Akshat/TeaCache-flux-identity-StableAudioOpen

Identity vs TangoFlux Coefficients with TeaCache Acceleration

Stable Audio Open 1.0 Evaluation (Tesla T4)

---

Abstract

This experiment evaluates Identity vs TangoFlux coefficients under TeaCache acceleration for Stable Audio Open 1.0.

Initial benchmarks showed that Identity provided better raw speedup.
However, to investigate quality trade-offs, we conducted structured stress tests across rhythmic, spectral, temporal, and multi-layer scenarios.

Prompts were selected to intentionally stress known diffusion and caching weaknesses.

Testing is done via feeding the output to gemini3(thinking) and then asking it to rate the generation!

---

Test Prompts & Rationale

The following prompts were intentionally designed to stress different structural and spectral weaknesses under TeaCache acceleration.

---

1️⃣ Rhythmic "Jitter" Test

Prompt:

"140 BPM sharp techno drum loop, heavy kick, crisp hi-hats, 44.1kHz."

Why:

Stable Audio Open uses a timing-conditioned latent diffusion model to generate variable-length audio.
When TeaCache skips updates, Identity coefficients may fail to preserve transient alignment precisely, potentially introducing rhythmic jitter or subtle beat drift.

Evaluation Criteria:

• BPM accuracy
• Timing jitter (ms)
• Peak consistency (kick uniformity)
• High-frequency crispness

---

2️⃣ High-Frequency Clarity (Transient Test)

Prompt:

"Glass shattering on a concrete floor, high-pitched shards, long reverb tail."

Why:

The model targets high-fidelity 44.1kHz stereo synthesis, where high-frequency energy is critical.
Identity coefficients may over-smooth sharp spectral peaks, turning crisp impacts into broadband noise-like "whooshes."

Evaluation Criteria:

• Transient detection count
• Average peak width (sharpness)
• Spectral high-frequency ratio

---

3️⃣ Ghosting & Artifact (Drone Test)

Prompt:

"Ambient drone, slow evolving cinematic pads, deep bass, mystical atmosphere."

Why:

TeaCache reuses transformer residuals from previous timesteps to achieve ~1.5×–2.0× speedup.
Slow-moving, low-variation signals may suffer from:

• Residual ghosting
• Rhythmic wobble
• Amplitude pumping

if the cache update threshold is too aggressive.

Evaluation Criteria:

• Amplitude variance (pulsing detection)
• Spectral flatness
• High-frequency noise ratio

---

4️⃣ Complex Layering Test

Prompt:

"A man speaking in a crowded cafe with clinking plates and background jazz."

Why:

Stable Audio Open is known to struggle with prompts involving connectors (multiple simultaneous sound sources).
This test evaluates whether TeaCache acceleration exacerbates:

• Foreground/background blending
• Stereo image collapse
• Layer loss

Evaluation Criteria:

• Stereo width (correlation)
• Speech presence band (300Hz–3kHz energy)
• Dynamic variance (layer separation clarity)

---

---

Hardware & Configuration

• GPU: Tesla T4 16GB
• Total Steps: 50
• Max Audio Length: 10 seconds
• Precision: float16
• TeaCache r1_threshold values tested:
  • 0.2 (Stable / Conservative)
  • 0.6 (Aggressive / Stress Test)

---

Results

---

Rhythm Test Results

File	BPM	Timing Jitter	Peak Consistency ↓	HF Crispness	Rating
rhythm_02_identity	140.0	21.2 ms	1066	62.8%	7/10
rhythm_02_tangoflux	140.5	23.1 ms	266	37.4%	9/10
rhythm_06_identity	140.1	6.9 ms	1164	53.1%	6/10
rhythm_06_tangoflux	143.7	40.1 ms	3663	66.1%	3/10

Analysis

Threshold 0.2

• Identity (7/10)
  On-grid tempo but high peak variance → inconsistent kick dynamics.

• TangoFlux (9/10)
  4× better peak consistency → uniform professional thump.
  Slight BPM drift but stronger rhythmic lock.

Threshold 0.6 (Stress)

• Identity (6/10)
  Stable timing but reduced crispness → muffled sound.

• TangoFlux (3/10)
  Severe instability: extra beats, 40ms jitter, massive variance.
  Audible rhythmic breakdown.

Rhythm Summary

TangoFlux excels at low thresholds.
Identity is more robust under aggressive caching.

---

Transient Test Results

File	Transients	Avg Peak Width ↓	HF Ratio	Rating
transient_02_identity	7	0.100 ms	37.5%	5/10
transient_02_tangoflux	3	0.076 ms	42.9%	8/10
transient_06_identity	3	0.060 ms	42.8%	7/10
transient_06_tangoflux	4	0.057 ms	53.4%	9/10

Analysis

Threshold 0.2

• Identity: More peaks but blurred.
• TangoFlux: Sharper and brighter impacts.

Threshold 0.6

• Identity: Clean but simplified.
• TangoFlux: Best sharpness and HF preservation.

Transient Summary

TangoFlux clearly superior for high-frequency bite and sharp impacts.

---

Layering Test Results

File	Stereo Width ↓	Speech Presence	Dynamic Variance	Rating
layering_02_identity	0.78	71.9%	0.081	8/10
layering_02_tangoflux	0.69	51.3%	0.040	6/10
layering_06_identity	0.54	65.2%	0.060	7/10
layering_06_tangoflux	1.00 (Mono)	40.1%	0.031	1/10

Analysis

Threshold 0.2

• Identity preserves voice clarity and depth.
• TangoFlux compresses layers.

Threshold 0.6

• Identity fails gracefully.
• TangoFlux collapses to mono and loses subject separation.

Layering Summary

Identity is significantly more reliable for complex environments.

---

🌊 Drone / Temporal Test Results

File	Amplitude Variance ↓	Spectral Flatness	HF Noise	Rating
drone_02_identity	0.038	0.081	14.2%	9/10
drone_02_tangoflux	0.045	0.076	13.5%	8/10
drone_06_identity	0.052	0.104	16.1%	7/10
drone_06_tangoflux	0.071	0.092	22.4%	4/10

Analysis

Threshold 0.2

Identity produces the most stable amplitude profile.

Threshold 0.6

TangoFlux introduces noticeable wobble and HF artifacts.

Drone Summary

Identity is superior for atmospheric and evolving pads.

---

Final Verdict (Tesla T4 Context)

Use Case	Recommended Coefficient
Sharp transients (glass, snares, impacts)	TangoFlux (r1=0.2)
Drum loops at conservative thresholds	TangoFlux (r1=0.2)
Multi-layered environments	Identity
Drones & atmospheric audio	Identity
Aggressive caching (r1=0.6+)	Identity (more stable)

---

Overall Rating: 8.5 / 10

The experiment demonstrates a clear trade-off:

• TangoFlux preserves spectral sharpness and transient attack.
• Identity provides better stability and graceful degradation under high caching thresholds.

On a Tesla T4 (limited compute), Identity offers a superior speed-to-quality stability balance for complex audio scenes.

---

Coefficient Comparison Summary

Feature	Identity	TangoFlux
Transient Response	Blurs sharp sounds	Preserves HF detail
Rhythmic Stability	Stable but dynamic variance	Excellent at low threshold, unstable high
Layer Separation	Strong foreground/background separation	Risk of muddiness or mono collapse
Temporal Flow	Minimal pulsing	Prone to wobble artifacts

---

Conclusion

TangoFlux is best for high-frequency, sharp, single-subject sounds.

Identity is more reliable for:

• Multi-layered environments
• Atmospheric content
• High-threshold TeaCache usage
• T4-constrained deployments

This experiment highlights how coefficient selection directly impacts structural vs spectral fidelity in accelerated diffusion audio pipelines.