Remove the LLM compose step (Inference Providers cost)

Lyrics + structured caption are now editable reference nodes wired
directly into generate_song, with the official-guide defaults. Drops the
openai dep, the MiniMax-M3 prompts, OAuth token plumbing, and the
inference-api scope.

README.md

@@ -9,25 +9,19 @@ app_file: app.py
 pinned: true
 suggested_hardware: zero-a10g
 hf_oauth: true
-hf_oauth_scopes:
- - inference-api
 ---
 
 # MiniMax Music 3 — diffusers demo as a `gr.Workflow`
 
-Generates full songs from a one-line idea using the `MiniMaxMusic3Pipeline` diffusers port, presented as a
-visual, node-based `gr.Workflow` canvas (`workflow.json`):
+Generates full songs from tagged lyrics + a structured caption using the `MiniMaxMusic3Pipeline` diffusers
+port, presented as a visual, node-based `gr.Workflow` canvas (`workflow.json`):
 
-**Song Idea + Duration → `compose_song` (MiniMax-M3 LLM: tagged lyrics + structured caption per the official
+**Lyrics + Global Metadata + Vocal Details + Arrangement (editable reference nodes, defaults per the official
 prompting guide) → `generate_song` (`@spaces.GPU` ZeroGPU worker: AR frames → windowed DiT decode → vocoder →
 wav) → Output Song / Seed Used / Stats → `make_video` (CPU ffmpeg visualizer) → Share Video.**
 
-Seed, randomize-seed, flow-matching steps, guidance scale, duration, and video title are editable reference
-nodes. `hf_oauth: true` lets the owner edit the canvas; visitors sign in to run the pipeline, and the
-`compose_song` LLM step uses the visitor's own OAuth token (`inference-api` scope), so inference bills the
-visitor directly. For logged-out visitors the LLM step is skipped (it's an add-on) and the pipeline
-generates from the built-in demo lyrics + structured caption; set an `HF_TOKEN` secret to let anonymous
-runs use the owner's token instead. The original Blocks app
+Duration, seed, randomize-seed, flow-matching steps, guidance scale, and video title are also editable
+reference nodes. `hf_oauth: true` lets the owner edit the canvas; visitors can run the pipeline. The original Blocks app
 (with the live-streaming PCM player) is preserved as `app_blocks.py` — workflow fn nodes are plain callables,
 so chunk-by-chunk streaming is replaced by a final audio subject.
 

app.py

@@ -470,66 +470,6 @@ DEFAULT_ARRANGEMENT = (
     "long, starlit reverb wash."
 )
 
-_CAPTION_CONTRACT = """The three caption fields follow the exact labeled style the model was trained on. Be concrete and musical; describe an energy arc and instrument lifecycles, never a static equipment list or decorative adjectives. Never contradict an explicit user constraint: instrumental stays instrumental; never reverse a required vocal gender, tempo limit, required instrument, or exclusion. Do not quote or paraphrase lyric lines inside the caption. Total caption length roughly 250-400 words.
-
-global_metadata: one paragraph, in order: "Basic Attributes: bpm is <number>. key is <letter>, and scale is <major|minor>. <Genre / Subgenre>." then "Global Emotional Progression: <how the emotion evolves from the opening through the final section>." then "Application Scenarios & Imagery: <two or three vivid listening scenarios>." then "Sonics & Production Profile: <soundstage, frequency balance, dynamics, production character>."
-
-vocal_details: one paragraph: "Vocal Gender & Timbre: Singer A (<Male|Female>), <timbre and register>." then "Vocal Style: <delivery, and how it shifts per section>." then "Harmony/Backing Vocals: <where harmonies or doubles appear and their character>." then "Vocal FX: <restrained treatment: reverb, delay, light compression>." For instrumental pieces write "Instrumental, no vocals." and name the instrument or texture carrying the lead melodic role.
-
-arrangement: one paragraph: "Instrument Lifecycle Description (Primary/Secondary Layering): Primary: <core instruments present start to finish and their role>. Secondary: <instruments that enter, exit or intensify, and in which sections>." then "Groove & Foundation Progression: <how drums, bass and groove develop across sections>." then "Embellishments, Textures & Spatial FX: <fills, textures, transitional gestures, stereo and space treatment where relevant>." State what enters, exits, changes or intensifies for every section of the song, aligned with the lyric section tags."""
-
-_LYRICS_RULES = """lyrics: singable lyrics using ONLY these section tags, each ALWAYS ALONE on its own line: [intro] [verse] [pre-chorus] [chorus] [post-chorus] [bridge] [instrumental] [solo] [outro]. Never put words on the same line as a tag. Size the structure to the duration: <=30s: one verse + one chorus; ~60s: verse/pre-chorus/chorus/verse/chorus; >=120s: full structure with bridge and outro. Roughly 12-16 sung words per 10 seconds. Musical instructions (tempo, instruments, dynamics) never belong in the lyrics. If the song is instrumental, use [instrumental] sections with no words."""
-
-_COMPOSER_SYSTEM = f"""You write inputs for MiniMax Music 3, a lyrics+description music generation model.
-Given a song description and a target duration, produce:
-1. {_LYRICS_RULES}
-2-4. global_metadata, vocal_details, arrangement — a structured caption. {_CAPTION_CONTRACT}
-Answer with ONLY a JSON object with keys: lyrics, global_metadata, vocal_details, arrangement."""
-
-_LYRICS_SYSTEM = f"""You write lyrics for MiniMax Music 3, a lyrics+description music generation model.
-Given a lyrics instruction, the current structured prompt (global metadata, vocal details, arrangement) and a target duration, write lyrics coherent with that structured prompt.
-{_LYRICS_RULES}
-Answer with ONLY a JSON object with key: lyrics."""
-
-_PROMPT_SYSTEM = f"""You write the structured caption for MiniMax Music 3, a lyrics+description music generation model.
-Given a sound instruction and/or lyrics, produce global_metadata, vocal_details and arrangement. Build the arrangement timeline around the lyric section tags when lyrics are provided. {_CAPTION_CONTRACT}
-Answer with ONLY a JSON object with keys: global_metadata, vocal_details, arrangement."""
-
-
-def _llm_json(system, user, token=None):
-    import json as _json
-
-    from openai import OpenAI
-
-    # Prefer the visitor's OAuth token (hf_oauth + inference-api scope bills the visitor
-    # directly); fall back to the Space owner's HF_TOKEN secret for local/anonymous runs.
-    api_key = getattr(token, "token", None) or os.environ.get("HF_TOKEN")
-    if not api_key:
-        return None  # logged-out visitor, no owner secret: caller skips the LLM step
-    client = OpenAI(base_url="https://router.huggingface.co/v1", api_key=api_key)
-    completion = client.chat.completions.create(
-        model="MiniMaxAI/MiniMax-M3:fastest",
-        messages=[{"role": "system", "content": system}, {"role": "user", "content": user}],
-    )
-    text = completion.choices[0].message.content.strip()
-    if text.startswith("```"):
-        text = text.split("```")[1].removeprefix("json").strip()
-    return _json.loads(text)
-
-
-def compose_song(description, duration, token: gr.OAuthToken = None):
-    # `token` is injected by gr.Workflow (OAuthToken params are excluded from the canvas
-    # ports and filled with the signed-in visitor's OAuth token at run time).
-    if not description.strip():
-        raise gr.Error("Describe the song you want first.")
-    data = _llm_json(_COMPOSER_SYSTEM, f"Song description: {description}\nTarget duration: {int(duration)} seconds.", token)
-    if data is None:
-        # No OAuth token (logged out) and no HF_TOKEN secret: the LLM compose step is an
-        # add-on — skip it and generate from the built-in demo lyrics + structured caption.
-        return DEFAULT_LYRICS, DEFAULT_GLOBAL, DEFAULT_VOCALS, DEFAULT_ARRANGEMENT
-    return data["lyrics"], data["global_metadata"], data["vocal_details"], data["arrangement"]
-
-
 def render_video(wav_path, title):
     # Social share visualizer: warm citrus bars on a dark gradient, rendered via numpy -> ffmpeg pipe (CPU).
     if not wav_path:
@@ -637,13 +577,12 @@ def render_video(wav_path, title):
 
 
 # ---------------------------------------------------------------------------
-# gr.Workflow app. The canvas (workflow.json) wires three fn operators:
-#   compose_song   — MiniMax-M3 LLM writes lyrics + structured caption from a one-line idea
+# gr.Workflow app. The canvas (workflow.json) wires two fn operators:
 #   generate_song  — @spaces.GPU ZeroGPU worker: AR frames -> windowed DiT decode -> vocoder -> wav
 #   make_video     — CPU ffmpeg visualizer for the share video
-# Workflow fn nodes are plain callables (no streaming), so the live PCM player of the
-# Blocks version becomes a final audio subject; everything else (compose assist, structured
-# caption fields, duration/seed/steps/guidance knobs, share video, seed reporting) is kept.
+# Lyrics + structured caption are editable reference nodes (defaults from the official
+# prompting guide). Workflow fn nodes are plain callables (no streaming), so the live PCM
+# player of the Blocks version becomes a final audio subject.
 # ---------------------------------------------------------------------------
 
 import tempfile
@@ -677,7 +616,7 @@ def _friendly_gpu_error(err):
 def _generate_song_gpu(lyrics, global_meta, vocal_details, arrangement, duration, seed, randomize_seed, steps, guidance):
     caption = "\n".join(s.strip() for s in (global_meta, vocal_details, arrangement) if s and s.strip())
     if not caption:
-        raise gr.Error("Fill in the structured prompt (or run compose_song) first.")
+        raise gr.Error("Fill in the structured prompt (Global metadata / Vocal details / Arrangement) first.")
     if not lyrics or not lyrics.strip():
         raise gr.Error("Lyrics are required (section tags like [verse] must be on their own line).")
     if randomize_seed:
@@ -751,7 +690,6 @@ def make_video(audio, title: str):
 demo = gr.Workflow(
     graph="workflow.json",
     bind={
-        "compose_song": compose_song,
         "generate_song": generate_song,
         "make_video": make_video,
     },

requirements.txt

@@ -3,4 +3,3 @@ transformers
 accelerate
 scipy
 spaces
-openai

workflow.json

@@ -1,7 +1,7 @@
 {
   "schema_version": "2",
   "name": "MiniMax Music 3",
-  "description": "Full-song generation with MiniMax Music 3 on ZeroGPU: a MiniMax-M3 LLM composes tagged lyrics + a structured caption from a one-line idea, then the diffusers pipeline (AR frames -> windowed DiT decode -> vocoder) renders the song, plus an ffmpeg share-video visualizer. Tip: sign in with Hugging Face (optional) to have the compose_song LLM step write custom lyrics + caption from your Song Idea on your own inference quota \u2014 logged-out runs skip it and use the built-in demo lyrics.",
+  "description": "Full-song generation with MiniMax Music 3 on ZeroGPU: tagged lyrics + a structured caption (Global metadata / Vocal details / Arrangement, per the official prompting guide) feed the diffusers pipeline (AR frames -> windowed DiT decode -> vocoder), plus an ffmpeg share-video visualizer. Edit the Lyrics and caption reference nodes to shape the song.",
   "runtime": {
     "default": "client"
   },
@@ -10,30 +10,111 @@
   },
   "references": [
     {
-      "id": "ref_description",
-      "label": "Song Idea",
+      "id": "ref_lyrics",
+      "label": "Lyrics",
       "role": "reference",
       "asset_type": "text",
       "inputs": [
         {
           "id": "in",
-          "label": "Song Idea",
+          "label": "Lyrics",
           "type": "text"
         }
       ],
       "outputs": [
         {
           "id": "out",
-          "label": "Song Idea",
+          "label": "Lyrics",
           "type": "text"
         }
       ],
       "x": 40,
-      "y": 120,
-      "width": 260,
-      "height": 130,
+      "y": 80,
+      "width": 300,
+      "height": 200,
+      "data": {
+        "out": "[intro]\n\n[verse]\nRiding on a beam of light tonight\nEvery little star is burning bright\n[pre-chorus]\nHold your breath, the sky is opening\n[chorus]\nWe are made of sound and time\nEvery heartbeat keeps the rhyme\n[outro]"
+      }
+    },
+    {
+      "id": "ref_global",
+      "label": "Global Metadata",
+      "role": "reference",
+      "asset_type": "text",
+      "inputs": [
+        {
+          "id": "in",
+          "label": "Global Metadata",
+          "type": "text"
+        }
+      ],
+      "outputs": [
+        {
+          "id": "out",
+          "label": "Global Metadata",
+          "type": "text"
+        }
+      ],
+      "x": 40,
+      "y": 310,
+      "width": 300,
+      "height": 160,
+      "data": {
+        "out": "Basic Attributes: bpm is 120. key is C, and scale is major. Synth-Pop / Electropop. Global Emotional Progression: The track opens in shimmering anticipation, a filtered pulse like city lights coming on at dusk. The verse glides forward with hopeful momentum, the pre-chorus holds its breath as the arrangement tightens and rises, and the chorus bursts open into wide-screen euphoria \u2014 bright, weightless, celebratory. The outro drifts back down into a starry afterglow, ending on air and quiet wonder. Application Scenarios & Imagery: a night drive under neon overpasses with the windows down; a planetarium dome igniting as the lights dim; a rooftop countdown at midnight. Sonics & Production Profile: a polished, modern pop mix with a wide stereo image \u2014 airy sparkling highs, present mid-range vocals, and a tight, punchy low end; side-chained compression gives the chorus a gentle pumping lift, and the outro dissolves into long reverb tails."
+      }
+    },
+    {
+      "id": "ref_vocals",
+      "label": "Vocal Details",
+      "role": "reference",
+      "asset_type": "text",
+      "inputs": [
+        {
+          "id": "in",
+          "label": "Vocal Details",
+          "type": "text"
+        }
+      ],
+      "outputs": [
+        {
+          "id": "out",
+          "label": "Vocal Details",
+          "type": "text"
+        }
+      ],
+      "x": 40,
+      "y": 500,
+      "width": 300,
+      "height": 140,
       "data": {
-        "out": "a shimmering synth-pop anthem about a night drive under neon city lights, warm female vocals, euphoric chorus"
+        "out": "Vocal Gender & Timbre: Singer A (Female), a warm mezzo-soprano with an intimate, breathy texture in her low register and a clear, ringing brightness when she lifts. Vocal Style: soft and close-miked through the verse, phrasing like a secret; the pre-chorus rises with held, urgent notes, and the chorus opens into a confident, soaring belt with sustained tones riding the beat; over the outro she dissolves into wordless, airy ad-libs echoing the chorus melody. Harmony/Backing Vocals: a single ghost double shadows the pre-chorus; stacked parallel harmonies in thirds widen the chorus into a glowing wall; the verse stays solo and intimate. Vocal FX: light plate reverb throughout, tempo-synced delay throws on chorus line endings, subtle saturation for chorus presence, and a longer, washier reverb on the outro ad-libs."
+      }
+    },
+    {
+      "id": "ref_arrangement",
+      "label": "Arrangement",
+      "role": "reference",
+      "asset_type": "text",
+      "inputs": [
+        {
+          "id": "in",
+          "label": "Arrangement",
+          "type": "text"
+        }
+      ],
+      "outputs": [
+        {
+          "id": "out",
+          "label": "Arrangement",
+          "type": "text"
+        }
+      ],
+      "x": 40,
+      "y": 670,
+      "width": 300,
+      "height": 160,
+      "data": {
+        "out": "Instrument Lifecycle Description (Primary/Secondary Layering): Primary: a round, side-chained analog-style synth bass anchors the harmony from the first verse through the chorus, under a soft pad bed that opens the intro and never fully leaves. Secondary: a shimmering arpeggio enters at the pre-chorus and runs through the chorus; wide analog pads and a bright synth counter-melody appear only in the chorus to lift it; a sparse felt piano takes over the outro as the synths fall away. Groove & Foundation Progression: the intro pulses on a filtered four-on-the-floor kick; the verse keeps drums minimal \u2014 kick, soft clap, ticking closed hat; the pre-chorus adds open hats and a rising snare build, and the chorus lands with the full kit: punchy kick on every beat, layered claps, driving crash accents. After the chorus the drums drop out entirely, leaving piano, pad, and air for the outro. Embellishments, Textures & Spatial FX: a white-noise riser and reverse swell launch the chorus; glittering bell accents answer the vocal there; and the final piano chord rings into a long, starlit reverb wash."
       }
     },
     {
@@ -56,7 +137,7 @@
         }
       ],
       "x": 40,
-      "y": 280,
+      "y": 860,
       "width": 200,
       "height": 90,
       "data": {
@@ -82,8 +163,8 @@
           "type": "number"
         }
       ],
-      "x": 40,
-      "y": 400,
+      "x": 260,
+      "y": 860,
       "width": 200,
       "height": 90,
       "data": {
@@ -110,7 +191,7 @@
         }
       ],
       "x": 40,
-      "y": 520,
+      "y": 980,
       "width": 200,
       "height": 90,
       "data": {
@@ -136,8 +217,8 @@
           "type": "number"
         }
       ],
-      "x": 40,
-      "y": 640,
+      "x": 260,
+      "y": 980,
       "width": 200,
       "height": 90,
       "data": {
@@ -164,7 +245,7 @@
         }
       ],
       "x": 40,
-      "y": 760,
+      "y": 1100,
       "width": 200,
       "height": 90,
       "data": {
@@ -200,58 +281,6 @@
     }
   ],
   "operators": [
-    {
-      "id": "op_compose",
-      "label": "compose_song",
-      "role": "operator",
-      "kind": "fn",
-      "source": "fn",
-      "fn": "compose_song",
-      "inputs": [
-        {
-          "id": "in_0",
-          "label": "description",
-          "type": "text",
-          "required": true
-        },
-        {
-          "id": "in_1",
-          "label": "duration",
-          "type": "number"
-        }
-      ],
-      "outputs": [
-        {
-          "id": "out_0",
-          "label": "lyrics",
-          "type": "text",
-          "output_index": 0
-        },
-        {
-          "id": "out_1",
-          "label": "global_metadata",
-          "type": "text",
-          "output_index": 1
-        },
-        {
-          "id": "out_2",
-          "label": "vocal_details",
-          "type": "text",
-          "output_index": 2
-        },
-        {
-          "id": "out_3",
-          "label": "arrangement",
-          "type": "text",
-          "output_index": 3
-        }
-      ],
-      "x": 380,
-      "y": 150,
-      "width": 260,
-      "height": 180,
-      "data": {}
-    },
     {
       "id": "op_generate",
       "label": "generate_song",
@@ -327,8 +356,8 @@
           "output_index": 2
         }
       ],
-      "x": 720,
-      "y": 300,
+      "x": 560,
+      "y": 400,
       "width": 280,
       "height": 300,
       "data": {}
@@ -471,56 +500,40 @@
     }
   ],
   "edges": [
-    {
-      "id": "e_desc",
-      "from_node_id": "ref_description",
-      "from_port_id": "out",
-      "to_node_id": "op_compose",
-      "to_port_id": "in_0",
-      "type": "text"
-    },
-    {
-      "id": "e_dur_c",
-      "from_node_id": "ref_duration",
-      "from_port_id": "out",
-      "to_node_id": "op_compose",
-      "to_port_id": "in_1",
-      "type": "number"
-    },
     {
       "id": "e_lyr",
-      "from_node_id": "op_compose",
-      "from_port_id": "out_0",
+      "from_node_id": "ref_lyrics",
+      "from_port_id": "out",
       "to_node_id": "op_generate",
       "to_port_id": "in_0",
       "type": "text"
     },
     {
       "id": "e_gm",
-      "from_node_id": "op_compose",
-      "from_port_id": "out_1",
+      "from_node_id": "ref_global",
+      "from_port_id": "out",
       "to_node_id": "op_generate",
       "to_port_id": "in_1",
       "type": "text"
     },
     {
       "id": "e_vd",
-      "from_node_id": "op_compose",
-      "from_port_id": "out_2",
+      "from_node_id": "ref_vocals",
+      "from_port_id": "out",
       "to_node_id": "op_generate",
       "to_port_id": "in_2",
       "type": "text"
     },
     {
       "id": "e_arr",
-      "from_node_id": "op_compose",
-      "from_port_id": "out_3",
+      "from_node_id": "ref_arrangement",
+      "from_port_id": "out",
       "to_node_id": "op_generate",
       "to_port_id": "in_3",
       "type": "text"
     },
     {
-      "id": "e_dur_g",
+      "id": "e_dur",
       "from_node_id": "ref_duration",
       "from_port_id": "out",
       "to_node_id": "op_generate",