Exemplary improvements for in-band wavelet video coding with spatial scalability are described. In one implementation an encoder uses leaky prediction to balance a trade-off between reducing drifting errors at lower resolutions while maintaining coding performance at higher resolutions. In an alternative implementation, a trade-off technique defines two coding modes at the macroblock level. The first mode includes predictions employed at a low quality reference while the second mode includes predictions employed at a high quality reference. The method then adaptively selects the proper coding mode for each macroblock according to a rate-distortion criterion.