Skip to content

Add RMS renormalization option (fixes magnitude-collapse quality loss)#2

Open
huwhitememes wants to merge 1 commit into
nova452:mainfrom
huwhitememes:fix/quality-preserving-renormalize
Open

Add RMS renormalization option (fixes magnitude-collapse quality loss)#2
huwhitememes wants to merge 1 commit into
nova452:mainfrom
huwhitememes:fix/quality-preserving-renormalize

Conversation

@huwhitememes

@huwhitememes huwhitememes commented Jun 25, 2026

Copy link
Copy Markdown

@/tmp/pr2_body.md

Adds a non-breaking `renormalize` option that rescales the per-layer-weighted
tensor to its input RMS, so tap ratios can be rebalanced without inflating the
overall conditioning magnitude. Off by default; existing behavior is unchanged.

Background: the global multiplier amplifies the whole tensor, so with the default
weights + multiplier=4.0 the magnitude inflates ~8.7x (4x mult x ~2.2x from the
per-layer gains) -- matching the community-reported quality collapse (likeness,
prompt adherence, colour). With renormalize on, the multiplier becomes the sole
magnitude control.

Co-Authored-By: Claude <noreply@anthropic.com>
@huwhitememes huwhitememes force-pushed the fix/quality-preserving-renormalize branch from 0c51575 to 9b5c71d Compare June 25, 2026 07:06
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant