Restore the LoRA fits, and correct what this README claims
K. S. Ernest (iFire) Lee
e2aa8c0
Three claims did not hold. lora/ has four views on disk, not eight. Not
every directory carries the JSON its run wrote -- bench/ never did, and
lora/ stopped when four images were dropped and both JSONs went with
them. The slopes were rounded to 0.04 and 0.10 against 0.042 and 0.099.
The JSONs are restored from 4931850. A render is regenerated by rerunning
the generator; a measurement is not, and deleting the fit deleted the
only record of what the LoRA arm scored.
The four dropped views carried three of that arm slowest four errors --
155.9, 163.2 and 83.0 degrees -- so a reader who took the surviving four
images for the arm would have read a better result than it earned. That
is the failure the paragraph about keeping failures already warns about,
applied to itself.
The az180/225/270/315 A-view renders and the azimuth_recovery /
ladder JSONs go alongside the earlier prune. Content is regeneratable
from the Live2D renders + the generator's recorded checkpoint per
CLAUDE.md condition 1.
OmniGen2 outputs, manifested apart from the renders
