cardsorting/cats

CATS is a research project that has been officially retired and archived. The README states plainly that it should not be revived, redesigned, or extended, and that it is being kept only as a record of completed exploratory research. The project set out to answer a specific question: whether image generation models could benefit from something similar to a reusable memory system that language models already use to speed up their work. Two lines of research came out of this question. The first looked at whether an image generation model called Stable Diffusion 1.5 could be sped up by changing how it plans its generation steps partway through, using saved checkpoints and adjusted trajectories. Two controlled experiments, named E001 and E002, found that this approach did not perform better than already strong, standard fixed schedules when given the same amount of computing time. The second line of research tried to design a memory system that would track objects and their relationships across images for a different model called FLUX.2, but this line was also retired before completion. The project's overall conclusion is that trying to improve older image generation scheduling methods is unlikely to lead to a real breakthrough, and that future research in this area should instead focus directly on newer multimodal generation systems. The README points to four documents that record the project's findings in more depth: a post mortem explaining why the project stopped, a research brief summarizing the lasting insights, an archive index cataloging what was tried and what failed, and a longer whitepaper covering all of it together. Although the specific experiments are archived, a handful of general purpose Python utilities built along the way, covering things like saving state, tracking a computing budget, measuring quality, timing, and checking reproducibility, are kept active for use in future projects, and can be tested by running pytest.