Accurate modeling of non-linear gravitational dynamics is essential for constraining extensions to the standard cosmological model using large-scale structure observations. While high-resolution $N$-body simulations provide the required fidelity, they are computationally prohibitive for the large ensembles needed to analyze Modified Gravity (MG) scenarios. We present , a field-level emulator based on a convolutional neural network that upgrades fast, approximate simulations to near–$N$-body accuracy at a fraction of the computational cost. Trained on a suite of simulations for $f(R)$ gravity, achieves nearly sub-percent accuracy ($\lesssim 1\%$) in both the matter power spectrum and bispectrum up to $k\simeq 1h{Mpc}^{-1}$ on its training domain. Crucially, although being trained on a fixed cosmology with variations restricted to the modified gravity sector, the network generalizes robustly to cosmologies outside its training manifold keeping the error below $10\%$. Furthermore, within the standard gravity regime, it successfully recovers the General Relativity limit $\Lambda$CDM without introducing spurious MG signals and accurately captures the power suppression induced by massive neutrinos ($M_{\nu }\le 0.4$ eV), despite being trained on cosmologies with massless neutrinos. The pipeline delivers a speed-up factor of $\sim 31\times$ relative to full $N$-body runs, generating a high-fidelity realization in $\sim 16.3$ CPU-hours compared to $\sim 500$ CPU-hours for the baseline. This accuracy–efficiency trade-off establishes as a powerful tool for generating the massive mock catalogs required for next-generation galaxy surveys.