DMD LoRa trained by Lodestones which has been put through svd rank pruning at 90% to remove overcooking amd allow better compatibility to other lora compared to original.

A DMD LoRa is a destillation of a checkpoint that allows it to output similiar results at lower steps.

This original DMD lora has been created for Kroma V30 by Lodestones , a Krea 2 finetune on the Chroma dataset of 5 million training images selected out of 20 million total. The original is overcooked requiring a strength at 0.65.

You can find the original DMD LoRa here: https://tensor.art/models/1056882394004495813

The SVD rank prune at 90% accuracy reduces the 900Mb lora to 600Mb in size.

This SVD approximation , which I have created , is an alternative to the original lora for allowing better compatibility with other lora , while retaining the destillation effects .

The approximation helps against overcooking , allowing the svd90% lora to be used at normal strength without artifacts found in the original.

The DMD lora creates comparable results at 5 steps generation to the base Kroma V30 model 's 15 steps.