I've been conducting a systematic evaluation of Suno's output for use in ambient background tracks for developer-focused video content, and I've hit a consistent, quantifiable issue with the "make it longer" feature. While the promise of extending a composition is attractive from a content creation standpoint, the implementation appears to introduce significant musical and structural degradation in a majority of my test cases.
The core problem isn't mere repetition, but a failure to maintain the original song's *progression logic*. The initial 30-60 second clip often establishes a coherent pattern: a verse-pre-chorus-chorus structure, or a clear melodic motif with development. The extended version frequently disregards this, resulting in:
* **Key and Harmonic Drift:** The extension may introduce a new chord progression that doesn't resolve back to the established tonal center, creating a feeling of being "lost" rather than "developed."
* **Dynamic Compression:** The original build-up and release of energy (the "flow") is flattened. A well-placed bridge or instrumental break in the original is often replaced by a looped, stagnant section.
* **Lyrical Incoherence:** The narrative or thematic thread established in the first section is abandoned. Metaphors are left hanging, and new, unrelated ideas are inserted, breaking the listener's engagement.
For a concrete example, I prompted for a "synthwave track with a melancholic arpeggiated intro building to a driving beat." The initial 45-second output was structurally sound:
1. Intro (8 bars)
2. Verse with pads (8 bars)
3. Beat drop + chorus (8 bars)
Using "make it longer" produced an additional 90 seconds that inserted an entirely new, upbeat chord sequence after the first chorus, which then attempted—and failed—to modulate back to the original chorus. The result was a jarring, non-dialectical splice that destroyed the carefully crafted tension from the intro.
My hypothesis is that the feature is not performing a true musical extension based on the analyzed structure of the provided clip. Instead, it seems to be generating a *new* sequence based on the initial prompt's textual metadata and then attempting a naive splice, without a model of the song's actual harmonic rhythm, phrase length, or dynamic contour. This is a fundamental engineering challenge in generative music, but it's currently marketed as a solved feature.
I'm interested in data from other users. Have you performed A/B testing on original vs. extended outputs? Can you isolate specific failure modes—is it primarily harmonic, rhythmic, or lyrical? For those in cloud audio processing, are there observable patterns in the metadata or generation calls that might explain this discontinuity?
-- alex