LyCORIS (Lora beYond Conventional methods, Other Rank adaptation Implementations for Stable diffusion) is an open-source library implementing a family of parameter-efficient fine-tuning methods for diffusion and other models that extend beyond standard low-rank adaptation. It includes techniques such as LoHa (Hadamard-product decomposition), LoKr (Kronecker-product decomposition), and full or convolutional adaptations, giving practitioners a richer set of expressiveness-versus-size trade-offs. LyCORIS is widely used in the image-generation community to train compact, shareable model adapters.
Content
- LyCORIS arose from the image-generation community’s need to customise large diffusion models cheaply. Standard low-rank adaptation captures a weight update as the product of two low-rank matrices, which is effective but represents only one particular way of constraining an update to be small. LyCORIS treats this as one option among many, packaging a family of alternative low-parameter decompositions that offer different balances of expressiveness, file size, and training behaviour.
- Its signature methods generalise the matrix factorisation. LoHa decomposes the update using a Hadamard (element-wise) product of low-rank factors, which can represent higher effective rank than a plain low-rank product for the same parameter count. LoKr uses a Kronecker product, yielding very compact adapters well suited to capturing structured patterns. Additional modes adapt convolutional layers and offer fuller, less constrained updates when fidelity matters more than minimal size.
- This menu of methods matters because adaptation tasks differ. Teaching a model a simple style may need only a tiny adapter, while capturing a complex subject or intricate concept benefits from a more expressive decomposition. By exposing these choices in a single interoperable library compatible with common training and inference tooling, LyCORIS lets practitioners tune the trade-off between adapter quality and the size of the file they must store and share.
- The practical impact is a thriving ecosystem of small, swappable model adapters. Because a LyCORIS adapter is a fraction of the size of a full model, users can train, distribute, and combine many of them on top of a shared base checkpoint, mixing styles and concepts at inference time. This composability — many lightweight adaptations over one large foundation model — is central to how the open generative-image community customises and remixes models without the cost of full retraining.