While the study of colour patterns is a traditional subject of evolutionary ecology, there are various hypotheses which suffer from a lack of experimental evidence. One intriguing possibility is a trade-off between warning efficiency and detectability. After a certain size threshold, the detrimental effect of increased detectability can outweigh the benefits of warning colouration. One may thus expect corresponding patterns at the level of ontogenetic development: as juveniles grow, they should first acquire warning colouration, and then lose it again. We analysed this possibility in Orgyia antiqua, a moth species with hairy larvae which are polyphenic with respect to the intensity of warning colouration. We detected a regular change in colour patterns through larval life. Indeed, the larvae tend to display warning colouration at intermediate sizes while dull colours dominate in fully grown larvae. In aviary experiments, we confirmed that the colourful phenotype is the one that causes the strongest aversion in birds. Nevertheless, the effect was rather weak and most of the larvae were still eventually consumed when found. Unexpectedly, for human subjects, the warningly coloured larvae were harder, and not easier to find among natural vegetation, most likely due to the disruptive effect of the aposematic colour pattern. Importantly, the trend was reversed in the largest size class, suggesting that the disruptive colouration loses its advantage as the larva grows. This is consistent with the actual patterns of size-dependence of colouration. We present evidence against an alternative explanation which relates size-related change in colouration to behavioural changes prior to pupation. We conclude that even if the efficiency of the warning effect plays a role in determining the size-dependence of colouration, the pattern may be largely explained by the effects of size-dependent detectability alone.