Abstract: To address the problems of spatial discontinuity, blurred edges, and loss of structural details in seismic data reconstruction, a new algorithm is proposed based on multi-scale feature fusion and generative adversarial network (MSF-GAN). The algorithm designs a multi-scale feature fusion generator in the GAN for effective seismic feature extraction and multi-scale fusion. A feature splicing module is designed in the generator for adaptively adding masks to seismic data, so as to splice the reconstructed data and the original intact data and improve computational efficiency. A multi-dimensional adversarial discriminator is designed in the GAN to improve reconstruction accuracy. Furthermore, a hybrid loss function integrating Smooth L₁ reconstruction loss and adversarial loss is proposed to update the generator and improve reconstruction reliability. Public data and seismic data from Daqing oilfield are reconstructed to validate the algorithm in different scenarios: continuous data loss, random data loss, and regular data loss. MSF-GAN performs better than orthogonal matching pursuit, projection onto convex sets, and spectrally normalized generative adversarial network in structural details and spatial continuity.