Update README.md

README.md

@@ -27,67 +27,97 @@ Prompts were selected to intentionally stress known diffusion and caching weakne
 
 # Test Prompts & Rationale
 
-1️⃣ Rhythmic "Jitter" Test
+The following prompts were intentionally designed to stress different structural and spectral weaknesses under TeaCache acceleration.
 
-Prompt:
+---
+
+## 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)
 
-    "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 handle variable-length audio. Identity coefficients may fail to capture transient timing precisely when skipping steps, potentially causing rhythmic jitter or "beat drift".
-    Evaluation Criteria:
+**Prompt:**
 
-    BPM accuracy
+> "Glass shattering on a concrete floor, high-pitched shards, long reverb tail."
 
-    Timing jitter
+**Why:**
 
-    Peak consistency
+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."
 
-    High-frequency crispness
+**Evaluation Criteria:**
 
-2️⃣ High-Frequency Clarity (Transient Test)
+- Transient detection count  
+- Average peak width (sharpness)  
+- Spectral high-frequency ratio  
 
-Prompt:
+---
 
-    "Glass shattering on a concrete floor, high-pitched shards, long reverb tail."
-    Why:
-    The model targets high-fidelity 44.1kHz output with energy in high frequencies. Identity coefficients may over-smooth these peaky regions, turning sharp "shards" into generic white-noise "whooshes."
-    Evaluation Criteria:
+## 3️⃣ Ghosting & Artifact (Drone Test)
 
-    Transient detection count
+**Prompt:**
 
-    Peak width (sharpness)
+> "Ambient drone, slow evolving cinematic pads, deep bass, mystical atmosphere."
 
-    Spectral high-frequency ratio
+**Why:**
 
-3️⃣ Ghosting & Artifact (Drone Test)
+TeaCache reuses transformer residuals from previous timesteps to achieve ~1.5×–2.0× speedup.  
+Slow-moving, low-variation signals may suffer from:
 
-Prompt:
+- Residual ghosting  
+- Rhythmic wobble  
+- Amplitude pumping  
 
-    "Ambient drone, slow evolving cinematic pads, deep bass, mystical atmosphere."
-    Why:
-    TeaCache reuses transformer residuals from previous steps to provide a 1.5x-2.0x speedup. Slow-moving signals may suffer from "ghosting" or rhythmic "wobbles" if the cache is updated too aggressively.
-    Evaluation Criteria:
+if the cache update threshold is too aggressive.
 
-    Amplitude variance (Pulsing)
+**Evaluation Criteria:**
 
-    Spectral flatness
+- Amplitude variance (pulsing detection)  
+- Spectral flatness  
+- High-frequency noise ratio  
+
+---
 
-    High-frequency noise
+## 4️⃣ Complex Layering Test
 
-4️⃣ Complex Layering Test
+**Prompt:**
 
-Prompt:
+> "A man speaking in a crowded cafe with clinking plates and background jazz."
 
-    "A man speaking in a crowded cafe with clinking plates and background jazz."
-    Why:
-    Stable Audio Open specifically finds it challenging to generate prompts with "connectors" and may omit sounds or merge layers. This tests if caching exacerbates the "muddying" of foreground speech against background noise.
-    Evaluation Criteria:
+**Why:**
 
-    Stereo width
+Stable Audio Open is known to struggle with prompts involving connectors (multiple simultaneous sound sources).  
+This test evaluates whether TeaCache acceleration exacerbates:
 
-    Speech presence band (300Hz–3kHz)
+- Foreground/background blending  
+- Stereo image collapse  
+- Layer loss  
+
+**Evaluation Criteria:**
+
+- Stereo width (correlation)  
+- Speech presence band (300Hz–3kHz energy)  
+- Dynamic variance (layer separation clarity)  
+
+---
 
-    Dynamic variance
 
 ---