Rust port of LyCORIS (advanced LoRA-family algorithms for diffusion models) by KohakuBlueleaf, targeting flame-core for inference-time weight merging.
The actual crate is at lycoris-rs/ — see its README for full status, API, parity results, and usage.
- 4 algorithms shipped: LoCon, LoHa, LoKr, Full — Linear + Conv2d + Tucker variants
- Inference-only weight-merge mode (load → materialize ΔW → add into base → run inference)
- Wired into inference-flame as a path dep with per-model name mappers (FLUX, Z-Image, Chroma, Klein, Qwen-Image, SDXL, SD 1.5)
fuse_split_qkvhelper for fused-QKV models — validated against a real Z-Image LoKr (240 split-QKV adapters → 30 fused triples → 180 adapters)- 31 tests in the crate (lib + smoke + loader + parity vs upstream Python) plus 32 in the inference-flame integration
Original LyCORIS: KohakuBlueleaf/LyCORIS. This Rust port keeps the math; ships the loader + per-model integration around it.
MIT.