i2L predicts LoRA weights in one forward pass, no per-style training

"Diffusion-based style transfer currently faces a trade-off: adapter methods are fast but often copy too much content, while optimization-based methods like LoRA are better at internalizing style but require a separate training loop for each new reference. i2L (image-to-LoRA) breaks this deadlock by amortizing LoRA training into a single forward pass. Given one or more reference images, an image encoder feeds into learnable LoRA queries, which then pass through compressed decoding heads to generate the adapted weight matrices. Training on semantically diverse style pairs teaches the model to preserve appearance cues while suppressing content copying from the reference. The approach works with popular models like Z-Image, FLUX.2, and Hidream-O1, delivering higher style fidelity and prompt alignment than existing baselines.

Because i2L produces explicit LoRA weights rather than conditioning vectors, it unlocks several practical advantages. Users can apply asymmetric classifier-free guidance to better separate style from content, fuse multiple reference styles into a single set of weights, and compose style transfer with other controllable-generation modules. This makes i2L not just faster, but more flexible than prior work. The paper demonstrates strong results across both artistic and photorealistic styles, suggesting that amortized style personalization could become a standard component in next-generation image generation pipelines. For practitioners, this means no more per-style fine-tuning cycles: upload a reference, run one forward pass, and generate images in that style immediately."