Spaces:
Running
Running
2B-Studio: DAW-Studio, neue Modelle (YuE2, TangoFlux, AudioGen, AudioLDM2), Lizenz-Badges, HF-Login, Einswarm Schwarz/Gold
c600f78 verified Download studio/ui.py from 2Btainment/2b-studio: direct link, hf CLI and curl.
- Browser
- Download file 109 kB
-
https://huggingface.co/spaces/2Btainment/2b-studio/resolve/main/studio/ui.py
- Command line
-
hf download hf://spaces/2Btainment/2b-studio/studio/ui.py
-
curl -L -o ui.py https://huggingface.co/spaces/2Btainment/2b-studio/resolve/main/studio/ui.py
109 kB
from __future__ import annotations
import os
import random
import time
from hashlib import sha256
from html import escape
from pathlib import Path
from typing import Any
import gradio as gr
from studio.adapters import (
AdapterError,
AdapterRouter,
GenerateRequest,
OnusAdapter,
OnusRequest,
SpaceMasterAdapter,
route_id_for_model,
)
from studio.adapters.health import set_visitor_ip_token
from studio.agent import FieldProposal, MasterProposal, ProducerAgent, propose_master
from studio.audio_analysis import AudioAnalysis, analyze_wav
from studio.library import add_track, delete_track, import_wav, library_choices, select_track
from studio.model_status import (
is_unusable,
model_state,
model_status_html_live,
refresh_all_model_labels,
refresh_model_labels_cached,
)
from studio.registry import ModelRegistry, load_registry
from studio.remix import build_remix_plan
from studio.state import ChatTurn, StudioSession, new_session
from studio.templates import CATEGORIES, KEYS, LANGUAGES, MOODS, VOICES, filter_choices, make_custom
from studio.theme import CSS, HEAD, THEME
from studio.daw.api import register as daw_register
from studio.studio_room.handlers import (
studio_chat,
studio_chip_click,
studio_load_track,
studio_open,
studio_upload,
)
from studio.studio_room.html_ui import chips_html as studio_chips_html
from studio.studio_room.html_ui import empty_state_html as studio_empty_html
from studio.studio_room.html_ui import chat_bubble_html as studio_chat_html
from studio.studio_room.agent_3b import default_chips as studio_default_chips
from studio.ui_extras import (
history_html,
model_chips_html,
resolve_template,
reuse_values,
template_detail_html,
template_field_defaults,
template_vars,
render_any,
)
__all__ = ["CSS", "HEAD", "THEME", "build_app"]
# Buy Me a Coffee: schlichter statischer Link (kein Fremd-JS), im Look des Studios
BMC_LINK_HTML = (
'<a href="https://www.buymeacoffee.com/2btainment" target="_blank" rel="noopener noreferrer" '
'class="bmc-link tb-bmc" title="Buy me a coffee – 2Btainment unterstützen" aria-label="Buy me a coffee – 2Btainment unterstützen">'
'<i aria-hidden="true">☕</i><span>Buy me<br>a coffee</span></a>'
)
KEYWORD_LEXICON = {
"◉ Genres": ("genre", ("Dark Riddim", "Afro House", "Neo Soul", "Cinematic")),
"♬ Instrumente": (
"genre",
("Analog Bass", "Rhodes Piano", "808 Drums", "Warm Strings"),
),
"◐ Moods": ("genre", ("Hypnotic", "Euphoric", "Melancholic", "Raw & Intimate")),
"✦ Meta-Formulierungen": (
"genre",
("radio-ready hook", "dynamic arrangement", "human timing", "wide stereo image"),
),
"⊘ Negativ-Prompts": (
"negative",
("muddy low end", "harsh sibilance", "clipping", "generic AI haze"),
),
}
PROMPT_RECIPES = {
"🔥 Club-Druck": (
"club-ready low end, punchy transient design, energetic drop, controlled stereo width",
"weak kick, muddy sub bass, flat dynamics",
),
"✧ Vocal-Klarheit": (
"intimate lead vocal, present center image, clear diction, supportive arrangement",
"harsh sibilance, buried vocals, boxy mids, excessive reverb",
),
"◒ Analog-Wärme": (
"warm analog character, subtle tape saturation, human timing, organic dynamics",
"brittle highs, sterile digital tone, over-compression",
),
"◈ Cinematic-Weite": (
"cinematic depth, evolving arrangement, wide orchestral layers, emotional arc",
"static arrangement, narrow image, synthetic strings, abrupt transitions",
),
}
LOADER_HTML = """
<div class="studio-loader tb-loader" role="status" aria-live="polite">
<div class="tb-loader__eq" aria-hidden="true"><i></i><i></i><i></i><i></i><i></i><i></i><i></i></div>
<div class="tb-loader__txt"><strong>2B formt deinen Take</strong>
<small><span data-loader-phase>Worker wird geweckt</span> · <span data-loader-time>0:00</span></small></div>
<ol class="tb-loader__steps" aria-hidden="true"><li>Wecken</li><li>Rechnen</li><li>Analyse</li></ol>
</div>
"""
SMART_REMIX_HELP = (
"<span class='remix-help-text'>"
"✦ <strong>Smart Remix</strong> analysiert den gewählten Track (Lautstärke, "
"Dynamik, Klangcharakter), behält Melodie & Rhythmus bei, verfeinert das "
"Arrangement und ergänzt automatisch Negativregeln gegen technische Schwächen. "
"Der Seed wird aus dem Referenztrack abgeleitet – gleiche Referenz, gleiches Ergebnis."
"</span>"
)
def _ensure_session(session: StudioSession | None) -> StudioSession:
return session or new_session()
def _source_label(session: StudioSession) -> str:
track = session.current_track()
if track is None:
return "**Quelle:** Noch kein Track gewählt"
return f"**Quelle:** `{Path(track.path).name}` · {track.engine}"
def _analysis_for_track(track: Any) -> AudioAnalysis:
analysis = analyze_wav(track.path)
track.metrics["analysis"] = analysis.as_dict()
return analysis
def _analysis_html(track: Any, analysis: AudioAnalysis | None = None) -> str:
result = analysis or _analysis_for_track(track)
channel_label = "Stereo" if result.channels == 2 else f"{result.channels} Kanal"
return f"""
<section class="track-analysis">
<div class="track-analysis-head"><strong>Referenz-Analyse</strong>
<span>{escape(Path(track.path).name)}</span></div>
<div class="analysis-grid">
<div class="analysis-metric">Dauer<b>{result.duration_s:.1f} s</b></div>
<div class="analysis-metric">Format<b>{result.sample_rate // 1000} kHz · {channel_label}</b></div>
<div class="analysis-metric">RMS<b>{result.rms_dbfs:.1f} dBFS</b></div>
<div class="analysis-metric">Crest<b>{result.crest_db:.1f} dB</b></div>
</div>
<p class="analysis-character">{escape(result.character)}</p>
</section>
"""
def _download_update(track: Any | None) -> Any:
if track is None:
return gr.DownloadButton(value=None, interactive=False)
return gr.DownloadButton(value=track.path, interactive=True)
def _model_status_html(model_id: str, registry: ModelRegistry) -> str:
return model_status_html_live(model_id, registry)
def _tuning_for(model_id: str, registry: ModelRegistry) -> dict[str, tuple[float, float, float]]:
"""Modellspezifische Feineinstellungen aus der Registry (steps/guidance)."""
tuning = registry.get(model_id).capabilities.get("tuning", {})
result: dict[str, tuple[float, float, float]] = {}
if isinstance(tuning, dict):
for key in ("steps", "guidance"):
spec = tuning.get(key)
if isinstance(spec, (list, tuple)) and len(spec) == 3:
default, minimum, maximum = (float(v) for v in spec)
result[key] = (minimum, maximum, default)
return result
def _model_selection_update(
model_id: str,
current_duration: float,
registry: ModelRegistry,
) -> tuple[Any, Any, Any, Any, str]:
model = registry.get(model_id)
preview_limit = max(5, min(180, model.max_duration_s))
duration = max(5, min(preview_limit, round(current_duration)))
tuning = _tuning_for(model_id, registry)
steps_spec = tuning.get("steps")
guidance_spec = tuning.get("guidance")
steps_update = (
gr.Slider(
minimum=steps_spec[0],
maximum=steps_spec[1],
value=steps_spec[2],
step=1,
visible=True,
)
if steps_spec
else gr.Slider(visible=False)
)
guidance_update = (
gr.Slider(
minimum=guidance_spec[0],
maximum=guidance_spec[1],
value=guidance_spec[2],
step=0.1,
visible=True,
)
if guidance_spec
else gr.Slider(visible=False)
)
return (
gr.HTML(value=_model_status_html(model_id, registry)),
gr.Slider(minimum=5, maximum=preview_limit, value=duration, step=1),
steps_update,
guidance_update,
f"Routing: {model.label} → {route_id_for_model(model)}",
)
def _empty_library_update(has_tracks: bool) -> Any:
return gr.Markdown(
value="Keine Tracks in dieser Session." if not has_tracks else "",
visible=not has_tracks,
)
def _chat_messages(session: StudioSession, agent_role: str) -> list[dict[str, str]]:
allowed = {"user", agent_role}
messages: list[dict[str, str]] = []
for turn in session.chat:
if turn.role not in allowed:
continue
role = "user" if turn.role == "user" else "assistant"
messages.append({"role": role, "content": turn.text})
return messages
def _append_keyword(current_value: str | None, keyword: str) -> str:
current_value = current_value or ""
current = current_value.strip()
if keyword.casefold() in current.casefold():
return current_value
if not current:
return keyword
return f"{current.rstrip(' ,')} · {keyword}"
def _apply_prompt_recipe(
genre_value: str, negative_value: str, recipe_name: str
) -> tuple[str, str, str]:
genre_phrase, negative_phrase = PROMPT_RECIPES[recipe_name]
return (
_append_keyword(genre_value, genre_phrase),
_append_keyword(negative_value, negative_phrase),
f"✓ Rezept „{recipe_name}“ wurde auf Sound und Negativ-Prompt verteilt.",
)
def _instrumental_update(enabled: bool, current_lyrics: str) -> tuple[Any, str]:
if enabled:
return (
gr.Textbox(value="[Instrumental]", interactive=False),
"Instrumental aktiv · Lyrics sind deaktiviert.",
)
current_lyrics = current_lyrics or ""
lyrics = "" if current_lyrics.strip().casefold() == "[instrumental]" else current_lyrics
return (
gr.Textbox(value=lyrics, interactive=True),
"Instrumental aus · Lyrics können wieder bearbeitet werden.",
)
def _clear_lyrics(instrumental: bool) -> Any:
return gr.Textbox(
value="[Instrumental]" if instrumental else "",
interactive=not instrumental,
)
def _proposal_html(proposal: FieldProposal) -> str:
lyrics = escape(proposal.lyrics).replace("\n", "<br>")
warnings = "".join(f"<li>{escape(item)}</li>" for item in proposal.warnings)
ask = (
f'<div class="proposal-ask"><b>❓ Rückfrage:</b> {escape(proposal.ask_user)}'
"<br><span>Beantworte die Frage im Chat — ich verfeinere den Vorschlag dann.</span></div>"
if proposal.ask_user
else ""
)
return f"""
<section class="proposal-shell">
<div class="proposal-head">
<strong>Agent-Vorschlag</strong>
<span class="proposal-model">{proposal.duration_s}s · {escape(proposal.model_id or "Aktuelles Modell")}</span>
</div>
{ask}
<div class="proposal-grid">
<div class="proposal-item"><b>Sound & Instrumente</b>{escape(proposal.genre_style_instruments)}</div>
<div class="proposal-item"><b>Lyrics</b>{lyrics}</div>
<div class="proposal-item"><b>Negativ</b>{escape(proposal.negative)}</div>
<div class="proposal-item"><b>Routing</b>{escape(proposal.rationale)}</div>
<div class="proposal-item proposal-wide"><b>Agentenquelle</b>{escape(proposal.source)} · Generierung startet erst nach deiner Bestätigung.</div>
</div>
{f'<ul class="proposal-warnings">{warnings}</ul>' if warnings else ''}
</section>
"""
DEFAULT_SOUND_PROMPT = "warm lo-fi hip hop, 85 BPM, dusty drums, rhodes chords, soft vinyl texture"
def _friendly_error(error: Exception | str, model_label: str = "") -> str:
"""Technische Worker-Fehler in klare deutsche Hinweise mit nächstem Schritt übersetzen."""
raw = str(error)
low = raw.casefold()
who = f"„{model_label}“" if model_label else "Das Modell"
if "zerogpu" in low and ("quota" in low or "limit" in low or "exceeded" in low):
return (f"{who}: ZeroGPU-Kontingent gerade aufgebraucht. Bei Hugging Face eingeloggt "
"gibt es mehr Kontingent – sonst in ein paar Minuten erneut versuchen. Alle ZeroGPU-Modelle "
"teilen sich dieses Kontingent.")
if "paused" in low or "pausiert" in low or "stage=paused" in low or "stopped" in low:
return (f"{who} ist vom Betreiber pausiert und gerade nicht verfügbar. "
"Bitte ein Modell mit Status „Live“ oder „Schläft“ wählen.")
if "nicht geantwortet" in low or "timeout" in low or "timed out" in low:
return (f"{who} hat zu lange gebraucht (Worker überlastet oder Kaltstart). "
"Bitte erneut versuchen – der zweite Versuch ist meist schneller.")
if "token" in low and ("braucht" in low or "unauthorized" in low or "401" in low):
return f"{who} ist privat und braucht ein gültiges HF_TOKEN im Space – bitte anderes Modell wählen."
if "nicht bereit" in low or "unavailable" in low or "unreachable" in low:
return f"{who} ist gerade nicht erreichbar. Bitte kurz warten oder ein anderes Live-Modell wählen."
return f"{who}: {raw}"
def _visitor_token(request: gr.Request | None) -> bool:
"""X-IP-Token des Besuchers für ZeroGPU-Worker setzen; True, wenn eins vorhanden ist."""
try:
headers = request.headers if request is not None else {}
token = headers.get("x-ip-token") if headers else None
except Exception: # noqa: BLE001
token = None
set_visitor_ip_token(token)
return bool(token)
def _is_quota_status(value: Any) -> bool:
return isinstance(value, str) and "ZeroGPU-Kontingent" in value
def _show_loading() -> tuple[str, Any]:
return "Job eingereiht · Worker wird vorbereitet …", gr.HTML(value=LOADER_HTML, visible=True)
def _job_failed() -> tuple[str, Any]:
"""Nach einer Exception: Loader schließen, Status ehrlich setzen (Details stehen im Toast)."""
return (
"⚠ Job fehlgeschlagen – Details in der Meldung oben rechts. Modell-Status prüfen oder anderes Modell wählen.",
gr.HTML(visible=False),
)
def _stop_loading() -> tuple[str, Any]:
return (
"Generierung gestoppt. Bereits erzeugtes Audio bleibt erhalten.",
gr.HTML(visible=False),
)
def _generate(
session: StudioSession | None,
registry: ModelRegistry,
model_id: str,
genre: str,
lyrics: str,
negative: str,
instrumental: bool,
duration_s: int,
seed: int,
steps: int | None = None,
guidance: float | None = None,
) -> tuple[Any, ...]:
current = _ensure_session(session)
registry.get(model_id)
# Felder in eingeklappten Akkordeons kommen als None – nie an Pydantic weiterreichen.
genre, lyrics, negative = (genre or "").strip(), (lyrics or "").strip(), (negative or "").strip()
instrumental = bool(instrumental)
state = model_state(model_id, registry)
if is_unusable(state):
return (
gr.skip(), gr.skip(), gr.skip(), gr.skip(),
"✕ " + _friendly_error("paused" if state == "paused" else "unavailable", registry.get(model_id).label),
_source_label(current), current, gr.HTML(visible=False),
gr.skip(), gr.skip(), gr.skip(), gr.skip(), gr.skip(), gr.skip(), gr.skip(), gr.skip(),
)
used_default_prompt = not genre
if used_default_prompt:
genre = DEFAULT_SOUND_PROMPT
effective_lyrics = "[Instrumental]" if instrumental else lyrics
current.create.model_id = model_id
current.create.genre_style_instruments = genre
current.create.lyrics = effective_lyrics
current.create.negative = negative
current.create.duration_s = int(duration_s)
router = AdapterRouter(registry, Path(current.runtime_dir))
request = GenerateRequest(
model_id=model_id,
genre_style_instruments=genre,
lyrics=effective_lyrics,
negative=negative,
duration_s=int(duration_s),
advanced={
"seed": int(seed),
"instrumental": instrumental,
"studio": current.create.studio,
"project_name": "2B-Studio",
"steps": steps,
"guidance": guidance,
},
)
try:
track = router.generate(request)
except (AdapterError, ValueError) as error:
return (
gr.skip(),
gr.skip(),
gr.skip(),
gr.skip(),
"✕ " + _friendly_error(error, registry.get(model_id).label),
_source_label(current),
current,
gr.HTML(visible=False),
gr.skip(),
gr.skip(),
gr.skip(),
gr.skip(),
gr.skip(),
gr.skip(),
gr.skip(),
gr.skip(),
)
add_track(current, track)
analysis = _analysis_for_track(track)
analysis_markup = _analysis_html(track, analysis)
choices = library_choices(current)
model_label = registry.get(model_id).label
mode = "Instrumental" if instrumental else "Song"
route_id = str(track.metadata["route_id"])
render_mode = "Live-Track" if str(route_id).endswith("-real") else "Worker offline"
used_seed = track.metrics.get("used_seed", track.metrics.get("seed", 0))
seed_note = f"Seed {used_seed}" if used_seed else "Seed zufällig"
status = (
f"✓ {mode} als {render_mode} mit {model_label} · {track.duration_s:.1f} s · "
f"{route_id} · {seed_note}"
)
if used_default_prompt:
status += f" · Kein Sound-Prompt angegeben → Beispiel verwendet: „{DEFAULT_SOUND_PROMPT}“"
status += " · Weiter: anhören, ⬇ Export oder Master →"
download_update = gr.DownloadButton(value=track.path, interactive=True)
return (
track.path,
gr.Radio(choices=choices, value=track.id),
gr.Radio(choices=choices, value=track.id),
download_update,
status,
_source_label(current),
current,
gr.HTML(visible=False),
_download_update(track),
_download_update(track),
gr.HTML(value=analysis_markup, visible=True),
gr.HTML(value=analysis_markup, visible=True),
_empty_library_update(True),
_empty_library_update(True),
track.path,
track.path,
)
def _load_track(
track_id: str, session: StudioSession | None
) -> tuple[str | None, str | None, Any, Any, str, str, StudioSession, Any]:
current = _ensure_session(session)
if not track_id:
return (
None,
None,
_download_update(None),
_download_update(None),
"Kein Track gewählt.",
_source_label(current),
current,
gr.HTML(value="", visible=False),
)
track = select_track(current, track_id)
return (
track.path,
track.path,
_download_update(track),
_download_update(track),
f"▶ {Path(track.path).name} ist im Player geladen.",
_source_label(current),
current,
gr.HTML(value=_analysis_html(track), visible=True),
)
def _delete_selected(track_id: str, session: StudioSession | None) -> tuple[Any, ...]:
current = _ensure_session(session)
if not track_id:
raise gr.Error("Bitte zuerst einen Track in der Bibliothek auswählen.")
deleted = delete_track(current, track_id)
fallback = current.current_track()
choices = library_choices(current)
fallback_id = fallback.id if fallback else None
analysis_markup = _analysis_html(fallback) if fallback else ""
analysis_update = gr.HTML(value=analysis_markup, visible=fallback is not None)
return (
fallback.path if fallback else None,
gr.Radio(choices=choices, value=fallback_id),
gr.Radio(choices=choices, value=fallback_id),
_download_update(fallback),
_download_update(fallback),
_download_update(fallback),
f"🗑 {Path(deleted.path).name} wurde aus der Session gelöscht.",
_source_label(current),
current,
analysis_update,
gr.HTML(value=analysis_markup, visible=fallback is not None),
_empty_library_update(bool(current.tracks)),
_empty_library_update(bool(current.tracks)),
fallback.path if fallback else None,
fallback.path if fallback else None,
)
def _show_remix_loading() -> tuple[str, Any]:
return (
"Smart Remix analysiert den Referenztrack und perfektioniert die Prompts …",
gr.HTML(value=LOADER_HTML, visible=True),
)
def _smart_remix(
track_id: str,
session: StudioSession | None,
registry: ModelRegistry,
) -> tuple[Any, ...]:
current = _ensure_session(session)
source = current.track_by_id(track_id) if track_id else current.current_track()
if source is None:
raise gr.Error("Für Smart Remix muss zuerst ein Referenztrack ausgewählt werden.")
source_analysis = _analysis_for_track(source)
plan = build_remix_plan(source, source_analysis)
try:
registry.get(plan.model_id)
model_id = plan.model_id
except KeyError:
model_id = "minimax-music3"
reference_hash = sha256(Path(source.path).read_bytes()).hexdigest()
seed = int(reference_hash[:8], 16)
router = AdapterRouter(registry, Path(current.runtime_dir))
try:
remix = router.generate(
GenerateRequest(
model_id=model_id,
genre_style_instruments=plan.genre_style_instruments,
lyrics=plan.lyrics,
negative=plan.negative,
duration_s=plan.duration_s,
advanced={
"seed": seed,
"workflow": "smart_remix",
"reference_track_id": source.id,
"reference_sha256": reference_hash,
"reference_analysis": source_analysis.as_dict(),
},
)
)
except AdapterError as error:
instrumental = current.create.lyrics.strip().casefold() == "[instrumental]"
return (
gr.skip(),
gr.skip(),
gr.skip(),
gr.skip(),
gr.skip(),
gr.skip(),
"✕ Smart Remix: " + _friendly_error(error, registry.get(model_id).label),
_source_label(current),
current,
gr.HTML(visible=False),
gr.skip(),
gr.skip(),
gr.skip(),
gr.skip(),
current.create.model_id,
current.create.genre_style_instruments,
gr.Textbox(value=current.create.lyrics, interactive=not instrumental),
current.create.negative,
instrumental,
gr.skip(),
gr.skip(),
)
remix.parent_id = source.id
remix.metadata.update(
{
"workflow": "smart_remix",
"reference_track_id": source.id,
"reference_analysis": source_analysis.as_dict(),
"remix_rationale": plan.rationale,
}
)
remix.metrics["reference_analysis"] = source_analysis.as_dict()
add_track(current, remix)
current.create.model_id = model_id
current.create.genre_style_instruments = plan.genre_style_instruments
current.create.lyrics = plan.lyrics
current.create.negative = plan.negative
current.create.duration_s = plan.duration_s
remix_analysis = _analysis_for_track(remix)
analysis_markup = _analysis_html(remix, remix_analysis)
choices = library_choices(current)
instrumental = plan.lyrics.strip().casefold() == "[instrumental]"
status_text = (
f"✓ Smart Remix fertig · Referenz {Path(source.path).name} · {plan.summary or plan.rationale}"
)
return (
remix.path,
gr.Radio(choices=choices, value=remix.id),
gr.Radio(choices=choices, value=remix.id),
_download_update(remix),
_download_update(remix),
_download_update(remix),
status_text,
_source_label(current),
current,
gr.HTML(visible=False),
gr.HTML(value=analysis_markup, visible=True),
gr.HTML(value=analysis_markup, visible=True),
_empty_library_update(True),
_empty_library_update(True),
model_id,
plan.genre_style_instruments,
gr.Textbox(value=plan.lyrics, interactive=not instrumental),
plan.negative,
instrumental,
remix.path,
remix.path,
)
def _master_selected(
track_id: str,
session: StudioSession | None,
) -> tuple[Any, ...]:
current = _ensure_session(session)
track = current.track_by_id(track_id) if track_id else current.current_track()
if track is None:
raise gr.Error("Bitte zuerst einen Track für das Mastering auswählen.")
select_track(current, track.id)
return (
gr.Column(visible=False),
gr.Column(visible=True),
gr.Column(visible=False),
"Track als Master-Quelle gesetzt · lokales ONUS ist startbereit.",
_source_label(current),
current,
track.path,
gr.Radio(choices=library_choices(current), value=track.id),
_download_update(track),
_download_update(track),
gr.HTML(value=_analysis_html(track), visible=True),
track.path,
track.path,
)
def _import_library_wav(
source_path: str | None,
session: StudioSession | None,
) -> tuple[Any, ...]:
current = _ensure_session(session)
if not source_path:
raise gr.Error("Bitte zuerst eine WAV-Datei auswählen.")
try:
track = import_wav(current, source_path)
except ValueError as error:
raise gr.Error(str(error)) from error
choices = library_choices(current)
analysis_markup = _analysis_html(track)
return (
track.path,
track.path,
track.path,
gr.Radio(choices=choices, value=track.id),
gr.Radio(choices=choices, value=track.id),
_download_update(track),
_download_update(track),
_download_update(track),
f"✓ {Path(source_path).name} wurde lokal importiert und analysiert.",
_source_label(current),
current,
gr.HTML(value=analysis_markup, visible=True),
gr.HTML(value=analysis_markup, visible=True),
_empty_library_update(True),
_empty_library_update(True),
)
def _producer_reply(
message: str,
session: StudioSession | None,
agent: ProducerAgent,
) -> tuple[str, list[dict[str, str]], dict[str, Any] | None, Any, Any, Any, StudioSession]:
current = _ensure_session(session)
clean_message = (message or "").strip()
if not clean_message:
return (
"",
_chat_messages(current, "prompt_agent"),
None,
gr.HTML(visible=False),
gr.Button(interactive=False),
gr.Button(visible=False),
current,
)
action = agent.propose(clean_message, current.create.model_dump())
proposal = action.payload
warning_note = f" {len(proposal.warnings)} Hinweis(e) stehen in der Vorschau." if proposal.warnings else ""
if proposal.ask_user:
response = (
f"Bevor ich den Vorschlag finalisiere: **{proposal.ask_user}** "
"Ich habe schon einen Entwurf vorbereitet — antworte kurz, dann verfeinere ich ihn."
)
else:
response = (
f"{action.preview}. Genre, Lyrics, Negativ-Prompt, Dauer und Modell sind vorbereitet."
f"{warning_note} Prüfe alles und klicke **Übernehmen**; ich starte keinen Job selbst."
)
current.create.idea = clean_message
current.chat.extend(
[
ChatTurn(role="user", text=clean_message),
ChatTurn(
role="prompt_agent",
text=response,
actions=[{"type": "fill_create_fields", "payload": proposal.model_dump()}],
),
]
)
return (
"",
_chat_messages(current, "prompt_agent"),
proposal.model_dump(),
gr.HTML(value=_proposal_html(proposal), visible=True),
gr.Button(interactive=True),
gr.Button(visible=True),
current,
)
def _apply_proposal(
raw_proposal: dict[str, Any] | None,
session: StudioSession | None,
registry: ModelRegistry,
) -> tuple[str, str, Any, str, bool, int, Any, Any, str, StudioSession, Any]:
current = _ensure_session(session)
if not raw_proposal:
instrumental = current.create.lyrics.strip().casefold() == "[instrumental]"
return (
current.create.model_id,
current.create.genre_style_instruments,
gr.Textbox(value=current.create.lyrics, interactive=not instrumental),
current.create.negative,
instrumental,
current.create.duration_s,
gr.Slider(visible=False),
gr.Slider(visible=False),
"Der Agent hat noch keinen Vorschlag erstellt.",
current,
gr.Button(interactive=False),
)
proposal = FieldProposal.model_validate(raw_proposal)
model_id = proposal.model_id or current.create.model_id
instrumental = proposal.lyrics.strip().casefold() == "[instrumental]"
current.create.model_id = model_id
current.create.genre_style_instruments = proposal.genre_style_instruments
current.create.lyrics = proposal.lyrics
current.create.negative = proposal.negative
current.create.duration_s = proposal.duration_s
current.create.idea = proposal.idea
current.create.studio = proposal.studio
current.create.renderings = proposal.renderings
tuning = _tuning_for(model_id, registry)
steps_spec = tuning.get("steps")
guidance_spec = tuning.get("guidance")
steps_update = (
gr.Slider(
minimum=steps_spec[0],
maximum=steps_spec[1],
value=steps_spec[2],
step=1,
visible=True,
)
if steps_spec
else gr.Slider(visible=False)
)
guidance_update = (
gr.Slider(
minimum=guidance_spec[0],
maximum=guidance_spec[1],
value=guidance_spec[2],
step=0.1,
visible=True,
)
if guidance_spec
else gr.Slider(visible=False)
)
return (
model_id,
proposal.genre_style_instruments,
gr.Textbox(value=proposal.lyrics, interactive=not instrumental),
proposal.negative,
instrumental,
proposal.duration_s,
steps_update,
guidance_update,
"✓ Agent-Vorschlag übernommen. Alle Felder bleiben editierbar.",
current,
gr.Button(interactive=False),
)
def _discard_proposal() -> tuple[None, Any, Any, Any, str]:
return (
None,
gr.HTML(value="", visible=False),
gr.Button(interactive=False),
gr.Button(visible=False),
"Agent-Vorschlag verworfen.",
)
def _master_reply(
message: str, session: StudioSession | None
) -> tuple[str, list[dict[str, str]], dict[str, Any] | None, Any, Any, StudioSession]:
current = _ensure_session(session)
clean_message = (message or "").strip()
if not clean_message:
return (
"",
_chat_messages(current, "master_agent"),
None,
gr.HTML(visible=False),
gr.Button(interactive=False),
current,
)
action = propose_master(clean_message)
proposal = action.payload
response = (
f"{action.preview}. {proposal.rationale} Prüfe den Vorschlag und übernimm ihn; "
"ONUS startet nicht automatisch."
)
current.chat.extend(
[ChatTurn(role="user", text=clean_message), ChatTurn(role="master_agent", text=response)]
)
approval = (
"<p class='master-warning'>Strict benötigt anschließend eine ausdrückliche Bestätigung.</p>"
if proposal.requires_strict_approval
else ""
)
preview = f"""
<section class="proposal-shell">
<div class="proposal-head"><strong>Master-Vorschlag</strong>
<span class="proposal-model">{escape(proposal.preview)}</span></div>
<div class="proposal-grid">
<div class="proposal-item"><b>Profil</b>{escape(proposal.profile)}</div>
<div class="proposal-item"><b>Purity</b>{escape(proposal.purity_mode)}</div>
<div class="proposal-item"><b>Air / Margin</b>{proposal.keep_air_db:g} dB · {proposal.margin:g}</div>
<div class="proposal-item"><b>Begründung</b>{escape(proposal.rationale)}</div>
</div>{approval}
</section>
"""
return (
"",
_chat_messages(current, "master_agent"),
proposal.model_dump(),
gr.HTML(value=preview, visible=True),
gr.Button(interactive=True),
current,
)
def _apply_master_proposal(
raw_proposal: dict[str, Any] | None,
session: StudioSession | None,
) -> tuple[str, str, float, float, int, bool, bool, str, StudioSession]:
current = _ensure_session(session)
if not raw_proposal:
return (
current.master.profile,
current.master.purity_mode,
current.master.keep_air_db,
current.master.margin,
current.master.intensity,
False,
False,
"Der Mastering-Agent hat noch keinen Vorschlag erstellt.",
current,
)
proposal = MasterProposal.model_validate(raw_proposal)
current.master.profile = proposal.profile
current.master.purity_mode = proposal.purity_mode
current.master.keep_air_db = proposal.keep_air_db
current.master.margin = proposal.margin
current.master.intensity = proposal.intensity
current.master.human_approval = "pending" if proposal.requires_strict_approval else "n/a"
return (
proposal.profile,
proposal.purity_mode,
proposal.keep_air_db,
proposal.margin,
proposal.intensity,
proposal.analyze_only,
False,
"✓ Master-Vorschlag übernommen. ONUS wartet auf deinen Start.",
current,
)
def _master_report_html(result: Any) -> str:
loudness = result.analysis.get("loudness", {})
master = result.master or {}
after = master.get("after", {}) if isinstance(master, dict) else {}
purity = result.purity or {}
return f"""
<section class="track-analysis master-report">
<div class="track-analysis-head"><strong>Mastering Report</strong>
<span>Human approval: pending</span></div>
<div class="analysis-grid">
<div class="analysis-metric">Input LUFS<b>{loudness.get('integrated_lufs', 'n/a')}</b></div>
<div class="analysis-metric">Master LUFS<b>{after.get('integrated_lufs', 'n/a')}</b></div>
<div class="analysis-metric">True Peak<b>{after.get('true_peak_dbtp', 'n/a')}</b></div>
<div class="analysis-metric">QA<b>{master.get('perceptual_verdict', 'Analyze only')}</b></div>
<div class="analysis-metric">Residual Ratio<b>{purity.get('residual_energy_ratio', 'off')}</b></div>
<div class="analysis-metric">Air erhalten<b>{purity.get('air_preserved_db', 'off')} dB</b></div>
</div>
<p class="analysis-character">Original unverändert: {result.original_unchanged}</p>
</section>
"""
def _show_master_loading() -> tuple[str, Any]:
return "Mastering analysiert den Track · Original bleibt unverändert …", gr.HTML(
value=LOADER_HTML, visible=True
)
def _run_onus(
track_id: str,
session: StudioSession | None,
profile: str,
purity_mode: str,
keep_air_db: float,
margin: float,
intensity: int,
analyze_only: bool,
strict_approved: bool,
) -> tuple[Any, ...]:
current = _ensure_session(session)
source = current.track_by_id(track_id) if track_id else current.current_track()
if source is None:
raise gr.Error("Bitte zuerst einen Master-Quelltrack auswählen.")
in_space = os.environ.get("STUDIO_SPACE_MODE") == "1"
try:
if in_space:
result = SpaceMasterAdapter(Path(current.runtime_dir)).process(
source,
OnusRequest(
profile=profile,
purity_mode=purity_mode,
keep_air_db=keep_air_db,
margin=margin,
intensity=intensity,
analyze_only=analyze_only,
strict_approved=strict_approved,
),
)
else:
result = OnusAdapter(Path(current.runtime_dir)).process(
source,
OnusRequest(
profile=profile,
purity_mode=purity_mode,
keep_air_db=keep_air_db,
margin=margin,
intensity=intensity,
analyze_only=analyze_only,
strict_approved=strict_approved,
),
)
except (AdapterError, ValueError) as error:
return (
gr.skip(),
source.path,
gr.skip(),
gr.skip(),
gr.skip(),
gr.skip(),
gr.skip(),
gr.skip(),
f"✕ Mastering fehlgeschlagen: {error}",
_source_label(current),
current,
gr.HTML(visible=False),
gr.HTML(value="", visible=False),
gr.skip(),
gr.skip(),
gr.skip(),
gr.skip(),
gr.HTML(visible=False),
gr.skip(),
gr.skip(),
)
report_html = gr.HTML(value=_master_report_html(result), visible=True)
if result.tracks:
current.tracks.extend(result.tracks)
output = result.tracks[-1]
current.master.source_track_id = output.id
current.master.last_report_id = Path(result.session_dir).name
choices = library_choices(current)
analysis_markup = _analysis_html(output)
engine = "Space-Mastering" if in_space else "ONUS-Mastering"
status = f"✓ {engine} fertig · {profile} · Original unverändert"
return (
output.path,
source.path,
output.path,
gr.Radio(choices=choices, value=output.id),
gr.Radio(choices=choices, value=output.id),
_download_update(output),
_download_update(output),
_download_update(output),
status,
_source_label(current),
current,
report_html,
gr.HTML(value=analysis_markup, visible=True),
gr.HTML(value=analysis_markup, visible=True),
output.path,
output.path,
gr.HTML(visible=False),
_empty_library_update(True),
_empty_library_update(True),
gr.HTML(visible=False),
)
status = "✓ Analyse abgeschlossen · keine Audiodatei verändert"
return (
source.path,
source.path,
None,
gr.skip(),
gr.skip(),
_download_update(source),
_download_update(source),
_download_update(source),
status,
_source_label(current),
current,
report_html,
gr.skip(),
gr.skip(),
source.path,
source.path,
gr.HTML(visible=False),
_empty_library_update(bool(current.tracks)),
_empty_library_update(bool(current.tracks)),
gr.HTML(visible=False),
)
def _open_master(
session: StudioSession | None,
) -> tuple[Any, Any, Any, str, str, StudioSession, str | None]:
current = _ensure_session(session)
track = current.current_track()
if track is None:
return (
gr.Column(visible=True),
gr.Column(visible=False),
gr.Column(visible=False),
"Erzeuge oder wähle zuerst einen Track.",
_source_label(current),
current,
None,
)
in_space = os.environ.get("STUDIO_SPACE_MODE") == "1"
current.master.source_track_id = track.id
engine = "Space-Mastering" if in_space else "ONUS-Mastering"
return (
gr.Column(visible=False),
gr.Column(visible=True),
gr.Column(visible=False),
f"Track an die {engine}-Ebene übergeben.",
_source_label(current),
current,
track.path,
)
def _open_create(session: StudioSession | None) -> tuple[Any, Any, Any, str, StudioSession]:
current = _ensure_session(session)
return (
gr.Column(visible=True),
gr.Column(visible=False),
gr.Column(visible=False),
"Create-Ebene aktiv.",
current,
)
# ===========================================================================
# Räume, Verlauf, Einstellungen wiederverwenden
# ===========================================================================
ROOMS_HTML = """
<a class="tb-skip" href="#tb-work">Zum Arbeitsbereich springen</a>
<div class="tb-brand" aria-label="2B Studio">
<svg viewBox="0 0 64 64" aria-hidden="true"><path d="M19 6h26l13 13v26L45 58H19L6 45V19z"/><g><rect x="16" y="27" width="4" height="10" rx="2"/><rect x="23" y="20" width="4" height="24" rx="2"/><rect x="30" y="14" width="4" height="36" rx="2"/><rect x="37" y="22" width="4" height="20" rx="2"/><rect x="44" y="28" width="4" height="8" rx="2"/></g></svg>
<span>2B<b>Studio</b></span>
</div>
<div class="tb-rail__label" aria-hidden="true">Räume</div>
"""
RAIL_FOOT_HTML = f"""
<div class="tb-rail__foot">
{BMC_LINK_HTML}
<span class="tb-by">by <b>EINswarm</b></span>
</div>
"""
TOPBAR_HTML = """
<header class="tb-top">
<div class="tb-top__title">
<span class="tb-tape" data-room-kicker>KI-Musikstudio · Raum Generate</span>
<h1 data-room-title>Generate</h1>
</div>
<nav class="tb-flow" aria-label="Arbeitsschritte">
<a href="#tb-step-model"><b>01</b>Modell</a>
<a href="#tb-step-params"><b>02</b>Einstellungen</a>
<a href="#tb-step-prompt"><b>03</b>Prompt</a>
<a href="#tb-generate-bar"><b>04</b>Generieren</a>
<a href="#tb-player"><b>05</b>Ergebnis</a>
<a href="#tb-side"><b>⟲</b>Verlauf</a>
</nav>
</header>
"""
DAW_HTML = '<div id="daw-root" class="daw-root" aria-label="2B Studio DAW"><p class="daw-hint" style="padding:24px">Studio wird geladen …</p></div>'
BG_LAYERS_HTML = """
<div class="tb-bg" aria-hidden="true"></div><div class="tb-grain" aria-hidden="true"></div>
<canvas class="tb-fx" aria-hidden="true"></canvas>
"""
def _section_head(number: str, title: str, hint: str, anchor: str) -> str:
return (
f'<div class="tb-sec__head" id="{anchor}"><span class="tb-sec__num">{number}</span>'
f"<h2>{escape(title)}</h2><p>{escape(hint)}</p></div>"
)
def _open_master_room(
session: StudioSession | None,
) -> tuple[Any, Any, Any, str, str, StudioSession, str | None]:
"""Raum Master öffnen – auch ohne Track (Upload ist dort möglich)."""
current = _ensure_session(session)
track = current.current_track()
in_space = os.environ.get("STUDIO_SPACE_MODE") == "1"
engine = "Space-Mastering" if in_space else "ONUS-Mastering"
if track is not None:
current.master.source_track_id = track.id
note = f"Raum Master · {engine} · Quelle: {Path(track.path).name}"
else:
note = f"Raum Master · {engine} · Noch kein Track – generiere einen Take oder importiere eine WAV."
return (
gr.Column(visible=False),
gr.Column(visible=True),
gr.Column(visible=False),
note,
_source_label(current),
current,
track.path if track else None,
)
def _histories(
session: StudioSession | None, selected_create: str | None, selected_master: str | None, registry: ModelRegistry
) -> tuple[str, str]:
return (
history_html(session, selected_create, registry, "create"),
history_html(session, selected_master, registry, "master"),
)
def _reuse_settings(track_id: str | None, session: StudioSession | None, registry: ModelRegistry) -> tuple[Any, ...]:
current = _ensure_session(session)
values = reuse_values(track_id, current)
if values is None:
raise gr.Error("Dieser Take enthält keine wiederverwendbaren Einstellungen (z. B. Import oder Master).")
model_id = values["model_id"] if values["model_id"] in {m.id for m in registry.models} else current.create.model_id
model = registry.get(model_id)
duration = max(5, min(int(model.max_duration_s), 180, int(values["duration_s"])))
instrumental = bool(values["instrumental"])
studio = values["studio"]
current.create.model_id = model_id
current.create.genre_style_instruments = values["genre"]
current.create.lyrics = values["lyrics"]
current.create.negative = values["negative"]
current.create.duration_s = duration
current.create.studio = {k: v for k, v in studio.items() if v}
return (
model_id,
values["genre"],
gr.Textbox(value=values["lyrics"], interactive=not instrumental),
values["negative"],
instrumental,
duration,
values["seed"],
studio.get("global_metadata", ""),
studio.get("vocal_details", ""),
studio.get("arrangement", ""),
f"↺ Einstellungen übernommen · {model.label} · Seed {values['seed'] or 'zufällig'} – gleicher Seed = reproduzierbar.",
current,
)
def _set_studio_fields(global_meta: str, vocal: str, arrangement: str, session: StudioSession | None) -> StudioSession:
current = _ensure_session(session)
fields = {"global_metadata": global_meta, "vocal_details": vocal, "arrangement": arrangement}
current.create.studio = {k: v.strip() for k, v in fields.items() if v and v.strip()}
return current
def _studio_visibility(model_id: str) -> Any:
return gr.Accordion(visible=model_id.startswith("minimax"))
def _gen_summary(model_id: str, seconds: float, instrumental: bool, seed_value: float, registry: ModelRegistry) -> str:
try:
entry = registry.get(model_id)
label = entry.label
vocals_ok = bool(entry.supports_vocals)
except KeyError:
label, vocals_ok = model_id, True
seed_txt = f"Seed {int(seed_value or 0)}" if seed_value else "Seed zufällig"
mode = "Instrumental" if (instrumental or not vocals_ok) else "Song mit Gesang"
try:
state = model_state(model_id, registry)
except Exception: # noqa: BLE001
state = "live"
if is_unusable(state):
label = f"{label} · ⚠ pausiert – anderes Modell wählen" if state == "paused" else f"{label} · ⚠ nicht verfügbar"
return (
f'<div class="tb-gensum"><span><b>{escape(label)}</b></span><span>{int(seconds or 0)} s</span>'
f"<span>{mode}</span><span>{seed_txt}</span>"
'<span class="tb-gensum__kbd"><kbd>Strg</kbd>+<kbd>Enter</kbd></span></div>'
)
# ===========================================================================
# Vorlagen-Bibliothek
# ===========================================================================
_TPL_VAR_COUNT = 10
def _tpl_choices_update(category: str, query: str, favs: list[str] | None, custom: list[dict] | None,
value: str | None = None) -> Any:
choices = filter_choices(category or "all", query or "", favs, custom)
ids = [c[1] for c in choices]
keep = value if value in ids else None
return gr.Radio(choices=choices, value=keep)
def _tpl_count_html(category: str, query: str, favs: list[str] | None, custom: list[dict] | None) -> str:
n = len(filter_choices(category or "all", query or "", favs, custom))
empty = {
"fav": "Noch keine Favoriten – Vorlage wählen und ★ antippen.",
"custom": "Noch keine eigenen Vorlagen – Felder ausfüllen und unten speichern.",
}.get(category or "all", "Keine Treffer – Suchbegriff kürzen.")
if n == 0:
return f'<p class="tb-tpl-count is-empty">{escape(empty)}</p>'
return f'<p class="tb-tpl-count">{n} Vorlage{"n" if n != 1 else ""}</p>'
def _fav_label(template_id: str | None, favs: list[str] | None) -> Any:
on = bool(template_id) and template_id in set(favs or [])
return gr.Button(value="★ Favorit" if on else "☆ Merken", interactive=bool(template_id))
def build_app() -> gr.Blocks:
registry = load_registry()
producer_agent = ProducerAgent(registry)
model_choices = refresh_all_model_labels(registry)
lexicon_bindings: list[tuple[gr.Button, str, str]] = []
recipe_bindings: list[tuple[gr.Button, str]] = []
wave_opts = gr.WaveformOptions(
waveform_color="#7a5f1c", waveform_progress_color="#d4af37", trim_region_color="rgba(212, 175, 55,.25)"
)
with gr.Blocks(title="2B Studio · KI-Musikstudio · by EINswarm", fill_width=True) as app:
gr.HTML(BG_LAYERS_HTML, elem_classes="tb-layers", padding=False)
session_state = gr.State(value=new_session)
proposal_state = gr.State(value=None)
master_proposal_state = gr.State(value=None)
tpl_favs = gr.BrowserState([], storage_key="2bstudio-template-favs")
tpl_custom = gr.BrowserState([], storage_key="2bstudio-template-custom")
with gr.Row(elem_id="tb-shell", equal_height=False):
# ---------------------------------------------------------------- Raum-Leiste
with gr.Column(elem_id="tb-rail", scale=0, min_width=96):
gr.HTML(ROOMS_HTML, padding=False)
room_generate = gr.Button("Generate", elem_id="room-generate", elem_classes=["tb-room", "is-active"])
room_master = gr.Button("Master", elem_id="room-master", elem_classes="tb-room")
room_studio = gr.Button("Studio", elem_id="room-studio", elem_classes="tb-room")
gr.LoginButton("HF Login", size="sm", elem_id="tb-login")
gr.HTML(RAIL_FOOT_HTML, padding=False)
# ---------------------------------------------------------------- Bühne
with gr.Column(elem_id="tb-stage", scale=1):
gr.HTML(TOPBAR_HTML, padding=False)
with gr.Row(elem_id="tb-chips-row"):
model_chips = gr.HTML(model_chips_html(registry, "minimax-music3"), padding=False, elem_id="tb-mchips")
refresh_chips = gr.Button("↻ Status", size="sm", elem_id="tb-refresh-chips", scale=0, min_width=90)
# ============================================================ Raum GENERATE
# Aufbau wie vor v2 (6181121a): Modell oben, darunter Einstellungen | Prompt | Bibliothek.
with gr.Column(visible=True, elem_id="tb-create-panel") as create_panel:
# ---------------------------------------------- 01 Modell
with gr.Column(elem_classes=["tb-sec", "tb-sec--model"], elem_id="tb-work"):
gr.HTML(_section_head("01", "Modell", "Wähle die Engine. Der Status zeigt, ob sie sofort bereit ist.", "tb-step-model"), padding=False)
with gr.Row(elem_classes="model-controls", equal_height=False):
model = gr.Dropdown(
choices=model_choices,
value="minimax-music3",
label="Modellauswahl",
info="Live = sofort bereit · Schläft = wird beim Start geweckt · Pausiert = nicht verfügbar",
scale=3,
elem_id="tb-model",
)
instrumental = gr.Checkbox(
label="Instrumental",
value=False,
info="Ohne Gesang und Lyrics",
elem_classes="instrumental-switch",
scale=1,
)
model_status = gr.HTML(value=_model_status_html("minimax-music3", registry), elem_classes="model-wrap")
with gr.Row(elem_classes="tb-gen-grid", equal_height=False):
# ------------------------------------------ 02 Einstellungen (links, wie vorher)
with gr.Column(elem_classes=["tb-sec", "tb-sec--params"], elem_id="tb-params-col", scale=1, min_width=260):
gr.HTML(_section_head("02", "Einstellungen", "Länge und Seed. Profi-Regler stecken unter „Erweitert“.", "tb-step-params"), padding=False)
duration = gr.Slider(5, 180, value=12, step=1, label="Dauer · Sekunden", info="Wird aufs Maximum des Modells begrenzt")
with gr.Row(elem_classes="seed-row"):
seed = gr.Number(
value=0, precision=0, label="Seed",
info="0 = jedes Mal neu · gleiche Zahl = gleiches Ergebnis", scale=3,
)
random_seed = gr.Button("🎲", size="sm", min_width=44, elem_classes="seed-dice", scale=0)
with gr.Accordion("Erweitert", open=False, elem_classes="tb-acc", elem_id="tb-advanced"):
steps = gr.Slider(1, 64, value=24, step=1, label="Schritte (Steps)", info="Mehr = detaillierter, aber langsamer", visible=True)
guidance = gr.Slider(1.0, 4.0, value=1.7, step=0.1, label="Guidance", info="Wie streng das Modell dem Prompt folgt", visible=True)
with gr.Accordion("MiniMax · Structured Caption", open=False, elem_classes="tb-acc", visible=True) as studio_acc:
gr.HTML('<p class="tb-hint">Wird von Vorlagen automatisch gefüllt. Leer = sinnvolle Standardwerte aus dem Sound-Feld.</p>', padding=False)
studio_global = gr.Textbox(label="Global Metadata", lines=2)
studio_vocal = gr.Textbox(label="Vocal Details", lines=2)
studio_arrangement = gr.Textbox(label="Arrangement", lines=2)
# ------------------------------------------ 03 Prompt (Mitte, wie vorher)
with gr.Column(elem_classes=["tb-sec", "tb-sec--prompt"], elem_id="tb-prompt-col", scale=2, min_width=320):
gr.HTML(_section_head("03", "Prompt", "Beschreibe den Sound. Lyrics und Ausschlüsse sind optional.", "tb-step-prompt"), padding=False)
with gr.Column(elem_classes="prompt-stack"):
with gr.Accordion("Sound · Genre, Stil & Instrumente", open=True, elem_classes=["prompt-card", "tb-field", "tb-field--sound"]):
genre = gr.Textbox(
label="Sound · Genre, Stil & Instrumente", show_label=False,
placeholder="z. B. dark riddim, 140 BPM, analog bass …",
lines=5, elem_classes="prompt-box", elem_id="tb-genre",
)
with gr.Row(elem_classes="tb-field__foot"):
gr.HTML('<span class="tb-hint">Klangbild, Tempo, Arrangement</span>', padding=False)
clear_genre = gr.Button("Leeren", size="sm", min_width=70, elem_classes="tb-clear")
with gr.Accordion("Lyrics · Text & Songstruktur (optional)", open=False, elem_classes=["prompt-card", "tb-field"], elem_id="tb-lyrics-acc") as lyrics_acc:
lyrics = gr.Textbox(
label="Lyrics · Text & Songstruktur", show_label=False,
placeholder="[Verse]\n…\n[Chorus]\n…",
lines=6, elem_classes="prompt-box", elem_id="tb-lyrics",
)
with gr.Row(elem_classes="tb-field__foot"):
gr.HTML('<span class="tb-hint">[Verse] [Chorus] [Bridge] – leer = Modell improvisiert</span>', padding=False)
clear_lyrics = gr.Button("Leeren", size="sm", min_width=70, elem_classes="tb-clear")
with gr.Accordion("Vermeiden · Negativ-Prompt (optional)", open=False, elem_classes=["prompt-card", "tb-field", "tb-field--avoid"]) as negative_acc:
negative = gr.Textbox(
label="Vermeiden · Negativ-Prompt", show_label=False,
placeholder="muddy low end, clipping …",
lines=3, elem_classes="prompt-box", elem_id="tb-negative",
)
with gr.Row(elem_classes="tb-field__foot"):
gr.HTML('<span class="tb-hint">Artefakte & Ausschlüsse</span>', padding=False)
clear_negative = gr.Button("Leeren", size="sm", min_width=70, elem_classes="tb-clear")
gr.HTML('<div class="tb-help-label">Hilfen für den Prompt</div>', padding=False)
with gr.Accordion("Vorlagen-Bibliothek · 30 Vorlagen", open=False, elem_id="tb-templates", elem_classes="tb-acc"):
tpl_category = gr.Radio(
choices=CATEGORIES, value="all", show_label=False,
elem_classes="tb-seg", container=False,
)
with gr.Row(elem_classes="tb-tpl-bar"):
tpl_search = gr.Textbox(
placeholder="Suchen: Trailer, Lo-Fi, 808, Podcast …", show_label=False,
container=False, scale=4, elem_id="tb-tpl-search",
)
tpl_count = gr.HTML(_tpl_count_html("all", "", [], []), padding=False, elem_classes="tb-tpl-countbox")
tpl_pick = gr.Radio(
choices=filter_choices("all", "", [], []), value=None, show_label=False,
label="Vorlage", elem_classes="tb-tpl-grid", container=False,
)
with gr.Row(elem_classes="tb-tpl-body", equal_height=False):
with gr.Column(scale=5, min_width=280, elem_classes="tb-tpl-vars"):
with gr.Row():
tpl_bpm = gr.Slider(0, 200, value=120, step=1, label="BPM", info="0 = ohne Tempo")
tpl_duration = gr.Slider(3, 300, value=30, step=1, label="Dauer (s)", info="wird aufs Modell-Maximum begrenzt")
with gr.Row():
tpl_key = gr.Dropdown(KEYS, value="beliebig", label="Tonart")
tpl_mood = gr.Dropdown(list(MOODS), value="Neutral", label="Stimmung", allow_custom_value=True)
tpl_instruments = gr.Textbox(label="Instrumente", lines=1, placeholder="z. B. 808 bass, rhodes, taiko")
tpl_vocals = gr.Checkbox(label="Mit Gesang (Lyrics-Gerüst in Sprache & Stimmung)", value=False, elem_classes="tb-check-tile")
with gr.Row():
tpl_voice = gr.Dropdown(list(VOICES), value="Weiblich", label="Stimme", scale=3, min_width=120)
tpl_language = gr.Dropdown(list(LANGUAGES), value="Deutsch", label="Sprache", scale=3, min_width=120)
with gr.Row():
tpl_title = gr.Textbox(label="Titel / Marke", placeholder="wird im Hook gesungen", lines=1)
tpl_extra = gr.Textbox(label="Zusatz", placeholder="z. B. vinyl warmth, female choir", lines=1)
with gr.Column(scale=6, min_width=280):
tpl_detail = gr.HTML(
template_detail_html(None, [], "minimax-music3", template_vars(120, "beliebig", "Neutral", "", 30, False, "Weiblich", "Deutsch", "", ""), registry, []),
padding=False, elem_id="tb-tpl-detail",
)
with gr.Row(elem_classes="tb-tpl-actions"):
tpl_insert = gr.Button("⤵ In Prompt einsetzen", variant="primary", interactive=False, elem_id="tb-tpl-insert")
tpl_append = gr.Button("+ Nur Sound anhängen", interactive=False)
tpl_fav = gr.Button("☆ Merken", interactive=False, elem_id="tb-tpl-fav")
tpl_delete = gr.Button("Eigene löschen", visible=False, variant="stop", size="sm")
with gr.Row(elem_classes="tb-tpl-save"):
tpl_save_name = gr.Textbox(placeholder="Name für eigene Vorlage …", show_label=False, container=False, scale=3)
tpl_save = gr.Button("✦ Aktuelle Felder als Vorlage speichern", scale=2)
with gr.Accordion("Bausteine & Schnellrezepte", open=False, elem_classes="tb-acc tb-acc--chips"):
with gr.Row(elem_classes="recipe-grid"):
for recipe_name in PROMPT_RECIPES:
recipe_button = gr.Button(recipe_name, size="sm", elem_classes="recipe-button")
recipe_bindings.append((recipe_button, recipe_name))
for category, (target, keywords) in KEYWORD_LEXICON.items():
with gr.Row(elem_classes="chip-row"):
gr.HTML(f'<span class="tb-chip-cat">{escape(category)}</span>', padding=False)
for keyword in keywords:
chip = gr.Button(keyword, size="sm", elem_classes="tag-chip")
lexicon_bindings.append((chip, target, keyword))
with gr.Accordion("KI-Producer · Song in Alltagssprache beschreiben", open=False, elem_classes=["tb-acc", "agent-card"], elem_id="tb-agent"):
producer_chat = gr.Chatbot(value=[], height=230, show_label=False, elem_id="tb-producer-chat")
with gr.Row(elem_classes="agent-composer"):
producer_input = gr.Textbox(
placeholder="Beschreibe deinen Song in Alltagssprache …",
show_label=False, scale=5, container=False,
)
producer_send = gr.Button("Senden →", variant="primary", scale=0, min_width=110)
producer_preview = gr.HTML(value="", visible=False, elem_classes="agent-preview")
with gr.Row(elem_classes="proposal-actions"):
discard_proposal = gr.Button("Verwerfen", size="sm", visible=False)
apply_proposal = gr.Button("Vorschlag übernehmen", variant="primary", interactive=False)
# -------------------------------------------------- Seitenleiste: Verlauf
with gr.Column(elem_id="tb-side", scale=1, min_width=300):
history_create = gr.HTML(history_html(None, None, registry, "create"), padding=False, elem_classes="tb-history-wrap")
library_create = gr.Radio(
choices=[], label="Session-Tracks", show_label=False,
elem_classes=["library-radio", "tb-lib-radio"], elem_id="lib-create",
)
library_empty_create = gr.Markdown("Keine Tracks in dieser Session.", elem_classes="library-empty", visible=False)
with gr.Column(elem_classes="tb-side__card"):
gr.HTML('<div class="tb-side__title">Gewählter Take</div>', padding=False)
library_preview_audio_create = gr.Audio(
label="Vorhören", type="filepath", interactive=False,
elem_classes="library-preview", waveform_options=wave_opts,
)
library_analysis_create = gr.HTML(value="", visible=False)
with gr.Row(elem_classes="library-actions"):
library_preview_create = gr.Button("▶ In Player", size="sm")
library_download_create = gr.DownloadButton("⬇ WAV", interactive=False, size="sm")
with gr.Row(elem_classes="library-actions"):
reuse_create = gr.Button("↺ Einstellungen", size="sm", elem_id="tb-reuse")
library_remix_create = gr.Button("✦ Smart Remix", size="sm", elem_classes="remix-action")
with gr.Row(elem_classes="library-actions"):
library_master_create = gr.Button("Master →", size="sm", elem_classes="master-arrow")
library_delete_create = gr.Button("Löschen", size="sm", elem_classes="delete-action", variant="stop")
gr.Markdown(SMART_REMIX_HELP, elem_classes="remix-help")
library_upload_create = gr.File(
label="WAV importieren", file_types=[".wav"], type="filepath",
elem_classes="library-upload", height=120,
)
# ---------------------------------------------- 04 Generieren
with gr.Row(elem_id="tb-generate-bar", equal_height=True):
gen_summary = gr.HTML(_gen_summary("minimax-music3", 12, False, 0, registry), padding=False, elem_id="tb-gensum")
generate = gr.Button("▶ Generieren", variant="primary", scale=0, min_width=220, elem_id="tb-generate")
# ============================================================ Raum MASTER
with gr.Column(visible=False, elem_id="tb-master-panel") as master_panel:
with gr.Row(elem_classes="tb-room-grid", equal_height=False):
with gr.Column(elem_id="tb-master-work", scale=3, min_width=320):
back_create = gr.Button("← Zurück zu Generate", size="sm", elem_id="tb-back")
with gr.Column(elem_classes=["tb-sec", "master-workspace"]):
gr.HTML(_section_head("A/B", "Mastering", "Original bleibt unverändert – Ergebnis als neuer Take.", "tb-step-master"), padding=False)
master_source = gr.Markdown("**Quelle:** Noch kein Track gewählt")
with gr.Row():
master_source_audio = gr.Audio(label="A · Original", type="filepath", interactive=False, waveform_options=wave_opts)
master_result_audio = gr.Audio(label="B · Master", type="filepath", interactive=False, waveform_options=wave_opts)
master_report = gr.HTML(value="", visible=False)
master_run = gr.Button("Mastering starten · Analyse & Run", variant="primary", elem_id="tb-master-run")
with gr.Column(elem_classes=["tb-sec", "tb-sec--params"]):
gr.HTML(_section_head("⚙", "Master-Parameter", "Profil, Purity und Intensität.", "tb-step-master-params"), padding=False)
with gr.Row():
master_profile = gr.Dropdown(
[
("Streaming Safe", "streaming_safe"),
("Streaming Loud", "streaming_loud"),
("Club", "club"),
("Dynamic Hi-Fi", "dynamic_hifi"),
("Premium Clarity", "premium_clarity"),
("Audiophile Preserve", "audiophile_preserve"),
],
value="streaming_safe",
label="Profil",
info="Zielklang des Masters",
)
master_purity = gr.Dropdown(
[
("Soft · Air behalten", "soft_keep_air"),
("Purity aus", "off"),
("Strict · Bestätigung nötig", "strict"),
],
value="soft_keep_air",
label="Purity-Modus",
info="Soft ist für fast alles richtig",
)
with gr.Row():
master_intensity = gr.Slider(0, 100, value=50, step=1, label="Intensität", info="50 = ausgewogen · höher = lauter & dichter")
master_analyze_only = gr.Checkbox(label="Nur analysieren", value=False, info="Kein neuer Take, nur Messwerte")
with gr.Accordion("Erweitert · Purity-Feintuning", open=False, elem_classes="tb-acc"):
with gr.Row():
master_keep_air = gr.Slider(-48, 0, value=-20, step=1, label="Air erhalten · dB", info="Wie viel Höhen-Luft geschützt bleibt")
master_margin = gr.Slider(0.5, 2.9, value=2.0, step=0.1, label="HPSS Margin", info="Trennschärfe Ton/Transienten")
master_strict_approval = gr.Checkbox(label="Strict-Modus ausdrücklich bestätigen", value=False, info="Nur nötig, wenn Purity-Modus = Strict")
gr.Markdown("Das Original bleibt immer unverändert – das Ergebnis landet als neuer Take im Verlauf.", elem_classes="tb-hint")
with gr.Column(elem_classes=["tb-sec", "agent-card"]):
gr.HTML(_section_head("AI", "Mastering-Agent", "Sag in Alltagssprache, wie es klingen soll.", "tb-step-master-agent"), padding=False)
master_chat = gr.Chatbot(value=[], height=210, show_label=False)
with gr.Row(elem_classes="agent-composer"):
master_input = gr.Textbox(placeholder="Mehr Druck, aber Dynamik behalten …", show_label=False, scale=5, container=False)
master_send = gr.Button("Senden →", variant="primary", scale=0, min_width=110)
master_preview = gr.HTML(value="", visible=False)
apply_master_proposal = gr.Button("Master-Vorschlag übernehmen", interactive=False)
with gr.Column(elem_id="tb-master-side", scale=1, min_width=300):
history_master = gr.HTML(history_html(None, None, registry, "master"), padding=False, elem_classes="tb-history-wrap")
library_master = gr.Radio(
choices=[], label="Session-Tracks", show_label=False,
elem_classes=["library-radio", "tb-lib-radio"], elem_id="lib-master",
)
library_empty_master = gr.Markdown("Keine Tracks in dieser Session.", elem_classes="library-empty", visible=False)
with gr.Column(elem_classes="tb-side__card"):
gr.HTML('<div class="tb-side__title">Gewählter Take</div>', padding=False)
library_preview_audio_master = gr.Audio(
label="Vorhören", type="filepath", interactive=False,
elem_classes="library-preview", waveform_options=wave_opts,
)
library_analysis_master = gr.HTML(value="", visible=False)
with gr.Row(elem_classes="library-actions"):
library_preview_master = gr.Button("▶ In Player", size="sm")
library_download_master = gr.DownloadButton("⬇ WAV", interactive=False, size="sm")
with gr.Row(elem_classes="library-actions"):
library_remix_master = gr.Button("✦ Smart Remix", size="sm", elem_classes="remix-action")
library_delete_master = gr.Button("Löschen", size="sm", elem_classes="delete-action", variant="stop")
gr.Markdown(SMART_REMIX_HELP, elem_classes="remix-help")
library_upload_master = gr.File(
label="WAV importieren", file_types=[".wav"], type="filepath",
elem_classes="library-upload", height=120,
)
# ============================================================ Raum STUDIO
with gr.Column(visible=False, elem_id="tb-studio-panel") as studio_panel:
gr.HTML(DAW_HTML, padding=False, elem_id="daw-wrap")
with gr.Accordion("Waben-Kette & Studio-Agent (klassisch)", open=False, elem_id="st-classic"):
with gr.Row(elem_classes="st-layout", equal_height=False):
with gr.Column(elem_id="tb-studio-work", scale=3, min_width=320):
studio_hive = gr.HTML(studio_empty_html(), padding=False, elem_id="st-hive-wrap")
studio_chat_view = gr.HTML(studio_chat_html([]), padding=False, elem_id="st-chat-wrap")
studio_sel_label = gr.Markdown("Kein Abschnitt markiert", elem_classes="st-sel")
with gr.Row():
studio_sel_start = gr.Number(value=0, label="Abschnitt Start · s", precision=2, scale=1)
studio_sel_end = gr.Number(value=0, label="Abschnitt Ende · s", precision=2, scale=1)
studio_sel_apply = gr.Button("Abschnitt setzen", scale=0, min_width=140)
with gr.Column(elem_classes="st-prompt", elem_id="st-prompt"):
studio_chips = gr.HTML(
studio_chips_html(["Dynamik", "Streaming −14 LUFS", "Club Master", "Export"]),
padding=False,
)
with gr.Row(elem_classes="st-prompt__row"):
studio_input = gr.Textbox(
placeholder="▸ Sag dem Studio, was du brauchst …",
show_label=False,
container=False,
scale=5,
elem_id="st-input",
)
studio_send = gr.Button("→", variant="primary", scale=0, min_width=64, elem_id="st-send")
studio_download = gr.DownloadButton("⬇ Export", interactive=False, elem_id="st-download")
# versteckte Bridge für Chips / Waben-Aktionen
studio_bridge = gr.Textbox(value="", visible=False, elem_id="st-bridge")
studio_bridge_btn = gr.Button("bridge", visible=False, elem_id="st-bridge-btn")
with gr.Column(elem_id="tb-studio-side", scale=1, min_width=280):
gr.HTML('<div class="tb-side__title">Takes · Studio</div>', padding=False)
library_studio = gr.Radio(
choices=[], label="Session-Tracks", show_label=False,
elem_classes=["library-radio", "tb-lib-radio"], elem_id="lib-studio",
)
with gr.Column(elem_classes="st-side__card"):
studio_preview = gr.Audio(
label="Studio-Vorschau", type="filepath", interactive=False,
waveform_options=wave_opts,
)
studio_upload_file = gr.File(
label="WAV/MP3 laden", file_types=[".wav", ".mp3"], type="filepath",
elem_classes="library-upload", height=110,
)
gr.Markdown(
"Mehrere Takes laden für Transition · Chips sind Tipps von 3B.",
elem_classes="tb-hint",
)
# ============================================================ 05 Ergebnis (alle Räume)
with gr.Column(elem_classes="player-stage", elem_id="tb-player"):
gr.HTML(_section_head("05", "Ergebnis", "Anhören, exportieren oder direkt ins Mastering schicken.", "tb-step-result"), padding=False, elem_id="tb-result-head")
audio = gr.Audio(
label="Aktueller Track",
type="filepath",
interactive=False,
elem_classes="wave-audio",
waveform_options=wave_opts,
)
with gr.Row(elem_classes="transport"):
download = gr.DownloadButton("⬇ Export", interactive=False, scale=0, min_width=110, elem_id="tb-download")
stop = gr.Button("■ Stopp", scale=0, min_width=100, elem_id="tb-stop", variant="stop")
status = gr.Markdown(
"Bereit · Modell wählen, Sound beschreiben (oder leer lassen für ein Beispiel) und ▶ Generieren drücken.",
elem_classes="status-strip",
elem_id="tb-status",
)
master_arrow = gr.Button("Master →", scale=0, min_width=130, elem_classes="master-arrow", elem_id="tb-master-arrow")
loading = gr.HTML(value=LOADER_HTML, visible=False, elem_id="tb-loading")
# ===================================================================
# Ereignisse – bestehende Verdrahtung (unverändert in Logik)
# ===================================================================
private = {"api_visibility": "private"}
daw_register(app)
def _generate_handler(session, model_id, genre_text, lyrics_text, negative_text, instrumental_enabled,
seconds, seed_value, steps_value, guidance_value, request: gr.Request):
has_visitor_token = _visitor_token(request)
args = (session, registry, model_id, genre_text, lyrics_text, negative_text,
instrumental_enabled, seconds, seed_value, steps_value, guidance_value)
result = _generate(*args)
if has_visitor_token and _is_quota_status(result[4]):
# Besucher-Kontingent leer → einmal mit dem Space-Kontingent wiederholen
set_visitor_ip_token(None)
result = _generate(*args)
return result
def _remix_handler(track_id, session, request: gr.Request):
has_visitor_token = _visitor_token(request)
result = _smart_remix(track_id, session, registry)
if has_visitor_token and _is_quota_status(result[6]):
set_visitor_ip_token(None)
result = _smart_remix(track_id, session, registry)
return result
generation_start = generate.click(_show_loading, outputs=[status, loading], **private)
generation_event = generation_start.then(
fn=_generate_handler,
inputs=[session_state, model, genre, lyrics, negative, instrumental, duration, seed, steps, guidance],
outputs=[
audio, library_create, library_master, download, status, master_source, session_state, loading,
library_download_create, library_download_master, library_analysis_create, library_analysis_master,
library_empty_create, library_empty_master, library_preview_audio_create, library_preview_audio_master,
],
api_name="generate",
)
generation_event.failure(_job_failed, outputs=[status, loading], queue=False, **private)
stop.click(_stop_loading, outputs=[status, loading], cancels=[generation_event], **private)
master_arrow.click(
_open_master,
inputs=session_state,
outputs=[create_panel, master_panel, studio_panel, status, master_source, session_state, master_source_audio],
)
back_create.click(
_open_create,
inputs=session_state,
outputs=[create_panel, master_panel, studio_panel, status, session_state],
)
room_generate.click(
_open_create,
inputs=session_state,
outputs=[create_panel, master_panel, studio_panel, status, session_state],
**private,
)
room_master.click(
_open_master_room,
inputs=session_state,
outputs=[create_panel, master_panel, studio_panel, status, master_source, session_state, master_source_audio],
**private,
)
room_studio.click(
studio_open,
inputs=session_state,
outputs=[
create_panel, master_panel, studio_panel, status, session_state,
studio_preview, studio_hive, studio_chips, studio_chat_view, studio_sel_label, studio_download,
],
**private,
)
for trigger in (producer_send, producer_input):
event_method = trigger.click if isinstance(trigger, gr.Button) else trigger.submit
event_method(
lambda message, session: _producer_reply(message, session, producer_agent),
inputs=[producer_input, session_state],
outputs=[producer_input, producer_chat, proposal_state, producer_preview, apply_proposal, discard_proposal, session_state],
api_name="producer_chat" if trigger is producer_send else False,
)
apply_proposal.click(
lambda proposal, session: _apply_proposal(proposal, session, registry),
inputs=[proposal_state, session_state],
outputs=[model, genre, lyrics, negative, instrumental, duration, steps, guidance, status, session_state, apply_proposal],
api_name="_apply_proposal",
)
discard_proposal.click(
_discard_proposal,
outputs=[proposal_state, producer_preview, apply_proposal, discard_proposal, status],
)
instrumental.input(_instrumental_update, inputs=[instrumental, lyrics], outputs=[lyrics, status], **private)
model.change(
lambda model_id, seconds: _model_selection_update(model_id, seconds, registry),
inputs=[model, duration],
outputs=[model_status, duration, steps, guidance, status],
**private,
)
model.change(_studio_visibility, inputs=model, outputs=studio_acc, queue=False, **private)
model.change(lambda m: model_chips_html(registry, m), inputs=model, outputs=model_chips, **private)
def _refresh_status(m):
return model_chips_html(registry, m), gr.Dropdown(choices=refresh_model_labels_cached(registry), value=m), _model_status_html(m, registry)
refresh_chips.click(_refresh_status, inputs=model, outputs=[model_chips, model, model_status], **private)
for comp in (model, duration, instrumental, seed):
comp.change(
lambda m, d, i, s: _gen_summary(m, d, i, s, registry),
inputs=[model, duration, instrumental, seed],
outputs=gen_summary, queue=False, **private,
)
# Programmatisch gefüllte Lyrics/Ausschlüsse sichtbar machen (Vorlage, KI-Producer, Wiederverwenden)
lyrics.change(lambda t: gr.Accordion(open=True) if (t or "").strip() and (t or "").strip().casefold() != "[instrumental]" else gr.skip(),
inputs=lyrics, outputs=lyrics_acc, queue=False, **private)
negative.change(lambda t: gr.Accordion(open=True) if (t or "").strip() else gr.skip(),
inputs=negative, outputs=negative_acc, queue=False, **private)
random_seed.click(lambda: random.randint(1, 2**31 - 1), outputs=seed, **private)
clear_genre.click(lambda: "", outputs=genre, queue=False, **private)
clear_lyrics.click(_clear_lyrics, inputs=instrumental, outputs=lyrics, queue=False, **private)
clear_negative.click(lambda: "", outputs=negative, queue=False, **private)
for chip, target, keyword in lexicon_bindings:
target_component = genre if target == "genre" else negative
chip.click(
lambda current_value, word=keyword: _append_keyword(current_value, word),
inputs=target_component, outputs=target_component, queue=False, **private,
)
for recipe_button, recipe_name in recipe_bindings:
recipe_button.click(
lambda genre_value, negative_value, name=recipe_name: _apply_prompt_recipe(genre_value, negative_value, name),
inputs=[genre, negative], outputs=[genre, negative, status], queue=False, **private,
)
for studio_box in (studio_global, studio_vocal, studio_arrangement):
studio_box.change(
_set_studio_fields,
inputs=[studio_global, studio_vocal, studio_arrangement, session_state],
outputs=session_state, **private,
)
master_outputs = [master_input, master_chat, master_proposal_state, master_preview, apply_master_proposal, session_state]
master_send.click(_master_reply, inputs=[master_input, session_state], outputs=master_outputs)
master_input.submit(_master_reply, inputs=[master_input, session_state], outputs=master_outputs, api_name=False)
apply_master_proposal.click(
_apply_master_proposal,
inputs=[master_proposal_state, session_state],
outputs=[master_profile, master_purity, master_keep_air, master_margin, master_intensity,
master_analyze_only, master_strict_approval, status, session_state],
)
load_outputs_create = [audio, library_preview_audio_create, download, library_download_create, status,
master_source, session_state, library_analysis_create]
load_outputs_master = [audio, library_preview_audio_master, download, library_download_master, status,
master_source, session_state, library_analysis_master]
library_create.input(_load_track, inputs=[library_create, session_state], outputs=load_outputs_create)
library_master.input(_load_track, inputs=[library_master, session_state], outputs=load_outputs_master, api_name=False)
library_master.input(
lambda track_id, session: (
session.track_by_id(track_id).path if session and session.track_by_id(track_id) else None
),
inputs=[library_master, session_state],
outputs=master_source_audio,
**private,
)
library_preview_create.click(_load_track, inputs=[library_create, session_state], outputs=load_outputs_create, api_name=False)
library_preview_master.click(_load_track, inputs=[library_master, session_state], outputs=load_outputs_master, api_name=False)
delete_outputs = [
audio, library_create, library_master, download, library_download_create, library_download_master,
status, master_source, session_state, library_analysis_create, library_analysis_master,
library_empty_create, library_empty_master, library_preview_audio_create, library_preview_audio_master,
]
library_delete_create.click(_delete_selected, inputs=[library_create, session_state], outputs=delete_outputs)
library_delete_master.click(_delete_selected, inputs=[library_master, session_state], outputs=delete_outputs, api_name=False)
library_master_create.click(
_master_selected,
inputs=[library_create, session_state],
outputs=[create_panel, master_panel, studio_panel, status, master_source, session_state, audio, library_master,
download, library_download_master, library_analysis_master, library_preview_audio_master,
master_source_audio],
)
reuse_create.click(
lambda track_id, session: _reuse_settings(track_id, session, registry),
inputs=[library_create, session_state],
outputs=[model, genre, lyrics, negative, instrumental, duration, seed,
studio_global, studio_vocal, studio_arrangement, status, session_state],
api_name="reuse_settings",
)
master_start = master_run.click(_show_master_loading, outputs=[status, loading], **private)
master_event = master_start.then(
_run_onus,
inputs=[library_master, session_state, master_profile, master_purity, master_keep_air, master_margin,
master_intensity, master_analyze_only, master_strict_approval],
outputs=[audio, master_source_audio, master_result_audio, library_create, library_master, download,
library_download_create, library_download_master, status, master_source, session_state,
master_report, library_analysis_create, library_analysis_master, library_preview_audio_create,
library_preview_audio_master, loading, library_empty_create, library_empty_master, master_preview],
)
master_event.failure(_job_failed, outputs=[status, loading], queue=False, **private)
stop.click(_stop_loading, outputs=[status, loading], cancels=[master_event], **private)
remix_outputs = [
audio, library_create, library_master, download, library_download_create, library_download_master,
status, master_source, session_state, loading, library_analysis_create, library_analysis_master,
library_empty_create, library_empty_master, model, genre, lyrics, negative, instrumental,
library_preview_audio_create, library_preview_audio_master,
]
remix_create_start = library_remix_create.click(_show_remix_loading, outputs=[status, loading], **private)
remix_create_event = remix_create_start.then(
_remix_handler,
inputs=[library_create, session_state], outputs=remix_outputs, api_name="smart_remix",
)
remix_master_start = library_remix_master.click(_show_remix_loading, outputs=[status, loading], **private)
remix_master_event = remix_master_start.then(
_remix_handler,
inputs=[library_master, session_state], outputs=remix_outputs, api_name=False,
)
for _ev in (remix_create_event, remix_master_event):
_ev.failure(_job_failed, outputs=[status, loading], queue=False, **private)
stop.click(_stop_loading, outputs=[status, loading], cancels=[remix_create_event, remix_master_event], **private)
upload_outputs = [
audio, library_preview_audio_create, library_preview_audio_master, library_create, library_master,
download, library_download_create, library_download_master, status, master_source, session_state,
library_analysis_create, library_analysis_master, library_empty_create, library_empty_master,
]
library_upload_create.change(_import_library_wav, inputs=[library_upload_create, session_state], outputs=upload_outputs)
library_upload_master.change(_import_library_wav, inputs=[library_upload_master, session_state], outputs=upload_outputs, api_name=False)
# Verlauf mit Waveforms folgt jeder Änderung der Bibliothek
# Verlauf nur aus dem Session-State bauen: Radio-Werte als Input scheitern nach dem Löschen
# (gelöschte ID ist kurz noch gesetzt → „Value … is not in the list of choices“).
def _histories_from_session(session):
current = _ensure_session(session)
track = current.current_track()
sel = track.id if track else None
return _histories(current, sel, sel, registry)
for lib in (library_create, library_master):
lib.change(
_histories_from_session,
inputs=[session_state],
outputs=[history_create, history_master],
**private,
)
# ===================================================================
# STUDIO-Raum
# ===================================================================
def _studio_bridge(cmd: str, session: StudioSession | None):
from studio.studio_room.handlers import (
studio_chat,
studio_chip_click,
studio_remove_wabe,
studio_reorder_wabe,
studio_set_macro,
studio_toggle_wabe,
)
from studio.studio_room.catalog import validate_wabe_dict
from studio.studio_room.html_ui import honeycomb_html
cmd = (cmd or "").strip()
if not cmd:
return (gr.skip(),) * 10
if cmd.startswith("chip:"):
return studio_chip_click(cmd[5:], session)
if cmd.startswith("toggle:"):
hive, preview, status_txt, sess = studio_toggle_wabe(cmd[7:], session)
return ("", hive, gr.skip(), gr.skip(), gr.skip(), preview, status_txt, sess, gr.skip(), gr.skip(), gr.skip())
if cmd.startswith("remove:"):
hive, preview, status_txt, sess = studio_remove_wabe(cmd[7:], session)
return ("", hive, gr.skip(), gr.skip(), gr.skip(), preview, status_txt, sess, gr.skip(), gr.skip(), gr.skip())
if cmd.startswith("up:") or cmd.startswith("down:"):
direction, wid = ("up", cmd[3:]) if cmd.startswith("up:") else ("down", cmd[5:])
hive, preview, status_txt, sess = studio_reorder_wabe(wid, direction, session)
return ("", hive, gr.skip(), gr.skip(), gr.skip(), preview, status_txt, sess, gr.skip(), gr.skip(), gr.skip())
if cmd.startswith("collapse:"):
current = _ensure_session(session)
wid = cmd.split(":", 1)[1]
waben = []
for raw in current.studio_room.waben:
try:
w = validate_wabe_dict(raw)
if w.id == wid:
w.collapsed = not w.collapsed
waben.append(w)
except Exception:
continue
current.studio_room.waben = [w.model_dump() for w in waben]
return ("", honeycomb_html(waben, current.studio_room.analysis), gr.skip(), gr.skip(), gr.skip(), gr.skip(), "Wabe umgeschaltet", current, gr.skip(), gr.skip(), gr.skip())
if cmd.startswith("macro:"):
parts = cmd.split(":")
if len(parts) >= 4:
hive, preview, status_txt, sess = studio_set_macro(parts[1], parts[2], float(parts[3]), session)
return ("", hive, gr.skip(), gr.skip(), gr.skip(), preview, status_txt, sess, gr.skip(), gr.skip(), gr.skip())
return studio_chat(cmd, session)
studio_chat_outputs = [
studio_input, studio_hive, studio_chips, studio_chat_view, studio_sel_label,
studio_preview, status, session_state, studio_download, library_create, library_master,
]
studio_send.click(
lambda message, session, a, b: studio_chat(message, session, a, b),
inputs=[studio_input, session_state, studio_sel_start, studio_sel_end],
outputs=studio_chat_outputs,
api_name="studio_chat",
).then(_histories_from_session, inputs=[session_state], outputs=[history_create, history_master], **private)
studio_input.submit(
lambda message, session, a, b: studio_chat(message, session, a, b),
inputs=[studio_input, session_state, studio_sel_start, studio_sel_end],
outputs=studio_chat_outputs,
api_name=False,
).then(_histories_from_session, inputs=[session_state], outputs=[history_create, history_master], **private)
studio_bridge_btn.click(
_studio_bridge,
inputs=[studio_bridge, session_state],
outputs=studio_chat_outputs,
**private,
)
studio_upload_file.change(
studio_upload,
inputs=[studio_upload_file, session_state],
outputs=[
studio_preview, studio_hive, studio_chips, studio_chat_view, studio_sel_label,
status, session_state, library_create, library_master,
],
api_name="studio_upload",
).then(_histories_from_session, inputs=[session_state], outputs=[history_create, history_master], **private)
library_studio.input(
studio_load_track,
inputs=[library_studio, session_state],
outputs=[studio_preview, studio_hive, studio_chips, studio_chat_view, studio_sel_label, status, session_state],
api_name="studio_load_track",
)
studio_sel_apply.click(
lambda a, b, session: __import__("studio.studio_room.handlers", fromlist=["studio_set_selection"]).studio_set_selection(a, b, session),
inputs=[studio_sel_start, studio_sel_end, session_state],
outputs=[studio_sel_label, session_state],
**private,
)
# Keep studio library choices in sync with create/master libraries
for lib in (library_create, library_master):
lib.change(
lambda session: gr.Radio(choices=library_choices(session), value=None),
inputs=session_state,
outputs=library_studio,
**private,
)
# ===================================================================
# Vorlagen-Bibliothek
# ===================================================================
tpl_vars = [tpl_bpm, tpl_key, tpl_mood, tpl_instruments, tpl_duration, tpl_vocals,
tpl_voice, tpl_language, tpl_title, tpl_extra]
assert len(tpl_vars) == _TPL_VAR_COUNT
def _refresh_list(category, query, favs, custom, current):
return (
_tpl_choices_update(category, query, favs, custom, current),
_tpl_count_html(category, query, favs, custom),
)
# Auswahl lebt in tpl_sel (State): tpl_pick (Radio) wird nie als Input gelesen,
# sonst lehnt Gradio Werte ab, die nach dem Filtern nicht mehr in den Choices stehen.
tpl_sel = gr.State(None)
for trig in (tpl_category.change, tpl_search.input):
trig(_refresh_list, inputs=[tpl_category, tpl_search, tpl_favs, tpl_custom, tpl_sel],
outputs=[tpl_pick, tpl_count], queue=False, trigger_mode="always_last", **private)
app.load(_refresh_list, inputs=[tpl_category, tpl_search, tpl_favs, tpl_custom, tpl_sel],
outputs=[tpl_pick, tpl_count], queue=False, **private)
app.load(_refresh_status, inputs=model, outputs=[model_chips, model, model_status], **private)
def _select(template_id, custom, model_id, favs):
tv = template_field_defaults(template_id, custom) if template_id else None
if tv is None:
# Gefiltert/abgewählt: letzte Auswahl bleibt aktiv
return (gr.skip(), *[gr.skip()] * _TPL_VAR_COUNT, gr.skip(), gr.skip(), gr.skip(), gr.skip(), gr.skip())
_template, own = resolve_template(template_id, custom)
detail = template_detail_html(template_id, custom, model_id, tv, registry, favs)
return (
template_id, tv.bpm, tv.key, tv.mood, tv.instruments, tv.duration_s, tv.vocals, tv.voice, tv.language,
tv.title, tv.extra, detail, gr.Button(interactive=True), gr.Button(interactive=True),
_fav_label(template_id, favs), gr.Button(visible=own is not None),
)
tpl_pick.change(
_select, inputs=[tpl_pick, tpl_custom, model, tpl_favs],
outputs=[tpl_sel, *tpl_vars, tpl_detail, tpl_insert, tpl_append, tpl_fav, tpl_delete], queue=False, **private,
)
def _preview(template_id, custom, model_id, favs, *values):
if not template_id:
return gr.skip()
return template_detail_html(template_id, custom, model_id, template_vars(*values), registry, favs)
for comp in tpl_vars:
comp.input(_preview, inputs=[tpl_sel, tpl_custom, model, tpl_favs, *tpl_vars], outputs=tpl_detail,
queue=False, trigger_mode="always_last", **private)
model.change(_preview, inputs=[tpl_sel, tpl_custom, model, tpl_favs, *tpl_vars], outputs=tpl_detail,
queue=False, **private)
def _insert(template_id, custom, model_id, session, append_only, cur_genre, cur_negative, *values):
current = _ensure_session(session)
tv = template_vars(*values)
template, own, rendered = render_any(template_id, custom, model_id, tv, registry)
if rendered is None:
raise gr.Error("Bitte zuerst eine Vorlage wählen.")
name = template.name if template else str(own.get("name"))
notes = (" · " + " ".join(rendered.notes)) if rendered.notes else ""
if append_only:
merged = _append_keyword(cur_genre or "", rendered.genre)
neg = _append_keyword(cur_negative or "", rendered.negative) if rendered.negative else cur_negative
return (merged, gr.skip(), neg, gr.skip(), gr.skip(), gr.skip(), gr.skip(), gr.skip(),
f"+ Sound aus „{name}“ angehängt{notes}", current)
current.create.genre_style_instruments = rendered.genre
current.create.lyrics = rendered.lyrics
current.create.negative = rendered.negative
current.create.duration_s = rendered.duration_s
current.create.studio = dict(rendered.studio)
return (
rendered.genre,
gr.Textbox(value=rendered.lyrics, interactive=not rendered.instrumental),
rendered.negative,
rendered.instrumental,
min(rendered.duration_s, 180),
rendered.studio.get("global_metadata", ""),
rendered.studio.get("vocal_details", ""),
rendered.studio.get("arrangement", ""),
f"✓ Vorlage „{name}“ eingesetzt · formatiert für {registry.get(model_id).label}{notes}",
current,
)
insert_outputs = [genre, lyrics, negative, instrumental, duration, studio_global, studio_vocal,
studio_arrangement, status, session_state]
tpl_insert.click(
lambda tid, cu, m, s, g, n, *v: _insert(tid, cu, m, s, False, g, n, *v),
inputs=[tpl_sel, tpl_custom, model, session_state, genre, negative, *tpl_vars],
outputs=insert_outputs, **private,
)
tpl_append.click(
lambda tid, cu, m, s, g, n, *v: _insert(tid, cu, m, s, True, g, n, *v),
inputs=[tpl_sel, tpl_custom, model, session_state, genre, negative, *tpl_vars],
outputs=insert_outputs, **private,
)
def _toggle_fav(template_id, favs, category, query, custom, model_id, *values):
favs = list(favs or [])
if not template_id:
return favs, gr.skip(), gr.skip(), _fav_label(None, favs), gr.skip()
if template_id in favs:
favs.remove(template_id)
else:
favs.append(template_id)
detail = template_detail_html(template_id, custom, model_id, template_vars(*values), registry, favs)
choices, count = _refresh_list(category, query, favs, custom, template_id)
return favs, choices, count, _fav_label(template_id, favs), detail
tpl_fav.click(
_toggle_fav,
inputs=[tpl_sel, tpl_favs, tpl_category, tpl_search, tpl_custom, model, *tpl_vars],
outputs=[tpl_favs, tpl_pick, tpl_count, tpl_fav, tpl_detail], queue=False, **private,
)
def _save(name, genre_v, lyrics_v, negative_v, seconds, instr, model_id, custom, favs):
if not (genre_v or "").strip():
raise gr.Error("Erst das Sound-Feld füllen – dann als Vorlage speichern.")
custom = list(custom or [])[-39:]
entry = make_custom(name, genre_v, lyrics_v, negative_v, int(seconds), bool(instr), model_id,
str(int(time.time() * 1000)))
custom.append(entry)
return (
custom, "custom", entry["id"],
_tpl_choices_update("custom", "", favs, custom, entry["id"]),
_tpl_count_html("custom", "", favs, custom), "",
f"✦ Eigene Vorlage „{entry['name']}“ im Browser gespeichert.",
)
tpl_save.click(
_save,
inputs=[tpl_save_name, genre, lyrics, negative, duration, instrumental, model, tpl_custom, tpl_favs],
outputs=[tpl_custom, tpl_category, tpl_sel, tpl_pick, tpl_count, tpl_save_name, status], queue=False, **private,
)
def _delete_custom(template_id, custom, favs, category, query):
custom = [c for c in (custom or []) if c.get("id") != template_id]
favs = [f for f in (favs or []) if f != template_id]
choices, count = _refresh_list(category, query, favs, custom, None)
return (custom, favs, None, choices, count, gr.Button(visible=False), gr.Button(interactive=False),
gr.Button(interactive=False), _fav_label(None, favs), "Eigene Vorlage gelöscht.")
tpl_delete.click(
_delete_custom,
inputs=[tpl_sel, tpl_custom, tpl_favs, tpl_category, tpl_search],
outputs=[tpl_custom, tpl_favs, tpl_sel, tpl_pick, tpl_count, tpl_delete, tpl_insert, tpl_append, tpl_fav, status], queue=False, **private,
)
return app.queue(default_concurrency_limit=2)
from __future__ import annotations
import os
import random
import time
from hashlib import sha256
from html import escape
from pathlib import Path
from typing import Any
import gradio as gr
from studio.adapters import (
AdapterError,
AdapterRouter,
GenerateRequest,
OnusAdapter,
OnusRequest,
SpaceMasterAdapter,
route_id_for_model,
)
from studio.adapters.health import set_visitor_ip_token
from studio.agent import FieldProposal, MasterProposal, ProducerAgent, propose_master
from studio.audio_analysis import AudioAnalysis, analyze_wav
from studio.library import add_track, delete_track, import_wav, library_choices, select_track
from studio.model_status import (
is_unusable,
model_state,
model_status_html_live,
refresh_all_model_labels,
refresh_model_labels_cached,
)
from studio.registry import ModelRegistry, load_registry
from studio.remix import build_remix_plan
from studio.state import ChatTurn, StudioSession, new_session
from studio.templates import CATEGORIES, KEYS, LANGUAGES, MOODS, VOICES, filter_choices, make_custom
from studio.theme import CSS, HEAD, THEME
from studio.daw.api import register as daw_register
from studio.studio_room.handlers import (
studio_chat,
studio_chip_click,
studio_load_track,
studio_open,
studio_upload,
)
from studio.studio_room.html_ui import chips_html as studio_chips_html
from studio.studio_room.html_ui import empty_state_html as studio_empty_html
from studio.studio_room.html_ui import chat_bubble_html as studio_chat_html
from studio.studio_room.agent_3b import default_chips as studio_default_chips
from studio.ui_extras import (
history_html,
model_chips_html,
resolve_template,
reuse_values,
template_detail_html,
template_field_defaults,
template_vars,
render_any,
)
__all__ = ["CSS", "HEAD", "THEME", "build_app"]
# Buy Me a Coffee: schlichter statischer Link (kein Fremd-JS), im Look des Studios
BMC_LINK_HTML = (
'<a href="https://www.buymeacoffee.com/2btainment" target="_blank" rel="noopener noreferrer" '
'class="bmc-link tb-bmc" title="Buy me a coffee – 2Btainment unterstützen" aria-label="Buy me a coffee – 2Btainment unterstützen">'
'<i aria-hidden="true">☕</i><span>Buy me<br>a coffee</span></a>'
)
KEYWORD_LEXICON = {
"◉ Genres": ("genre", ("Dark Riddim", "Afro House", "Neo Soul", "Cinematic")),
"♬ Instrumente": (
"genre",
("Analog Bass", "Rhodes Piano", "808 Drums", "Warm Strings"),
),
"◐ Moods": ("genre", ("Hypnotic", "Euphoric", "Melancholic", "Raw & Intimate")),
"✦ Meta-Formulierungen": (
"genre",
("radio-ready hook", "dynamic arrangement", "human timing", "wide stereo image"),
),
"⊘ Negativ-Prompts": (
"negative",
("muddy low end", "harsh sibilance", "clipping", "generic AI haze"),
),
}
PROMPT_RECIPES = {
"🔥 Club-Druck": (
"club-ready low end, punchy transient design, energetic drop, controlled stereo width",
"weak kick, muddy sub bass, flat dynamics",
),
"✧ Vocal-Klarheit": (
"intimate lead vocal, present center image, clear diction, supportive arrangement",
"harsh sibilance, buried vocals, boxy mids, excessive reverb",
),
"◒ Analog-Wärme": (
"warm analog character, subtle tape saturation, human timing, organic dynamics",
"brittle highs, sterile digital tone, over-compression",
),
"◈ Cinematic-Weite": (