Image.open(out).convert("RGBA")
    skin.load()
    with open(out, "rb") as fh:
        skin_b64 = base64.b64encode(fh.read()).decode()
    return _upscale(skin, 8), str(out), _viewer_iframe(skin_b64)


# ---------------------------------------------------------------------------
# UI
# ---------------------------------------------------------------------------

_CSS = """
footer { display: none !important; }
.gradio-container { max-width: 1180px !important; }
"""

_CREDITS = """
---
### Credits and license

Model, dataset and pipeline © **EntropyDrop** (Shanghai EqualRank Technology
Co., Ltd.) — [entropydrop.com](https://entropydrop.com) ·
[huggingface.co/EntropyDrop](https://huggingface.co/EntropyDrop) ·
[github.com/EntropyDrop](https://github.com/EntropyDrop)

- `EntropyDrop/Sking` and `SKING_DDJ_v104` — **AGPL-3.0**. This Space is
  therefore AGPL-3.0 too; its complete source is this public repository.
- `black-forest-labs/FLUX.2-klein-base-4B` — **Apache-2.0**.
- BiRefNet and SigLIP2 keep their own licenses; see `NOTICE.md`.
- The images in `testdata/` are real entropydrop.com generations, **CC BY-NC 4.0**.

Not affiliated with or endorsed by EntropyDrop, Mojang or Microsoft.
Uploads are processed in memory and are not stored or used for training.
"""

# theme/css go to launch(), not here: Gradio 6 relocated them and warns loudly.
with gr.Blocks(title="Sking — Image to Minecraft Skin") as demo:
    gr.Markdown(
        """
        # Sking · image → Minecraft skin

        An open reimplementation of **EntropyDrop**'s two-stage pipeline.

        **Stage 1** — FLUX.2 [klein] 4B Base + the `sking_v73` LoRA turns a
        reference image into a combined front/back character render.
        **Stage 2** — `SKING_DDJ_v104` (BiRefNet foreground + Dense UV Parser)
        converts that render into a real **64×64 RGBA** skin with inner and
        outer layers, ready to drop into Minecraft.

        > ⏳ **Models download in the background on first start** (~19 GB, 1–4
        > min). Watch the status panel below and hit *Refresh* — the UI is live
        > before the weights are.
        >
        > ⚖️ **Two generate buttons, because true 100-step parity is a paid
        > feature.** entropydrop.com runs **100 steps @ 1024**. Free ZeroGPU
        > forbids declaring more than **120 s per call**. The free button
        > **splits** Flux and stage 2 into two GPU calls so Flux keeps the
        > whole 120 s envelope — measured ceiling **~70 steps @1024** /
        > **~100 steps @768** (was ~33 before the split). The *production
        > parity* button declares 300 s and runs 100 steps unclamped (PRO or
        > a paid GPU Space only).
        """
    )

    status = gr.Markdown(status_md())
    refresh = gr.Button("Refresh status", size="sm")
    refresh.click(fn=status_md, inputs=None, outputs=status)

    with gr.Tab("Image → Skin"):
        with gr.Row():
            with gr.Column(scale=5):
                inp = gr.Image(type="pil", label="Reference image", height=340)
                seed = gr.Number(value=42, label="Seed", precision=0)
                # Defaults mirror what entropydrop.com actually sends: the HAR
                # shows n_step=100 and guidance=4.0 in all 185 captured records,
                # and 1024px renders on the two-stage path.
                steps = gr.Slider(
                    20, 100, value=100, step=5,
                    label="Inference steps (production: 100)",
                )
                guidance = gr.Slider(
                    1.0, 8.0, value=4.0, step=0.1,
                    label="Guidance scale (production: 4.0)",
                )
                size = gr.Slider(
                    768, 1024, value=1024, step=64,
                    label="Render size (production: 1024)",
                    info="1024 = what production emits. 768 is the v73 LoRA's "
                         "published training setting and ~1.8x cheaper.",
                )
                with gr.Row():
                    run_prod = gr.Button(
                        "Generate · production parity\n100 steps @ 1024 · needs PRO / paid GPU",
                        variant="primary",
                    )
                    run_free = gr.Button(
                        "Generate · free-tier (split)\n~70 steps @1024 · two 120s GPU calls",
                        variant="secondary",
                    )
            with gr.Column(scale=7):
                out_preview = gr.Image(
                    type="pil", label="Skin (64×64, upscaled 8×, nearest)"
                )
                out_render = gr.Image(
                    type="pil", label="Stage 1 · combined front/back render"
                )
                out_file = gr.File(label="skin.png (64×64 RGBA) — download this")
                out_info = gr.Textbox(label="Output", interactive=False)
                out_3d = gr.HTML(_empty_viewer(), label="3D preview")

        _IO = [inp, seed, steps, guidance, size]
        _OUT = [out_render, out_preview, out_file, out_info, out_3d]
        # PRO: single 300s GPU call, unclamped (platform refuses this on free).
        run_prod.click(fn=image_to_skin_production, inputs=_IO, outputs=_OUT).then(
            fn=status_md, inputs=None, outputs=status
        )
        # FREE: chain two @spaces.GPU calls. Flux returns the render + the
        # resolved knobs under its own 120s envelope (~70 steps @1024); stage 2
        # consumes those Gradio-serialised values under a separate 60s envelope.
        # Handing state through Gradio (not /tmp or module STATE) is load-
        # bearing: ZeroGPU forks a fresh worker per @spaces.GPU call.
        # Hidden components carry the non-image knobs between the two calls.
        _h_steps = gr.State(0)
        _h_guidance = gr.State(4.0)
        _h_seed = gr.State(0)
        _h_note = gr.State("")
        run_free.click(
            fn=_stage1_free,
            inputs=_IO,
            outputs=[out_render, _h_steps, _h_guidance, _h_seed, _h_note],
        ).then(
            fn=_stage2_free,
            inputs=[out_render, _h_steps, _h_guidance, _h_seed, _h_note],
            outputs=_OUT,
        ).then(
            fn=status_md, inputs=None, outputs=status
        )

    with gr.Tab("Render → Skin"):
        gr.Markdown(
            "Runs **stage 2 only** — EntropyDrop's `SKING_DDJ_v104` UV parser, "
            "byte-identical to their production release. Input must be an **even "
            "square between 256 and 4096 px**, left half = front view, right "
            "half = back view (their `front_left` / `back_left` layout). "
            "Uploads are centre-cropped to a square.\n\n"
            "This tab works even when stage 1 is unavailable, and the examples "
            "below are **real renders captured from entropydrop.com**, so you "
            "can compare directly against their published output."
        )
        with gr.Row():
            with gr.Column(scale=5):
                d_in = gr.Image(type="pil", label="Combined front/back render", height=320)
                d_run = gr.Button("Parse to skin", variant="primary")
            with gr.Column(scale=7):
                d_preview = gr.Image(type="pil", label="Skin (upscaled 8×)")
                d_file = gr.File(label="skin.png (64×64 RGBA)")
                d_3d = gr.HTML(_empty_viewer(), label="3D preview")

        _fixtures = sorted(Path("testdata").glob("*_render.png"))
        if _fixtures:
            gr.Examples(
                examples=[[str(p)] for p in _fixtures],
                inputs=[d_in],
                label=(
                    "Real production renders from entropydrop.com "
                    "(CC BY-NC 4.0 — see testdata/README.md)"
                ),
            )
        d_run.click(
            fn=render_to_skin_only, inputs=[d_in], outputs=[d_preview, d_file, d_3d]
        ).then(fn=status_md, inputs=None, outputs=status)

    gr.Markdown(_CREDITS)


print(f"[sking] module executed as __name__={__name__!r} ZERO_GPU={ZERO_GPU}",
      flush=True)


def _launch() -> None:
    """Bind 7860. Gradio 6.27 dropped `show_api` from launch(), and `spaces`
    monkeypatches `gr.Blocks.launch` to forward kwargs verbatim, so an
    unsupported one is a fatal TypeError. Walk richest -> bare minimum; each
    rejection happens before binding, so retrying is safe.
    """
    print("[sking] calling demo.launch()", flush=True)
    try:
        demo.queue(default_concurrency_limit=1)
    except TypeError:
        demo.queue()

    base = dict(server_name="0.0.0.0", server_port=7860)
    attempts = [
        {**base, "show_api": False, "theme": gr.themes.Soft(), "css": _CSS},
        {**base, "theme": gr.themes.Soft(), "css": _CSS},
        {**base, "show_api": False},
        dict(base),
    ]
    last = None
    for kwargs in attempts:
        try:
            demo.launch(**kwargs)
            print(f"[sking] demo.launch() returned (accepted {sorted(kwargs)})",
                  flush=True)
            return
        except TypeError as exc:
            last = exc
            print(f"[sking] launch rejected {sorted(kwargs)}: {exc}", flush=True)
    raise last


_bootstrap()
_launch()