Free version: clean.
Paid version: Loraholic’s sliders.

“merged DesireX BF16 weights”

A .safetensors file starts with a short JSON header in front of the weights, and tools can write free text into its __metadata__ field. Below are the three checkpoint headers as they are stored in the files.

{
  "__metadata__": {
    "format": "pt",
    "source": "BSS_DesireX2_BF16.safetensors",
    "provenance": "Direct INT8-convrot export from merged DesireX BF16 weights."
  }
}

{
  "__metadata__": {
    "format": "pt",
    "source": "DesireX2_Amateur_00001_.safetensors",
    "provenance": "Direct INT8-convrot export from merged DesireX BF16 weights."
  }
}

{
  "__metadata__": {
    "format": "pt",
    "source": "BSS_DesireX2_Amateur.safetensors",
    "provenance": "Direct INT8-convrot export from the named source checkpoint."
  }
}

• X2 and Amateur (Hard): provenance = "Direct INT8-convrot export from merged DesireX BF16 weights."
• Amateur (Soft): provenance = "Direct INT8-convrot export from the named source checkpoint."
• The source field names the file each export was made from: Amateur (Hard) from DesireX2_Amateur_00001_.safetensors; Amateur (Soft) from BSS_DesireX2_Amateur.safetensors; X2_INT8_ConvRot (Fastest) from BSS_DesireX2_BF16.safetensors.

Subtract, project, count signs

A checkpoint is the whole model, so a slider merged into it can’t be compared file to file. The scan works on the difference between two checkpoints instead.

Two dots out of 41

Every one of Loraholic’s 41 sliders, scanned against both paid versions. Across: the share of layers with the same sign. Up: the slider’s strength, regardless of sign. A merged slider lands in the top right corner.

None of the other 39 candidates goes above |strength| 0.35 or reaches 82 % consistency in either scan.

The same sign, layer after layer

Each square is one layer of Amateur (Hard) minus X2: 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 left half is Innie / Labia Minora Slider, the right half Wetness Slider. Blue: the layer pushes the same way as the whole file. Red: the opposite way. Hover or tap a square for its numbers.

Below zero, almost everywhere

One bar per layer, hanging down from zero to that layer’s strength, for Amateur (Hard) minus X2. The dashed line is the strength for the file as a whole. Both sliders are applied at negative strength, i.e. their negative direction is baked in. The page does not interpret what that direction looks like.

Paid minus paid: nothing. Free minus official: nothing.

The same scan, run on two pairs where the sliders should not show up. Each bar below sits next to the paid-version value for comparison.

The versions with the sliders are the paid ones

Both Amateur versions are on Civitai's paid early access (until 14 Oct and 15 Oct 2026), and the model description says the Amateur version is also sold by subscription to the creator's private channel.

The model page as a whole: 5,711 downloads and 440 likes at analysis (11 Oct 2026).

63 days after the Labia slider

Loraholic’s Wetness Slider went up on 29 Jun 2026, 14:17 UTC and Innie / Labia Minora Slider on 27 Jul 2026, 06:58 UTC. The free X2 followed on 19 Sep 2026, 11:00 UTC, Amateur (Hard) on 29 Sep 2026, 00:02 UTC and Amateur (Soft) on 30 Sep 2026, 22:27 UTC. Paid early access runs until 14 Oct 2026 and 15 Oct 2026.

63 days after Loraholic uploaded Labia, it was in a paid version of [BSS] - DesireX - [Krea2].

Two specific sliders, one sign, 256 layers

For the technical reader

• Method: Delta = paid Amateur checkpoint minus the free X2 checkpoint from the same model page, layer by layer (256 layers). Delta is projected onto each of Loraholic's 41 comparable Krea 2 sliders. Strength = how much of the slider the delta carries; consistency = share of layers whose projection has the same sign as the total.
• Checkpoint scan: W_checkpoint - W_base per layer after dequantizing both (int8 ConvRot, comfy_kitchen). The projection of that delta onto each candidate LoRA's B·A direction is accumulated over all layers; a joint least-squares solve over all candidates gives the 'joint' strength, which splits credit between correlated sliders.
• Consistency: A slider that is really baked in pushes every layer the same way, so its per-layer projections share one sign. Noise and unrelated training land near 50-80%.
• File header: Read a safetensors header. The command is below.
• Tool: Scans run with LoRA Forensics' engine/scan.py (K:/TOOLS/lora-forensics); the four scan result files in data/ are its unmodified output.
• Verdicts: the scan calls a candidate detected at |naive strength| ≥ 0.15 and ≥ 90 % sign consistency, marginal at ≥ 0.08 and ≥ 78 %.
• Per-layer cosine: the scan files also list a cosine per layer, measured against the whole difference between the two checkpoints. Most of that difference is other changes, so the cosine is small; this page does not use it as a similarity measure.
• Files verified: SHA-256 of the three analysed checkpoint files computed locally; all three match the hashes Civitai publishes for those versions.

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