گاهی، نگاهی

Over the past year, I’ve been working at the intersection of a peculiar creative challenge: composing contemporary Persian poetry specifically for Western operatic delivery, collaborating with AI to transform these works into full musical productions. The result was “Migrant Cloud (Operatic Explorations)”: an eight-track album released globally on Spotify, Apple Music, YouTube, and other streaming platforms. This represents the first full-length operatic work in Persian created through creative collaboration with AI.

The technical demands were extreme: Composing and delivering sustained bel canto legato phrasing, controlled vibrato across full dynamic range (pianissimo to fortissimo), bass-baritone and tenor tessitura, cello-led orchestration with saxophone and piano, dramatic narrative arcs across 6-8 minute compositions, and perhaps most challenging: accurate phonetic rendering of Persian text through AI models trained predominantly on English and European languages.

But here’s the paradox that prompted this post: the AI music generation platform that enabled this work has recently released a “better” model that has made my entire workflow, and the creative achievements it produced, completely impossible to replicate. The creative capabilities of new and ‘better’ models are all gone!

This isn’t a story about bugs or temporary setbacks. It’s a technical case study of a well-documented phenomenon in Generative AI development: mode collapse, where optimization for quality metrics and performance benchmarks, systematically destroys creative diversity of Generative AI, this applies the same way to OpenAI as it does to SUNO and other creativeAI. And it raises urgent questions about how we evaluate “progress” in generative AI systems designed for creative work.

The Three Generations: A Systematic Trade-off

Working with leading AI music generation platforms over ten months across three model generations revealed a consistent pattern. (Version numbers below are generic references to aid clarity, not vendor-specific designations. All platforms I tested across multiple providers, demonstrated identical evolutionary trajectories.)

V1 (Early 2025): Wild creativity, poor execution. The model generated genuinely experimental compositions with novel melodic structures and unexpected instrumental combinations, but the sound quality was recognizably synthetic and unnatural. Instruments sounded artificial, voices lacked realism, production was thin and unconvincingly shallow. Songs could not be extended beyond 4 minutes. Creative range was extraordinary while the technical quality was amateur.

V2 (Mid-2025): The productive sweet spot. Creative range remained largely intact while execution improved dramatically. The model could still explore diverse interpretations of the same prompt, generating meaningfully different renderings across iterations. Hallucinatory singing decreased (but plagued increasingly complex productions), instruments approached realism, metatag (inline AI nudges) responsiveness increased. Critically, the model maintained workable reproducibility across iterations: if you specified “[Bass-baritone, pp, fluid phrasing],” the model would make a genuine attempt to deliver a close interpretation of that, even if it required dozens of iterations get there.

This was the generation of AI models that enabled “Migrant Cloud” as an album. Every track on that album passed through hundreds of iterations, but iteration was productive, each generation explored the prompt space differently, allowing artistic discovery and refinement, and I could assert creative control over each track and the entire album.

V3 (Late 2025): Superb sound quality, creative bankruptcy. The latest model produces technically pristine output: crisp instruments, realistic voices, professional-grade production. But it has collapsed to a narrow creative valley. Variations of diverse prompt strategies generates virtually identical outputs across iterations. Dynamic range has compressed. Genre diversity has narrowed. Metatag responsiveness has vanished. Musical innovation has been replaced by a competent mediocrity. the highest quality lame/boring/cliche music one could imagine: technically polished but creatively inert.

Specific Technical Regressions in Latest AI Music Generator Models

1. Dynamic Range Collapse:

A genuine natural human performance, especially Opera, requires voices that can traverse extreme dynamic ranges, from whispered intimacy (pianissimo, ppp) to powerful climaxes (fortissimo, fff), all within a single vocal instrument. This is fundamental to dramatic narrative: the arc from vulnerability to triumph, despair to transcendence.

V2 handled this routinely (proof is in the pudding: Migrant Cloud). When I prompted for “bass-baritone delivery spanning ppp to fff,” it generated a single male voice dynamically traversing the full range. Admittedly, this required multiple iterations to achieve the desired arc, but ultimately produced coherent results.

V3 cannot do this at all. Instead, when prompted for a dynamic range, the model generates duets: a female voice handles the soft passages, a tenor takes over for the loud sections. The system has learned that soft singing correlates with higher-pitched voices and loud singing correlates with lower-pitched voices, but it cannot produce a single vocal instrument capable of both ends of the spectrum.

This fundamentally breaks operatic delivery, which depends on a single voice’s transformation across dynamic states.

2. Singing Reduced to Recitation

More fundamentally, V3 has lost the ability to sustain melodic singing across varied dynamics. Vocal delivery has degraded to predominantly recitational singing. The voice frequently falls behind instrumental tracks, forcing the music to slow down and change tempo mid-stream to accommodate the recitation rather than maintaining sustained melodic phrasing.

In “Mirage (tracks 1 and 4 of the album” when I prompted for operatic legato with pianissimo decrescendo on a Persian phrase translating to “where are you, O mirror of hope”, V3 repeatedly abandoned singing altogether and simply spoke the words over orchestral accompaniment.

V2 handled this beautifully, and regularly, producing genuine vocal performance with controlled dynamics, sustained vibrato, and seamless legato phrasing. V3 produces monotone recitation with erratic tempo adjustments to mask the absence of genuine vocal performance.

This isn’t just an opera problem. Genuine vocal delivery across all genres requires dynamic range with sustained melodic phrasing. V3 cannot deliver this in any style.

3. Genre Capability Narrowing

Classical, opera, and orchestral styles have degraded severely in V3 models. The model excels at mainstream pop and simple rock arrangements but struggles with repertoire requiring instrumental complexity, dynamic variation, and structural sophistication.

For “Lighthouse Within Me, tracks 3 and 7” V2 produced atmospheric psychedelic rock fusion with Operatic delivery and neo-classical chamber elements: cello-heavy orchestration building to orchestral crescendos, elaborate saxophone solos, piano arpeggios creating ambient textures, all developing across a dramatic emotional arc (questioning → yearning → discovery → understanding) throughout the track. impossible to imagine it with V3 models.

V3 defaults to generic pop instrumentation regardless of style prompts. No amount of metatag specification can steer it toward the orchestral complexity V2 delivered routinely.

4. Metatag Unresponsiveness

V2 responded reliably to detailed compositional directives:

• Section-level mood tags: [Verse – Despair, Intimate]
• Dynamic markings: [pp, fluid phrasing]
• Vocal delivery: [Bass-baritone, powerful delivery]
• Instrumental cues: [Cello lines, piano, ambient textures]

While V2 required many iterations to achieve desirable musical interpretation as instructed by the style prompts and metatag nudges, it was achievable and reproducible. As a result, I developed specific phonetic notation systems (see example below) and reliable workarounds that consistently shaped the model’s rendering, enabling an album with cohesive artistic direction across all eight tracks.

[Instrumental Break 2 - Contemplative] (15 seconds)
[mp, Dark distorted Cello lines with orchestral piano harmonies, ambient saxophone, deep reverb]

[Verse 4 - Passionate]
[staccato delivery, mp→ff crescendo]
سَجْدِه بَر‌ْ سَخْرِیِه سَخْت‌ْ.
قَترِ‌ه بَر‌ْ داٰامَنِ سَنگْ.
اینْ بوسِ‌یِه gärrmm

[Refrain 2 - Triumphant, mp →f]
سَحْمِ زییباٰائی‌یِه نَستَرَن‌َستْ.

[Instrumental Bridge 2 - Soaring] (25 seconds)
[Extended saxophone solo, layered cello lines, ambient piano]

[Final Verse - Urgency, Desperation, f→ff crescendo]
حاٰنْ، ey اَبْرِ أَجُولْ
Tôô بیاٰ وُ دَریینْ دَشتْ بِماٰنْ/
Tôô بِماٰانُ‌ بَریین‌ْ دَشتْ بِ‌باٰارْ.

V3 frequently ignores style prompts and metatag nudges entirely, defaulting to generic delivery regardless of instructions. What required systematic workflows and clever nudges in V2 to reproduce and control the creative process, is now utterly impossible in V3, even for a single track.

5. Loss of Musical Innovation

V2 demonstrated genuine creative exploration by generative/creative AI models. Identical style prompts and lyric compositions produced meaningfully different renderings across iterations, exploring novel regions of musical and vocal delivery. This variability enabled artistic discovery, testing interpretations, refining directions, finding unexpected solutions that elevated the work beyond initial conception. A truly collaborative, co-production experience.

V3 stagnates on default interpretations, no matter the prompt. Diverse prompt strategies generate virtually identical interpretations across iterations: the same clichéd instrumentation, the same monotone vocal delivery, the same generic arrangements. Variation between generations is negligible, eliminating any justification for experimenting with styles or metatags.

The model has collapsed to a single mode of output that cannot be steered, refined, or evolved through iteration.

The Technical Mechanism: Mode Collapse in RLHF

The pattern I’ve observed aligns precisely with a well-documented phenomenon in generative AI research: mode collapse resulting from reinforcement learning (RL) from human feedback (RLHF) or AI feedback systems (AI Judges).

Here’s what appears to be happening:

The Optimization Problem: Modern AI music generators use RLHF to align outputs with human preferences. Users rate generated music, and these ratings train a reward model that guides future generation. The system learns to maximize expected reward.

The Trade-off: Standard RLHF approaches optimize for sample quality (how good is this particular output?) rather than distributional diversity (how much of the possible output space can the model explore?). Critically, there’s not merely an absence of reward for novelty, exploration actively gets penalized.

When models venture away from the center of the distribution where training data is densest and user preferences most concentrated, they receive lower rewards even if the output is technically competent. The reward signal is strongest for outputs that conform to the most frequently liked or repeated patterns. This creates a vicious cycle: models learn that safety (staying near the distribution center) consistently outperforms innovation (exploring underrepresented regions), so they progressively narrow their output range with each training iteration. The mode collapse we observe is not a bug, it’s the system working exactly as optimized, converging toward maximum expected reward by eliminating creative risk.

Furthermore, each sample is evaluated in isolation, typically as bite-sized fragments (2-5 second clips) randomly extracted from compositions and compared against gold-standard examples on narrow spectral characteristics. The reward model asks: “Does this fragment’s frequency spectrum match high-quality reference audio?” It does not ask: “Does this fragment serve the compositional intent specified in the user’s prompt?”

This fragment-based evaluation fundamentally cannot capture what makes music work as music: structural relationships, harmonic development, motivic recurrence, and dramatic pacing that operate at timescales far beyond individual samples. These are precisely the qualities that style prompts and metatag nudges target for creative control, the very dimensions the reward system ignores. Consequently, a model can score maximum reward on every 3-second fragment while producing an 8-minute composition that completely disregards prompt specifications and metatag directives. No reward signal exists for global coherence or conformance to compositional intent, only local spectral quality. The system optimizes brilliantly for the wrong objective.

The Mode Collapse: The optimizer converges toward outputs that reliably score well with the reward model, which means it converges toward the median preference of the median user. In music generation, this means English-language pop/rock content with conventional instrumentation, moderate dynamics, and familiar structures. Extreme dynamic ranges, complex orchestrations, non-mainstream genres, and structural sophistication all represent regions of the distribution that median users don’t request and don’t reward.

Classifier-Free Guidance: Diffusion-based architectures use classifier-free guidance to improve prompt adherence, but this technique explicitly trades “mode coverage” for “sample fidelity.” You get outputs that match prompts more closely (higher fidelity) but with severely reduced diversity (lower coverage). The model becomes better at producing what it already does well and worse at exploring new territory.

Why V3 sounds better but creates worse music: The V3 systems have learned to produce pristine versions of the most commonly requested content. For first-time users generating simple pop songs, V3 delivers excellent results immediately. But for specialized applications, professional opera, classical music, complex orchestrations, extended dynamic ranges, underrepresented languages, these models have lost their competence altogether. The trade-off only becomes visible when you examine the full corpus of possible outputs, not individual samples.

Operatic creativity sits at the worst intersection: It demands underserved genres (opera, classical), extreme performance characteristics (ppp to fff dynamics, extended legato phrasing), non-mainstream instrumentation (cello-led orchestration with saxophone), structural complexity (6-8 minute dramatic arcs), and non-english languages (Italian, Persian). Every dimension of this work exists precisely where mode collapse hurts most.

Why This Matters Beyond My Use Case

I’m sharing this not just as complaint but as technical documentation of a critical problem in generative AI development for creative applications.

The market for AI music generation is rapidly commoditizing. Multiple platforms (Suno, Udio, aisonggenerator.io, and others) are converging toward the same optimization target: easy-to-use tools producing high-quality generic content for median users. They’re racing toward the bottom, competing on price and convenience while offering increasingly indistinguishable capabilities.

But there’s a different strategic position available: Creative-AI solutions that can deliver on serious creative work while preserving innovation and quality. Platforms that can produce both competent pop songs and experimental opera. Systems that respond reliably to sophisticated compositional directives while maintaining creative exploration across iterations.

This requires different optimization objectives incorporated into RL with HITL approaches:

1. Multi-objective reward models that balance sample quality against distributional diversity, and prompt conformance
2. Innovation bonuses that explicitly reward exploration of underrepresented output regions
3. Metatag fidelity metrics that verify outputs actually match compositional specifications
4. Structural coherence assessment for long-form works requiring thematic development
5. Granular user feedback beyond binary like/dislike ratings, structured assessments of conformance to specific prompt elements

The technology to implement these approaches exists today. Sophisticated language models can perform automated spectral analysis of generated audio, parse style prompts to extract compositional specifications, and verify presence of requested musical characteristics. I personally used these technics to understand and characterize my own work. These could serve as specialized AI judges in RLHF, creating tight feedback loops between user intent and model optimization.

The Value Proposition

Professional composers, experimental artists, and serious creative practitioners are willing to invest hundreds of hours mastering platform capabilities, developing custom workflows, and pushing systems to their limits. When their work succeeds, they create high-profile projects that generate substantial brand value through explicit platform attribution and community evangelism.

“Migrant Cloud” exists because V2 offered capabilities no competitor could match. I explicitly credited aisonggenerator.io in every track’s production metadata precisely because the platform enabled work I couldn’t achieve elsewhere.

But when platforms abandon the capabilities that enable this work, when they optimize exclusively for first-time user experience at the expense of advanced functionality, they lose the users who provide the most valuable form of marketing: proof that the technology can support genuine creative achievement.

Conclusion: Progress Isn’t Always Forward

This experience demonstrates something important about evaluating “progress” in generative AI: Better-sounding outputs don’t mean a better system if the range of what can be expressed has collapsed.

For generative and creative tools, diversity of expression matters as much as quality of execution. A model that can explore 50 different interpretations of a prompt with 80% quality is more valuable for artistic work than a model that produces a single interpretation with 95% quality.

The challenge for AI music generation isn’t just making things sound natural, it’s preserving the creative possibility space that makes artistic discovery feasible.

I hope this technical post-mortem contributes to broader conversations about how we optimize generative models for creative applications. The solutions exist. The question is whether companies will implement them, or whether the race to serve median users will eliminate the capabilities that enable genuinely innovative work.

Parsa Mirhaji