Improving coding efficiency for vector quantization (VQ), e.g., reducing bit-rate attainable and increasing encoding speed, has attracted intensive attention. A new image VQ scheme of subsampling-based predictive vector quantization (SB-PVQ) is introduced for this aim, which fully exploits the neighboring-pixel smoothness in both intra-blocks and the boundary areas of inter-blocks. Firstly, all non-overlapped image blocks partitioned in an image are down-sampled two-dimensionally periodically. The sub-sampled blocks (sub-vectors) are then vector quantized with a lowerdimensional and smaller size of codebook, which will contribute to speeding up the VQ encoding and reducing the bitrate. Finally, the original image blocks are constructed with decimated pixels predicted by their intra-or inter-block neighboring subsampled pixels. It should be highlighted that in the sub-block VQ encoding, multiple-candidates scheme is employed for each input in order to find the corresponding sub-codevector that can generate the best reconstructed image block. Compared with the peripheral prediction method, the new method achieves significant improvement in terms of rate-distortion performance while maintaining comparable computation complexity.