DiffRhythm
A full song in seconds, fast enough as a sketch machine; watch the structure yourself.
Model profile
- Vendor
- ASLP Lab（西北工业大学）
- Form
- 本地模型
- Input
- 歌词 · 文本
- Output
- 音频
- Licence
- 代码以 Apache 2.0 放出，权重与后续版本（含 DiffRhythm+）的商用范围另行标注，以仓库与模型卡为准
- Verified
- 2026-10-31
Related
Ten seconds and five minutes
DiffRhythm compresses a whole song into one diffusion pass: lyrics plus a style prompt, and about ten seconds later a finished track of roughly four minutes with vocals and backing together. It does not chase line-by-line polish; speed and completeness are the product. That suits short-video beds, three directions for a client, or a quick style sweep. It will not give you precise diction or a finished mix. The official ceiling is four minutes forty-five seconds, and actual length depends on lyric length and version.
Two inputs, one finished track
Lyrics go in by line, with tags to separate sections; the style prompt names genre, mood and instrumentation, and specificity helps. Output is a mixed stereo file with no stems. The model makes its own decisions about lyric placement, compresses long lines and does not always honour section tags. To steer form, even out the line count per section first, then state the verse-chorus contrast in the prompt.
| What you write | What often comes back | What to check |
|---|---|---|
| Four verse lines, four chorus lines | Roughly right, chorus lifts | Whether the second chorus is copied |
| One style sentence | Close instrumentation and mood | Low end and drum density |
| Long lines, polysyllables | Diction flattened | Line endings and high notes |
Speed belongs early in the process
Ten seconds a song means dozens of tries in an afternoon. Split lyrics, style and structure into variables, hold the rest fixed, and a usable direction appears quickly. Later versions can continue or edit a section, but the boundary of the edit is unclear and one move can replace the whole arrangement. Treat it as a sketch machine, not a final master; mixing, repair and stems still belong in a DAW or another tool.
Repetition, dropped lines, lost structure
- The second chorus is copied exactly, down to the drum fill.
- Long lyrics lose lines, especially final particles.
- Section tags are ignored and the bridge disappears.
- Vocal and backing levels do not match, so a batch needs per-file normalisation.
- Mandarin tones blur on fast lines; English legato is steadier.
Local runs and alternatives
The base model runs in 8GB of VRAM and generates quickly, which suits repeated trials. Code and weights are licensed separately, and later versions and DiffRhythm+ carry different commercial terms; check the model card before release. Alternatives: YuE for longer, steadier full songs; Suno for a browser service; TME Tianqin or a similar platform for stems and post-production.
Listen in the library
All tracks from midicn-lib — inline badge is the license tier (C1 commercial / C2 non-commercial / C3 study only).
Along the same line
Sources
- DiffRhythm 官方仓库与论文（ASLP Lab）关于潜在扩散、整曲时长与输入格式的说明。
- DiffRhythm 模型卡上的权重许可与版本差异。
- 公开用户报告中的重复、漏句与结构问题，只取现象，不引用具体评测数据。