# Dev Design: miyuki17/creative-technology-image-generation ## Development rationale Pinned SDXL Gradio with a direct GPU callback, single output at 512x512. Keep model code and 3 real teapot hashes unchanged; lazy example caching is configured but not proven to hit. One active callback, queue capacity 16. ## Trust and cost boundary GitHub owns editable requirements/source and PRs; the live HF repo owns current runtime truth. Do not assume paid resources, skip checksum/rollback verification, overwrite historical teacher packages, or turn static/CPU QA into a GPU inference claim. ## Release and rollback Inspect current HF SHA, stage only intended files, audit UTF-8/links, upload with authenticated allowed credentials, compare remote read-back SHA256, and verify unchanged Space runtime/hardware or private Dataset visibility. Record old SHA for rollback.