Sharing a script: cuts Krea2's LoRA down to 10% of its original size.

The Forge / 工坊
1

The credit for this method goes to an author on Civitai who came up with the idea of “slimming down” the Krea2 LoRA. I just used Opus 4.8 to whip up a script so everyone can run it locally with one click. This is only for Krea2 LoRAs, don’t try it on others. The idea is that Krea2 LoRA weights are split into two main parts: the DIT/UNET transformer blocks take up most of the size, and the other part is the text conditioning fusion layer.

The logic is that for many style/atmosphere/character LoRAs, the base model already “knows how to draw” the relevant content—what the LoRA really does is tweak the text conditioning path.

So if you strip away the DIT/UNET and keep only the text conditioning layer, you can cut the size by over 90%, often with very little loss in fidelity. From my own library: a 218MB one dropped to 13MB, a 1.46GB LoKr dropped to 40MB, consistently landing in the 90-94% range.

Gotta be clear though—this isn’t free lunch. It works great for style LoRAs, but not so much for characters or anything that needs fine detail. I’ve put the code and full instructions on GitHub, open source and free.

2

90% reduction sounds way too crazy, bookmarking for later.

3

Just keeping the text conditioning layer is a clever approach.

4

The character LoRA doesn’t work here, that’s a key point.

5

1.46G down to 40M, my hard drive is saved.

6

The base model already knows how to draw it, LoRA just steers it in the right direction—makes sense.

7

Let me grab a few style LoRAs and test the fidelity.

8

Open source is awesome, gonna clone it later.

9

LMAO

10

I tried the style LoRA, and honestly, the fidelity is basically lossless—you can barely tell the difference. But don’t get your hopes up for character ones; you’ve gotta keep that prerequisite in mind.

11

40M? I didn’t believe that number at first, but after the script finished and I checked the hard drive, it really does save that much space.