Forensic weight analysis · Civitai · KREA-2 LoRA

Same numbers.
Different name.

On June 26, 2026, alcaitiff published “KREA-2 Breast Slider” on Civitai. Inside that file are the trained weights of Loraholic’s “Breast Size Slider — (Krea-2 + ZIT)”, copied layer for layer. This page walks through the evidence. Every number is reproducible.

0 / 104

layers in the published file are bit-for-bit directional copies of the original

of the file is copied content

4.000×

the exact same scale factor on every single copied layer — the fingerprint of a merge tool

Scroll — the evidence tells the story ↓

01 · The two files

Two uploads. One set of weights.

Both models were published on Civitai. One was trained; the other was assembled. The SHA-256 hashes below identify the exact files analyzed — anyone can download them and verify.

The original

Loraholic

Breast Size Slider — (Krea-2 + ZIT)

Uploaded: June 26, 2026
breast_size_krea2_loraholic.safetensors · 256 layers · 20.5 MB

SHA256 39D0FDF1F76B557C21EF599A84615DE04179A5272F2B6FF2A254DBA26E7669CD

View on Civitai →

The copy

alcaitiff

KREA-2 Breast Slider

Uploaded: June 26, 2026
BreastSlider-KREA2.safetensors · 104 layers · 9.6 MB

SHA256 77B789BA95E24CCD450C535CC395D0A2120CA1EC9264B75D4BFC376155694DD6

View on Civitai →

02 · How you measure a copy

What identical looks like in math

A LoRA layer is a matrix of millions of numbers pointing in a direction in a very high-dimensional space. For each layer we reconstruct that matrix (B·A) in both files and measure the cosine similarity between them — how precisely the two directions align.

~0.01

Independent training. Two people training on the same concept — even with the same images and the same recipe — land on directions that barely overlap.

0.99

Heavy derivation. Fine-tuning on top of someone’s model leaves similarity high but never perfect — training always moves the numbers.

1.0000

A copy. The same numbers, possibly multiplied by a constant. This does not occur by training. It occurs by copying.

03 · The evidence, layer by layer

Scanning all 104 layers of alcaitiff’s file

Keep scrolling to run the comparison ↓

■ exact copies: 0 □ swapped-in content: 0

80 exact copies. 24 swapped layers. Zero coincidences.

04 · The smoking gun

Eighty layers. One number.

Each copied layer carries the original weights multiplied by a scale factor. If these layers had been trained, the factors would scatter like noise — training never produces a constant. Here is the measured scale of every layer in the file:

● copied layer ○ swapped-in layer · x-axis: the 104 layers in file order · y-axis: measured scale vs the original

Every one of the 80 copied layers sits at exactly 4.000× — the signature of loading someone’s LoRA into a merge tool at strength 4 and hitting “save”. Training does not do this. Ever.

05 · Could this be a coincidence?

No. Here is the scale of “no”.

Independent training on the same conceptsimilarity ≈ 0.01

Observed in alcaitiff’s file — 80 separate layerssimilarity = 1.0000

Hitting similarity 1.0000 in a single layer by chance is comparable to guessing a 40-digit number on the first try. This file does it in 80 layers simultaneously — and every one of them carries the identical 4.000× scale factor.

There is exactly one process that produces this result: copying the file.

06 · The 24 layers that differ

What was swapped in — and what it proves

Blocks 8, 9 and 16 don’t match the original. That is not evidence of independent work — it is evidence of editing.

They don’t come from the original creator’s method either

The 24 deviating layers were compared against four other sliders by Loraholic, trained with the same technique. Similarity against all of them: ≈ 0.000. The material is foreign — substituted in from elsewhere.

Selective replacement is a choice, not an accident

An independent training run differs everywhere. This file matches the original perfectly in 10 blocks and differs completely in exactly 3 — the pattern of someone splicing content into a copied file, not of someone training a model.

07 · Verify it yourself

For the technical reader

Method: for every layer, the delta matrix (B·A) is reconstructed in both files and the cosine similarity between them is computed. 1.0 = identical direction (same weights up to a scale factor). “Scale” is the norm ratio — the strength at which the copy carries the original.
The original is verified: downloaded directly from Loraholic’s Civitai account, SHA-256 checked against the hash Civitai itself publishes.
Reproducible: the table below is the complete dataset. Anyone with the two public files can regenerate every number in it.

Show all 104 layer pairs

Layer	Cosine similarity	Scale	Verdict

Same numbers.Different name.