A new arXiv paper, Pristine Models, addresses a central tradeoff in image generation. According to the abstract, training on broad, unfiltered data produces high-quality outputs, yet it can also teach models concepts that should not be generated. Restricting the training data is presented as a way to ensure safety, but that restriction comes with a quality penalty because examples are lost.
The abstract frames this as a direct tension: safety and quality pull in opposite directions. The phrase "because examples…" is where the provided excerpt ends, so the paper's proposed resolution is not available from the announcement alone.
As described, the work is a useful reminder that data filtering is not a free lunch. The tradeoff it names is familiar to practitioners, but the full contribution—and whether the paper offers a way to ease it—remains to be seen from the full text.