Most mergers (kohya, mergekit) only combine adapters that share the same low-rank format. Public Krea-2 realism packs are a mix of classic LoRA and LoKR, so those tools stop there.
Badaramoni Fusion works in weight-delta space instead:
- Reconstruct each adapter’s true ΔW
LoRA:ΔW = B·A
LoKR:ΔW = kron(w1, w2) - Normalize key conventions into one canonical module space
- Blend:
ΔW = Σ wᵢ · ΔWᵢ - Re-factor the combined delta back to one rank-128 LoRA with truncated SVD
You get a single portable .safetensors — not a naive alpha-average, and not a multi-adapter stack at inference.
No trigger word. Write a normal descriptive sentence.
Settings for the samples
- base:
krea/Krea-2-Turbo - LoRA scale 1.0–1.3 (start at 1.3)
- 8 steps, guidance 0–1, μ 1.15
- short side 1024–1536
This is a many-to-one, lossy-by-design compression. The contribution is the fusion algorithm that makes mixed-format realism adapters usable as one file.
License follows the Krea 2 Community License. You need access to krea/Krea-2-Turbo to run it.