Full-song context
Non-causal self-attention models the complete acoustic latent sequence rather than isolated local windows.
Music generation · Audio demo
Beyond Reconstruction:
Full-Context Generative DiT for Music Generation
FullDiT treats acoustic rendering as full-context generation from an imperfect discrete plan, combining frame-aligned RVQ conditions, independently encoded lyrics and captions, and non-causal full-song context.
+0.77ViSQOL under synthetic corruption
69.7%non-tied preference over no EMDC
15 / 18automatic metrics led
Top 3public vocals leaderboard
Method
Rendering beyond reconstruction
Error-Matched Distractor Conditioning (EMDC) exposes the renderer to realistic codec-token mistakes without changing the acoustic target. Four-way classifier-free guidance independently controls codec, lyric, and caption conditioning at inference time.
Error-matched training
Codebook-specific replacement rates and cosine-KNN distractors mimic plausible language-model errors.
Independent guidance
4-CFG separates codec, lyric, and caption guidance increments for controllable rendering.
Full-length samples
Complete music generations
Listen to the full tracks. Audio loads only after you press play.
01Complete vocal song ·
I Am Free
Prompt. Alternative metal driven by thick, distorted guitars, thunderous drums, gritty male vocals, and an anthemic narrative about breaking free and reclaiming one's identity.
02Complete vocal song ·
Still Waiting
Prompt. An intimate midnight vocal jazz track exploring the theme of lost love through smoky vocals and a sparse acoustic arrangement.
03Complete vocal song ·
New Year Fireworks
Prompt. An upbeat reggae-pop New Year's Eve celebration driven by acoustic guitar upstrokes, a syncopated electric bassline, punchy drums, bright brass, and a cheerful male lead vocal, with an energetic and uplifting mood.
04Instrumental ·
Changing Pace
Prompt. A mellow contemporary R&B instrumental built from syncopated electric bass, laid-back drums, rhythmic electric-piano chords, warm synth pads, and seamless shifts between standard and double-time feels, ending in a slow fade-out.
Controlled comparison
What changes the rendering?
Prompt, lyrics, upstream LM codec tokens, duration, random seed, sampler, and guidance tuple
(1, 2, 1) are fixed. Only the FullDiT configuration changes.
Shared prompt
A dreamy Folk Pop track evoking the feeling of falling asleep under the stars, featuring a soft male vocal and gentle acoustic instrumentation.
- Duration
- 3:17
- Vocal
- English · male
- Tempo
- 75 BPM
- Key
- G major
M1Complete
FullDiT
Full-song context, renderer-side caption and lyrics, and EMDC.
M2aAblation
Local context
Local 30-second training context; otherwise matched to M1.
M2bAblation
No renderer text
No renderer-side caption or lyric conditions; otherwise matched to M1.
M3Ablation
No EMDC
No Error-Matched Distractor Conditioning; otherwise matched to M1.
Paired reconstruction
Suno V5.5 original vs. FullDiT
Each pair has matching duration. Switch between the two players for a direct comparison.
Pair 01
Instrumental
AReference
Suno V5.5 original
BReconstruction
FullDiT reconstruction
Pair 02
Vocal
AReference
Suno V5.5 original
BReconstruction
FullDiT reconstruction
Results
External evaluation
Automatic comparisons across leading music-generation systems on a shared multilingual set of vocal music.
| Evaluator | Metric | Ours | Suno 5.5 | Suno 5 | Mureka 8 | Lyria 3 Pro | MiniMax 2.6 |
|---|---|---|---|---|---|---|---|
| SongBench | Melody | 7.2001 | 6.6939 | 6.9202 | 7.0660 | 7.0377 | 6.7245 |
| Arrangement | 7.3879 | 6.8406 | 7.0805 | 7.2601 | 7.1511 | 6.8167 | |
| Musicality | 6.3678 | 5.8297 | 6.0096 | 6.2559 | 6.1810 | 5.8823 | |
| Vocal | 7.6234 | 7.0981 | 7.3160 | 7.6248 | 7.4560 | 7.2513 | |
| Instrument | 7.3517 | 7.0080 | 7.1487 | 7.2585 | 7.0583 | 6.8458 | |
| Mixing | 7.3047 | 7.0332 | 7.1101 | 7.1433 | 7.0093 | 6.7397 | |
| Structure | 6.9650 | 6.6158 | 6.7363 | 7.0237 | 7.0565 | 6.5637 | |
| SongEval | Coherence | 4.5094 | 4.2905 | 4.3278 | 4.3870 | 4.4673 | 4.2772 |
| Musicality | 4.4098 | 4.1547 | 4.1970 | 4.2875 | 4.3466 | 4.1781 | |
| Memorability | 4.4663 | 4.2123 | 4.2473 | 4.3680 | 4.4153 | 4.2180 | |
| Clarity | 4.3838 | 4.1246 | 4.1741 | 4.2431 | 4.3099 | 4.1448 | |
| Naturalness | 4.2600 | 4.0046 | 4.0679 | 4.1767 | 4.2021 | 3.9919 | |
| Audiobox | Content Enjoyment | 7.7089 | 7.5289 | 7.5507 | 7.5921 | 7.6950 | 7.6844 |
| Content Usefulness | 7.9989 | 7.9535 | 7.8020 | 7.8634 | 7.8919 | 7.9578 | |
| Prod. Complexity | 6.8774 | 6.6486 | 6.8118 | 6.8768 | 6.7561 | 6.7514 | |
| Production Quality | 8.2986 | 8.2083 | 8.1411 | 8.1098 | 8.2780 | 8.2923 | |
| CMI-RM | Alignment | 2.1786 | 1.9216 | 1.9119 | 2.3089 | 2.0377 | 2.1546 |
| Musicality | 2.7611 | 2.3665 | 2.3756 | 2.7590 | 2.5280 | 2.6661 |