Skip to content

Latest commit

 

History

History
executable file
·
89 lines (53 loc) · 7.55 KB

update_alert.md

File metadata and controls

executable file
·
89 lines (53 loc) · 7.55 KB

✨ Major Update Alert: Text-Based Style Prompts & Wild Instrumental Generation — Unleash Your Musical Madness!

Hey DiffRhythm community,

Last week’s release of DiffRhythm hid two secret features—and you brilliant creators found them all! Thanks to sharp-eyed explorers in our community (shoutout to @Jourdelune for his epic code contributions!), we’re thrilled to officially announce: Text-to-Style prompts and Instrumental Music Generation are now live in the open-source code!

🎨 Craft Styles with Words

Describe your vision literally — from quick tags to cinematic scenes. DiffRhythm speaks your language:

Straightforward prompts:

(Vocal glitches acknowledged → Can't resist sharing this gorgeous instrumental track!)

etc.

Vivid storytelling:

  • Indie folk ballad, coming-of-age themes, acoustic guitar picking with harmonica interludes 📀 Demo Track

  • Buenos Aires tango-rap graffiti artist's rhymes over bandoneón syncopation 📀 Demo Track

No reference audio needed—type your imagination, and let Diffrhythm handle the rest.

🎻 Instrumental Mode: Lyrics Optional, Chaos Mandatory

Skip the lyrics and go wild with absurdly specific prompts:

  • Mountain cabin fireplace, acoustic guitar folk tunes wrapped in wool blankets 📀 Demo Track
  • Arctic research station, theremin auroras dancing with geomagnetic storms 📀 Demo Track

or merge them with wild prompts – it’s your playground：

  • Zombie apocalypse country-rock gas station escape, banjo shreds & shotgun reload beats 📀 Demo Track
  • AI love rival duet: human vs vocal assistant jealousy battle, vocoder vs raw screams (⚠️ Audio Safety Notice: This demo may redefine "loud".🎧 Lower volume → 😌 | Keep volume → 🤯 Your choice, brave listener!) 📀 Demo Track

Rules? We deleted them. Throw your weirdest ideas at DiffRhythm and see what sticks.

🌟 The Power of Open-Source Alchemy

This update began as a community-powered quest:

  • Developer @Jourdelune discovered our hidden API endpoints while tinkering with the code
  • They not only implemented the features but submitted a clean, elegant Pull Request
  • Now their code lives in the main repo—this is open-source magic at its finest!

🚀 Get Started Now

We dare you to: 👉 Share your “this-shouldn’t-work-but-it-does” creations in Community 👉 Join the #DiffRhythm channel on [Discord] to witness collective genius/insanity 👉 Star the repo or become the next code wizard!

🔧 Community Feedback Drives Progress

A massive shoutout to @corybrsn and everyone who shared sharp insights on current limitations. Here’s our deep dive into your questions — transparency is key!

  • Musicality Improvements 🎵 "Sometimes the musicality feels uncanny or off-key." Our Response: We agree! This stems from limitations in harmony/melody modeling. Next steps: Training a larger model with higher-quality, genre-diverse data. Think "less robotic, more Radiohead" vibes.
  • Lyric Consistency 📝 "Words get skipped or repeated occasionally." Root Cause: The timestamp constraint is a double-edged sword. If the time between lines is too short → words get squeezed; too long → repetition kicks in. Fixes in Progress: We’re rebalancing timestamp flexibility (less rigidity!) and testing adaptive duration prediction.
  • Structured Lyrics 🕒 "Can we label [verse]/[chorus] sections?" Plan: Absolutely! Inspired by tools like Suno, we’re adding lyric structure tags (e.g., [intro], [bridge]) for granular control. Bonus: This will simplify timestamp input — describe sections, not milliseconds!

🤖 Why This Matters

Your critiques aren’t just "issues" — they’re blueprints for evolution. We’re tackling these challenges head-on because:

  1. Better musicality = More professional-grade outputs
  2. Smarter timestamp handling = Less engineering, more creating
  3. Structured lyrics = Democratizing music production

To All Sound Rebels: You’re why DiffRhythm evolved from a “tool” to a movement. When we see tracks like Quantum Hotpot Folk or Sad Washing Machine Glitchcore—we know you’re rewriting the rules of music itself.

Next-gen features are brewing (spoilers: song editing? vocal cloning?), but here’s the real ask: Keep bringing the chaos. We’ll keep turning it into sound.

To Every Tester & Code Contributor: You’re not just users — you’re co-developers. When you dissect the model’s quirks, we gain superpowers to rebuild it. Keep the feedback flowing!

— The ASLP-lab DiffRhythm Team & All Open-Source Alchemists

P.S. To everyone who commented “this project will be abused”… You were right, and we’re proud. 🚀