We present a copula-based blind source separation (BSS) method in which the usual kernel estimate of c(U, V) is replaced by a compact implicit neural density model. The network is trained with a KL-oriented f-GAN loss and stabilized through exponential moving averages and short warm-start updates at each iteration. The substitution leaves the marginal transforms and derivative terms unchanged, but produces far more stable gradient evaluations. On standard image mixtures, the proposed variant reaches roughly 40–42 dB on the first source and about 30–32 dB on the second, whereas the kernel estimator levels off near 20 dB. The improvement is systematic, with smooth and steadily rising SNR curves and a marked reduction of residual interference in the separated images. Overall, the neural copula surrogate brings a substantial gain in numerical robustness at very low computational cost, making it a practical drop-in replacement for kernel-based copula BSS.