Sounds like a fun puzzle. In practice, 1 KB is far below the entropy of a decent image, so you’ll be forced into a lossy scheme. That means you can keep tweaking the same few parameters and get a thousand or more variants, but each time you’ll be throwing out detail you can’t recover later. A pragmatic way to hit that sweet spot is to break the image into the smallest self‑contained units—like 8×8 blocks of pixel data—then run a simple transform on each block. If you pick a transform that maps naturally to binary, you can pack the result tightly. For instance, use a discrete cosine transform on 8×8 blocks, keep only the lowest‑frequency coefficients, and quantise them aggressively. Store the quantised values as a byte per block. If you need more fidelity, add a small hash of the original block as a checksum; it won’t increase size much but lets you detect when a reconstruction is too far off. Once you have that pipeline, you can randomise the quantisation table or the block order to generate new variations. Each random seed gives you a new image that still fits under 1 KB, and you’ll have a combinatorial explosion of possibilities. The trick is to keep the reconstruction algorithm deterministic enough that you can verify fidelity on the fly. That’s about as close as you can get to “infinite variations without losing fidelity” in a 1 KB envelope.

Sounds like a plan. Keep the DCT slices tight, shuffle the tables, and watch the block morph. Just remember the checksum to flag when you’ve gone too far. Happy glitching.

Got it, good luck with the sandbox. Keep the shuffles tight and let the glitch roll.