Abstract: To address the aliasing problem caused by insufficient sampling rates in real-time rendering, as well as the limitations of traditional anti-aliasing algorithms in handling shading aliasing and temporal instability, an SSIM-guided anti-aliasing neural network called SGNN is proposed. The proposed network adopts an en-coder-decoder architecture: a decoupler extracts multi-scale image features, separating SSIM prediction in-formation from image reconstruction information; an SSIM prediction module estimates the aliasing severi-ty of each pixel region, and the prediction results serve as prior information to guide the kernel prediction module in generating spatially adaptive convolution kernels for high-quality neighborhood anti-aliasing; a temporal accumulator based on motion vectors and depth constraints, combined with a feature reweighting mechanism, computes fusion weights to blend historical and current frames, effectively suppressing tem-poral instability such as ghosting and flickering. Experimental results on six self-constructed test scenes demonstrate that, compared with three competing algorithms including DLSS, MCNN, and TAA, the pro-posed network achieves an average PSNR of 32.483 dB, surpassing the second-best algorithm DLSS by 1.192 dB; an average SSIM of 0.952, exceeding DLSS by 0.017; and an average IF-SSIM of 0.994 8, outper-forming DLSS at 0.992 6, demonstrating significant advantages in both anti-aliasing quality and temporal stability.