What does cwb stand for?

#2
by Jamerrone - opened

I am not sure which version to get.

https://github.com/silveroxides/Spec-CWB

CWB stands for Consensus-Weighted Blending, a technical merging method developed by silveroxides.

Unlike standard linear weighted-averaging when merging model checkpoints, LoRAs, or conditioning vectors, CWB treats the merge as a “cooperative vote” among the source models rather than a blind mathematical average. Standard averaging in high-dimensional spaces causes “Centroid Collapse” — the merged result loses distinct texture, color depth, and detail because vectors get diluted toward the origin as more sources are blended. CWB solves this using three techniques: a consensus baseline (mean or median) computed across aligned vectors, similarity-based soft-masking weights that boost well-aligned features while suppressing outliers, and norm rescaling to restore the original energy/magnitude of the merged vector so colors and detail don’t wash out.

 kroma-v0.2-base-to-turbo-cwb.safetensors  was built using this consensus-weighted merge algorithm to combine the base Kroma weights into the Turbo backbone, while  kroma-v0.2-base-to-turbo.safetensors  was likely produced via a simpler/standard linear merge instead. Since CWB specifically targets preserving texture detail, color saturation, and avoiding “muddy” or generic-looking merges, the  -cwb  version should theoretically retain sharper, more faithful characteristics from the base Kroma fine-tune compared to a standard linear-averaged merge.

Sign up or log in to comment