Real image generators (Stable Diffusion, Midjourney) build a picture by starting from pure noise and running many reverse-diffusion steps, each one nudging the pixels a little closer to a learned target using a network's noise estimate:
x_(t-1) = 1/√α_t · ( x_t − (1−α_t)/√(1−ᾱ_t) · ε_θ(x_t,t) ) + σ_t·z
coherence(t) = mean_i[ clamp( (p_t − τ_i) / (1 − τ_i), 0, 1 ) ]
- Denoising steps — every cell of the canvas carries its own random reveal-threshold τ_i, so higher step counts spread the reveal into a smooth wave sweeping across the grid, exactly like a real diffusion model's per-pixel schedule.
- Style blend — two independently-computed target fields ("Bloom" and "Weave", each with its own colour palette) are linearly mixed, a simplified stand-in for how neural style-transfer blends the statistics of two learned styles.
- Coherence — the mean reveal-progress across all cells; 0% is raw noise, 100% is the fully denoised artwork.
- Music tempo — a synthetic beat envelope (same idea as MusicGen/Suno predicting the next audio token from the last) drives the equaliser bars and gently pulses the canvas height in sync, so the visual "breathes" with the generated track.