GravityLLM-Space-Demo/app.py

from __future__ import annotations

import base64
import json
import os
import tempfile
from pathlib import Path
from typing import Any, Dict

import gradio as gr
from huggingface_hub import InferenceClient
from huggingface_hub.errors import HfHubHTTPError, InferenceTimeoutError

from utils.scene_tools import (
    SCHEMA,
    extract_first_json_block,
    heuristic_scene,
    parse_json_text,
    plot_scene,
    scene_markdown,
    scene_table,
    validate_scene,
)

ROOT = Path(__file__).resolve().parent
ASSETS = ROOT / "assets"
EXAMPLES = ROOT / "examples"

APP_TITLE = "GravityLLM"
DEFAULT_MODEL_ID = os.getenv("GRAVITYLLM_MODEL_ID", "your-namespace/GravityLLM-AutoPosition")
DEFAULT_BACKEND = os.getenv("GRAVITYLLM_BACKEND", "hybrid")
HF_TOKEN = os.getenv("HF_TOKEN") or os.getenv("HUGGINGFACEHUB_API_TOKEN")

SYSTEM_PREFIX = (
    "You are GravityLLM (Spatial9 AutoPosition SLM). "
    "Generate ONLY valid JSON matching the Spatial9Scene schema. "
    "No markdown. No explanation. No code fences.\n\n"
)

EXAMPLE_FILES = {
    "Club Drop": EXAMPLES / "club_drop.json",
    "Cinematic Break": EXAMPLES / "cinematic_break.json",
    "Podcast Voice": EXAMPLES / "podcast_voice.json",
}


def logo_data_uri() -> str:
    logo_bytes = (ASSETS / "spatial9_logo.png").read_bytes()
    return "data:image/png;base64," + base64.b64encode(logo_bytes).decode("utf-8")


def load_example(name: str) -> str:
    path = EXAMPLE_FILES.get(name, next(iter(EXAMPLE_FILES.values())))
    return path.read_text(encoding="utf-8")


def build_prompt(payload: Dict[str, Any]) -> str:
    return SYSTEM_PREFIX + "INPUT:\n" + json.dumps(payload, ensure_ascii=False, indent=2) + "\nOUTPUT:\n"


def remote_generate(
    payload: Dict[str, Any],
    model_id: str,
    temperature: float,
    top_p: float,
    max_new_tokens: int,
    use_grammar: bool,
) -> tuple[Dict[str, Any], str]:
    prompt = build_prompt(payload)
    client = InferenceClient(model=model_id, token=HF_TOKEN)

    call_kwargs = dict(
        prompt=prompt,
        model=model_id,
        max_new_tokens=max_new_tokens,
        temperature=temperature,
        top_p=top_p,
        repetition_penalty=1.05,
        return_full_text=False,
    )
    if use_grammar:
        call_kwargs["grammar"] = {"type": "json", "value": SCHEMA}

    try:
        response = client.text_generation(**call_kwargs)
        scene = parse_json_text(response)
        return scene, f"remote-model ({model_id})"
    except Exception as first_error:
        if use_grammar:
            try:
                call_kwargs.pop("grammar", None)
                response = client.text_generation(**call_kwargs)
                scene = parse_json_text(response)
                return scene, f"remote-model ({model_id}, grammar-fallback)"
            except Exception as second_error:
                raise RuntimeError(f"{type(first_error).__name__}: {first_error}\n\nFallback: {type(second_error).__name__}: {second_error}") from second_error
        raise


def write_download_file(scene: Dict[str, Any]) -> str:
    fd, path = tempfile.mkstemp(prefix="gravityllm_scene_", suffix=".json")
    os.close(fd)
    Path(path).write_text(json.dumps(scene, ensure_ascii=False, indent=2), encoding="utf-8")
    return path


def generate_scene(
    payload_text: str,
    model_id: str,
    backend: str,
    temperature: float,
    top_p: float,
    max_new_tokens: int,
    use_grammar: bool,
):
    try:
        payload = parse_json_text(payload_text)
    except Exception as exc:
        msg = f"### Invalid input JSON\n\n- {type(exc).__name__}: {exc}"
        return "", msg, None, [], None, "Fix the input JSON and try again."

    backend = backend or DEFAULT_BACKEND
    model_id = model_id.strip() or DEFAULT_MODEL_ID

    scene = None
    backend_used = "rules-engine demo"
    status = "Scene generated."

    if backend in {"remote-model", "hybrid"}:
        try:
            scene, backend_used = remote_generate(payload, model_id, temperature, top_p, max_new_tokens, use_grammar)
            status = f"Generated from remote model: {model_id}"
        except (InferenceTimeoutError, HfHubHTTPError, RuntimeError, ValueError, json.JSONDecodeError) as exc:
            if backend == "remote-model":
                msg = f"### Remote generation failed\n\n- {type(exc).__name__}: {exc}"
                return "", msg, None, [], None, "Remote inference failed."
            scene = heuristic_scene(payload)
            backend_used = "rules-engine demo (remote fallback)"
            status = f"Remote generation failed; heuristic fallback used. Details: {type(exc).__name__}: {exc}"

    if scene is None:
        scene = heuristic_scene(payload)

    valid, errors = validate_scene(scene)
    download_path = write_download_file(scene)
    figure = plot_scene(scene)
    table = scene_table(scene)
    summary = scene_markdown(scene, valid, errors, backend_used)
    return json.dumps(scene, ensure_ascii=False, indent=2), summary, figure, table, download_path, status


def validate_only(scene_text: str):
    try:
        scene = parse_json_text(scene_text)
    except Exception as exc:
        return f"### Invalid scene JSON\n\n- {type(exc).__name__}: {exc}", None, []
    valid, errors = validate_scene(scene)
    summary = scene_markdown(scene, valid, errors, "manual validation")
    return summary, plot_scene(scene), scene_table(scene)


def build_payload(target_format, style, section, bpm, energy, max_objects):
    payload = {
        "target_format": target_format,
        "max_objects": int(max_objects),
        "style": style,
        "section": section,
        "global": {"bpm": int(bpm), "energy": float(energy)},
        "stems": [
            {"id": "lead", "class": "lead_vocal", "lufs": -17.0, "transient": 0.25, "band_energy": {"low": 0.08, "mid": 0.67, "high": 0.25}, "leadness": 0.96},
            {"id": "kick", "class": "kick", "lufs": -10.6, "transient": 0.96, "band_energy": {"low": 0.82, "mid": 0.12, "high": 0.06}, "leadness": 0.22},
            {"id": "bass", "class": "bass", "lufs": -12.5, "transient": 0.58, "band_energy": {"low": 0.86, "mid": 0.10, "high": 0.04}, "leadness": 0.30},
            {"id": "pad", "class": "pad", "lufs": -21.5, "transient": 0.05, "band_energy": {"low": 0.20, "mid": 0.50, "high": 0.30}, "leadness": 0.08},
            {"id": "fx", "class": "fx", "lufs": -24.0, "transient": 0.22, "band_energy": {"low": 0.10, "mid": 0.24, "high": 0.66}, "leadness": 0.04},
        ],
        "rules": [
            {"type": "anchor", "track_class": "lead_vocal", "az_deg": 0, "el_deg": 10, "dist_m": 1.6},
            {"type": "mono_low_end", "hz_below": 120},
            {"type": "width_pref", "track_class": "pad", "min_width": 0.75},
        ],
    }
    return json.dumps(payload, ensure_ascii=False, indent=2)


hero_html = f"""
<div class="hero-wrap">
  <div class="hero-left">
    <div class="hero-logo-card">
      <img class="hero-logo" src="{logo_data_uri()}" alt="Spatial9 logo"/>
    </div>
  </div>
  <div class="hero-right">
    <div class="eyebrow">SPATIAL9 • HUGGING FACE SPACE</div>
    <h1>GravityLLM Studio</h1>
    <p class="hero-copy">
      Constraint-conditioned immersive scene generation with schema-guided JSON output,
      remote Hugging Face inference, heuristic fallback, and a live spatial preview.
    </p>
    <div class="hero-chips">
      <span>IAMF Ready</span>
      <span>Schema Validated</span>
      <span>Spatial Preview</span>
      <span>Branded Demo</span>
    </div>
  </div>
</div>
"""

css = """
:root {
  --g-bg: #f6f9ff;
  --g-panel: rgba(255,255,255,0.86);
  --g-panel-strong: rgba(255,255,255,0.96);
  --g-line: #dbe7f6;
  --g-ink: #15233d;
  --g-sub: #5f728f;
  --g-accent: #1f6fe5;
  --g-accent-2: #0f9bb9;
}
.gradio-container {
  background:
    radial-gradient(circle at top left, rgba(55,120,246,0.10), transparent 32%),
    radial-gradient(circle at bottom right, rgba(15,155,185,0.10), transparent 28%),
    var(--g-bg);
}
.hero-wrap {
  display: grid;
  grid-template-columns: 280px 1fr;
  gap: 28px;
  padding: 22px 8px 12px 8px;
  align-items: center;
}
.hero-logo-card {
  background: linear-gradient(180deg, rgba(255,255,255,0.96), rgba(247,250,255,0.90));
  border: 1px solid var(--g-line);
  box-shadow: 0 18px 42px rgba(31,58,114,0.08);
  border-radius: 28px;
  padding: 24px;
  display: flex;
  justify-content: center;
  align-items: center;
  min-height: 170px;
}
.hero-logo {
  width: 100%;
  max-width: 220px;
  object-fit: contain;
}
.hero-right h1 {
  font-size: 2.5rem;
  margin: 0;
  color: var(--g-ink);
}
.hero-copy {
  color: var(--g-sub);
  font-size: 1.06rem;
  line-height: 1.6;
  max-width: 780px;
}
.eyebrow {
  color: var(--g-accent);
  font-size: 0.92rem;
  letter-spacing: 0.14em;
  font-weight: 700;
  margin-bottom: 8px;
}
.hero-chips {
  display: flex;
  flex-wrap: wrap;
  gap: 10px;
  margin-top: 14px;
}
.hero-chips span {
  background: rgba(239,246,255,0.96);
  border: 1px solid #cfe0fb;
  color: #28558f;
  border-radius: 999px;
  padding: 8px 12px;
  font-size: 0.9rem;
  font-weight: 600;
}
.card-note {
  color: var(--g-sub);
}
.block-panel {
  background: var(--g-panel);
  border: 1px solid var(--g-line);
  border-radius: 22px;
  padding: 10px;
}
footer {visibility: hidden;}
@media (max-width: 900px) {
  .hero-wrap {grid-template-columns: 1fr;}
}
"""

with gr.Blocks(
    title=f"{APP_TITLE} Studio",
    fill_width=True,
    css=css,
    theme=gr.themes.Soft(
        primary_hue="blue",
        secondary_hue="cyan",
        neutral_hue="slate",
        radius_size="lg",
    ),
) as demo:
    gr.HTML(hero_html)

    with gr.Tabs():
        with gr.Tab("GravityLLM Studio"):
            with gr.Row():
                with gr.Column(scale=11):
                    example_name = gr.Dropdown(
                        choices=list(EXAMPLE_FILES.keys()),
                        value="Club Drop",
                        label="Example payload",
                    )
                    load_btn = gr.Button("Load Example", variant="secondary")
                    payload_box = gr.Code(
                        value=load_example("Club Drop"),
                        language="json",
                        label="Constraint + stem feature payload",
                        lines=26,
                    )

                with gr.Column(scale=6):
                    model_id = gr.Textbox(
                        value=DEFAULT_MODEL_ID,
                        label="Model repo or endpoint",
                        info="Set your Hugging Face model repo id or inference endpoint URL.",
                    )
                    backend = gr.Dropdown(
                        choices=["hybrid", "remote-model", "rules-engine demo"],
                        value=DEFAULT_BACKEND if DEFAULT_BACKEND in {"hybrid", "remote-model", "rules-engine demo"} else "hybrid",
                        label="Backend",
                    )
                    temperature = gr.Slider(0.0, 1.2, value=0.2, step=0.05, label="Temperature")
                    top_p = gr.Slider(0.1, 1.0, value=0.9, step=0.05, label="Top-p")
                    max_new_tokens = gr.Slider(128, 1400, value=900, step=16, label="Max new tokens")
                    use_grammar = gr.Checkbox(
                        value=True,
                        label="Use JSON schema grammar when remote backend supports it",
                    )
                    run_btn = gr.Button("Generate Spatial Scene", variant="primary")
                    status = gr.Textbox(label="Status", interactive=False)

            with gr.Row():
                with gr.Column(scale=9):
                    output_box = gr.Code(language="json", label="Generated Spatial9Scene JSON", lines=26)
                    download = gr.File(label="Download scene JSON")
                with gr.Column(scale=7):
                    summary = gr.Markdown("### Ready\n\nLoad an example or paste your own payload.")
                    plot = gr.Plot(label="Spatial scene preview")

            object_table = gr.Dataframe(
                headers=["id", "class", "az_deg", "el_deg", "dist_m", "width", "gain_db"],
                datatype=["str", "str", "number", "number", "number", "number", "number"],
                row_count=(0, "dynamic"),
                col_count=(7, "fixed"),
                label="Object inspector",
            )

        with gr.Tab("Prompt Builder"):
            gr.Markdown("Build a starter payload, then send it to GravityLLM Studio.")
            with gr.Row():
                target_format = gr.Dropdown(["iamf", "binaural", "5.1.4", "7.1.4"], value="iamf", label="Target format")
                style = gr.Dropdown(["club", "cinematic", "podcast", "live", "intimate"], value="club", label="Style")
                section = gr.Dropdown(["intro", "verse", "break", "drop", "full"], value="drop", label="Section")
            with gr.Row():
                bpm = gr.Slider(0, 200, value=128, step=1, label="BPM")
                energy = gr.Slider(0.0, 1.0, value=0.92, step=0.01, label="Energy")
                max_objects_builder = gr.Slider(1, 32, value=10, step=1, label="Max objects")
            build_btn = gr.Button("Build Payload", variant="primary")
            builder_output = gr.Code(language="json", label="Starter payload", lines=24)
            send_to_studio_btn = gr.Button("Send to Studio", variant="secondary")

        with gr.Tab("Validate Existing Scene"):
            scene_input = gr.Code(language="json", label="Paste a Spatial9Scene JSON", lines=24)
            validate_btn = gr.Button("Validate Scene", variant="primary")
            validate_summary = gr.Markdown()
            validate_plot = gr.Plot()
            validate_table = gr.Dataframe(
                headers=["id", "class", "az_deg", "el_deg", "dist_m", "width", "gain_db"],
                datatype=["str", "str", "number", "number", "number", "number", "number"],
                row_count=(0, "dynamic"),
                col_count=(7, "fixed"),
                label="Validated object inspector",
            )

        with gr.Tab("About"):
            gr.Image(value=str(ASSETS / "gravityllm_space_banner.png"), label="GravityLLM banner", show_download_button=False, show_fullscreen_button=False)
            gr.Markdown(
                """
### What this Space does

- Turns **constraints + stem descriptors** into **Spatial9Scene JSON**
- Can call a remote Hugging Face model repo through `InferenceClient`
- Falls back to a deterministic **rules engine** so the demo stays usable
- Validates outputs against the included JSON schema
- Renders a spatial top-down preview of object positions

### Environment variables

- `GRAVITYLLM_MODEL_ID` — model repo id or endpoint URL
- `HF_TOKEN` — required if the model is gated or private
- `GRAVITYLLM_BACKEND` — optional default: `hybrid`, `remote-model`, or `rules-engine demo`

### Recommended setup

1. Upload your GravityLLM model repo.
2. Train and push weights.
3. Upload this Space repo.
4. Set `GRAVITYLLM_MODEL_ID` in the Space settings.
                """
            )

    load_btn.click(fn=load_example, inputs=example_name, outputs=payload_box)
    build_btn.click(
        fn=build_payload,
        inputs=[target_format, style, section, bpm, energy, max_objects_builder],
        outputs=builder_output,
    )
    send_to_studio_btn.click(fn=lambda x: x, inputs=builder_output, outputs=payload_box)
    run_btn.click(
        fn=generate_scene,
        inputs=[payload_box, model_id, backend, temperature, top_p, max_new_tokens, use_grammar],
        outputs=[output_box, summary, plot, object_table, download, status],
    )
    validate_btn.click(
        fn=validate_only,
        inputs=scene_input,
        outputs=[validate_summary, validate_plot, validate_table],
    )

if __name__ == "__main__":
    demo.launch()