TMM, RLE, and quantile normalization aren't appropriate on data for which there are expected unidirectional global shifts. In such cases you either need a spike-in or prior knowledge about genes that aren't changing (these are then used with TMM/RLE/etc. to compute scaling factors).
scRNA-seq isn't any different in this regard, except if you're primarily doing it for finding cell-types you can hope there's more going on than simply transcriptional amplification and just ignore spike-ins. Spike-ins are probably more accurate in scRNA-seq, since it's more likely that you actually know how many cells you have.