Civitai's image CDN is a Cloudflare edge cache: the rendition redirect is
stored for 48 h and the blob for 1 h, so the same URL set is about 2.3x
faster on a second visit (measured 2.82 vs 6.6 req/s at concurrency 6).
The harness reused one URL set across modes and repeats, so the first mode
through it paid every cold miss while later modes looked faster.
- Resolve each tree's exact URLs with that tree's own builder (in a
subprocess, so the old and new modules cannot alias), warm them before
timing, then probe a sample and record the edge hit ratio per run.
- Print a per-run table and report cache_clean_runs; warn instead of
silently mixing cold runs into the medians.
- Pin the rendition width in the sandbox settings so the warm-up and the
server agree even when the user's own settings differ.
- Skip models outside folder_paths roots: their sidecars resolve alongside
the model file, so the old tree cannot read them and silently processed
them as failed. --any-root opts back in.
- Add reverify: populate the sandbox examples tree and time a library-wide
Re-process All (--force) or Download Missing (--no-force) pass in
ms/model, with --global to mirror the UI's global entry.
- Default the old worktree under --workdir.
Will Miao
d70232bb9d