I mean, they might have modified the midjourney code to patch out the error message. Presumably the programmers put the check in because results for negative we

I mean, they might have modified the midjourney code to patch out the error message. Presumably the programmers put the check in because results for negative weight sums become nonsensical, not because it's impossible. Thinking in terms of classifiers, an image is almost never categorized as exactly 0% something, instead it's a positive value. Negative weights would make the net optimize for the smallest possible percentage. For the birthday cake, the negative weight is not strong enough to favor any "anti-autumn-landscape" patterns, only to remove features associated with autumn landscapes. But for weights that are all negative, it's plausible that the system will produce all the features that, in the training data, are anticorrelated with the prompt.