We've talked about this problem 2 months ago.
The performance degradation was obvious for the specific scenario, where the same costume data was slightly modified and reused.
But there are no "low hanging fruits" for a generic case.
The performance penalty is caused by the implicit conversion of a string to a number by JS engine. So if you take into account both manipulating pixel list and setting costume there will be no substantial difference. Actually, the list of char version is 2 x faster because calling a custom reporter causes extra delay.
Filling the entire costume with color is faster this way, with a precalculated pixel list. But it still 500 ms for 500x600 costume (or ~100ms for 240x360).
So maybe some APL construct based on reshaping can help with this.
Or building an array exponentially by
