1ACE-Step 1.5best base
Text song, lyric-to-song, cover, repaint, style LoRA
- Access
- GitHub + Hugging Face
- License
- MIT
TJ Music Factory V7
Create like Mureka. Deliver like a workstation.
Prompt, style, engine, results, stems, voice, master, QA, and downloads now sit in one product flow. This is no longer a wall of text.
Creative promptCinematic mandopop film theme for a final scene, emotional strings, warm piano motif, modern trailer drums, subtle synth bass, soaring main melody.
Current pick: ACE-Step 1.5I will generate candidate takes from the selected style and engine, keeping cover art, duration, tags, downloads, and QA evidence.
StyleFilm Theme / 96 BPM / D minor / 128s
EngineUse as default local song engine target
Production Packtrailer cue / 3 candidates
Cinematic mandopop film theme for a final scene, emotional strings, warm piano motif, modern trailer drums, subtle synth bass, soaring main melody. Use case: movie trailer cue. Language: English primary. Style: movie trailer cue, cinematic, movie theme cue with emotional motif and widescreen final payoff, 96 bpm, D minor, instrumental, no vocals, pulse bed, riser, impact hits, hero theme, final slam, streaming-quality structure, memorable motif, professional mix depth, clear intro-to-payoff arc. Song form: Pulse intro -> Tension bed -> First impact -> Hero theme -> Drum escalation -> Final hit -> Tail. Target: 128 seconds, 96 BPM, D minor. Engine route: ACE-Step 1.5. Avoid: no toy loop, no 16-bar demo, no 32-bar demo, no nursery melody, no robotic vocal, no broken timing, no thin one-finger loop, no low bitrate artifact.
Stylemovie trailer cue, cinematic, movie theme cue with emotional motif and widescreen final payoff, 96 bpm, D minor, instrumental, no vocals, pulse bed, riser, impact hits, hero theme, final slam, streaming-quality structure, memorable motif, professional mix depth, clear intro-to-payoff arc
Avoidno toy loop, no 16-bar demo, no 32-bar demo, no nursery melody, no robotic vocal, no broken timing, no thin one-finger loop, no low bitrate artifact
Pulse introTension bedFirst impactHero themeDrum escalationFinal hitTail
duration is close to targetfile is not tiny or silenthas intro, lift, payoff, and clean endingdoes not sound like a 16/32-bar loop
Three.js live scene / mode: generate
Build songs, split stems, train voices.
A full AI music command desk with prompt-to-song, legal voice modeling, stem mixing, and export-ready production flow.
Click scene to switch modeCanvas labels are generated from HTML-like panels
Generated resultsV7
Generated song cards appear here instead of doing nothing.ACE-Step 1.5 / Film Theme
The Mureka pattern is prompt conversation, result cards, versions, tags, favorite actions, and continued editing. V7 now follows that structure.
01Prompt
02Engine
03Song
04Stems
05Voice
06Master
07QA
5 open engines
ACE-Step first, HeartMuLa and DiffRhythm 2 next, YuE for long-form research, Stable Audio 3 for BGM/SFX layers.
2HeartMuLacandidate
Lyrics plus tags to complete multilingual songs
- Access
- GitHub + weights
- License
- Apache-2.0
3DiffRhythm 2fast lane
Fast full-song generation with lyric alignment
- Access
- GitHub + weights
- License
- Apache-2.0
4YuEresearch
Long songs, lyrics-to-song, multilingual vocal style
- Access
- GitHub + weights
- License
- Apache-2.0 research review
5Stable Audio 3asset layer
Instrumental music, BGM, sound effects, inpainting, LoRA
- Access
- GitHub + open weights
- License
- Community license review
Generated results
No song yet. Run the full flow or generate a song.
Quality gate
Song filePendingmissing
MP3/WAV downloadPendingblocked until real audio
Stem countPending0/5
Voice versionPendingmissing
Master renderPendingblocked until real provider master
No local fake songPassprovider gated
Local mode can render downloadable audio, stem demos, and QA evidence. Natural Longge singing still needs a real singing model, dry vocal data, and listening approval.
Split stems
No stems yet. Split stems first.
Voice boundary
TTS reading, voice cloning, singing conversion, and a trained singer are different products. V7 does not sell reading as singing.
No voice version yet. Create a voice version.
Benchmark learning
MurekaPrompt chat, versioned track cards, style tags, library, studio, pricing.
ACE-Step 1.5Default local full-song target: metadata control, editing, stems, LoRA.
HeartMuLaLyrics and tags as first-class controls for multilingual song generation.
DiffRhythm 2Fast full-song lane with lyric alignment and coherent structure.
YuELong-form lyrics-to-song and multilingual vocal-style research.
Stable Audio 3BGM, SFX, continuation, inpainting, and LoRA asset layer.
Local substitutes
MusicGen nowPromptBrain -> MusicGen HF/local wrapper -> postprocessed real waveform; no chord/MIDI fallback.
Open model targetACE-Step is the first adapter target; HeartMuLa and DiffRhythm 2 are next wrapper candidates.
Longge assets nowGPT-SoVITS weights, clean Longge refs, and emotion refs are detected locally.
Truth boundaryDo not claim perfect Longge singing until singing model output and listening QA are proven.
Pricing
0Free first run1 song, 1 stem split, preview export
29Creator monthly20 songs, MP3, basic stems
99Single-song packWAV, stems, voice conversion, QA report
Evidence
0songs
0audio assets
0voices
0jobs
0exports
5priority engines