Abstract: We trained a generative diffusion model (DM) to simulate ultra-relativistic heavy-ion collisions end-to-end. The model takes initial entropy density profiles as input and produces final particle spectra, successfully reproducing integrated and differential observables like charged multiplicity, mean transverse momentum $\langle p_T\rangle$, anisotropy flows $v_n$, and $v_n(p_T)$ for $n=2,3,4$. It also captures higher-order fluctuations and correlations among observables, including $n$-particle ($n=2,3$) $p_T$ correlators, $n$-particle ($n=2,4,6,8$) flow cumulants, $n$-order ($n=2,3$) event-plane correlations, and correlations between mean transverse momentum $\langle p_T\rangle$ and flow $v_n$ ($n=2,3,4$). These findings suggest that the generative model has successfully learned the complex map from initial conditions to final particle spectra across various shear viscosities, as well as the fluctuations introduced during initial entropy production and hadronization stages, providing an efficient framework for resource-intensive physical goals.