Three uploads.
Zero training.

Each file names its own sources

A .safetensors file starts with a short JSON header in front of the weights, and tools can write free text into its __metadata__ field: the text below was written into Vince_AI’s files by the merge tool that built them, not by us.

{
  "__metadata__": {
    "source_a": "breast_size_v2_krea2_loraholic.safetensors",
    "source_b": "CrunchyBanana_Krea2_BreastSlider_V1.safetensors",
    "strength_a": "1.0",
    "strength_b": "1.0",
    "merge_type": "lora_plus_lora_to_lora",
    "rank_mode": "Exact",
    "output_dtype": "BF16",
    "vince_merge_lab": "1.3.3"
  }
}

{
  "__metadata__": {
    "source_a": "skindetails_krea2_loraholic.safetensors",
    "source_b": "real_3d_krea2_loraholic.safetensors",
    "strength_a": "1.0",
    "strength_b": "1.0",
    "merge_type": "lora_plus_lora_to_lora",
    "rank_mode": "Exact",
    "output_dtype": "BF16",
    "vince_merge_lab": "1.3.3"
  }
}

{
  "__metadata__": {
    "source_a": "brightness_krea2_loraholic.safetensors",
    "source_b": "light_krea2_loraholic.safetensors",
    "strength_a": "1.0",
    "strength_b": "1.0",
    "merge_type": "lora_plus_lora_to_lora",
    "rank_mode": "Exact",
    "output_dtype": "BF16",
    "vince_merge_lab": "1.3.3"
  }
}

python -c "import json,struct;f=open(PATH,'rb');n=struct.unpack('<Q',f.read(8))[0];print(json.loads(f.read(n))['__metadata__'])"

What “built from” 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 compare that matrix (B·A) in Vince_AI’s file with the same layer in each named source, and then ask how much of it the two sources explain together.

Every layer, accounted for

Each square is one layer of Vince_AI’s file: 32 rows for the blocks of the network (28 main blocks and 4 text-fusion blocks), 8 columns for the layers inside a block. The colour says whose file the layer came from. Hover or tap a square for its numbers.

The other two files

Same grid, same colours.

The header says 1.0. So does every layer.

For each layer we fit Vince_AI’s layer as c_a × source A + c_b × source B. If the layer had been trained, c would scatter like noise. Below, one bar per layer and one row per source: the bar’s height is c, in the owner’s colour. A flat wall at the dashed line means every layer carries the source at exactly the strength the header states.

Training never produces a constant. The header says strength_a = strength_b = 1.0, and that is exactly what we measure, layer after layer. In BustMorph Slider each layer takes 1.0 of one source and 0.0 of the other: a copy, not a blend.

From upload to re-upload in 7 h 52 min

Loraholic’s five sources went up between 24 Jun and 15 Jul 2026. CrunchyBanana’s slider went up on 9 Oct 2026, 21:37 UTC. Vince_AI’s three files followed between 9 Oct 2026, 21:33 UTC and 10 Oct 2026, 06:56 UTC.

7 h 52 min after CrunchyBanana uploaded [KREA2] Breast Slider, Vince_AI published BustMorph Slider with 24 of its layers inside, behind paid early access until 25 Oct 2026.

No. Here is the scale of “no”.

For the technical reader

• Cosine: Per layer, the delta matrix (B·A) of the target and of each source is compared by cosine similarity, computed on the low-rank factors without materialising the matrices. 1.0000 = identical direction.
• Norm ratio: ||target layer|| / ||source layer||. 1.0 on a copied layer means copied at full strength.
• Joint fit and R²: Per layer, least squares target ≈ Σ c_i · source_i over the sources named in the file header. R^2 is the share of the target layer's energy the sources explain. 1.000 = the layer is fully built from the sources; nothing was trained.
• Coefficients: The fitted c_i per layer. For the two sum-files every layer gives 1.0 and 1.0, matching strength_a = strength_b = 1.0 in the header. For BustMorph every layer gives (1.0, 0.0) or (0.0, 1.0): one source or the other.
• File header: The safetensors header (first 8 bytes = JSON length, then JSON) carries a __metadata__ object written by the merge tool. The command is below.
• Files verified: Sources downloaded from the creators' own Civitai pages, SHA-256 verified against the hashes Civitai publishes. Targets identified on Civitai by SHA-256 via the public API (model-versions/by-hash).
• Reproducible: evidence.py in this folder computes every per-layer number below from the public files (--target, --sources, --out); the tables are the complete dataset.

python -c "import json,struct;f=open(PATH,'rb');n=struct.unpack('<Q',f.read(8))[0];print(json.loads(f.read(n))['__metadata__'])"