HY-WU
Contributor · 2026
A functional neural memory framework that generates instance-conditioned LoRA adapters on the fly, instantiated for text-guided image editing.
- Parameter generation
- LoRA
- Image editing
- MoE
Personalising a large generative model usually means fine-tuning it, which costs time per instance and risks eroding the general capability you started with. HY-WU takes a different route: it synthesises the adapter weights themselves.
Given an image and an instruction, the framework generates a low-rank update conditioned on that specific request and injects it into a frozen backbone during the forward pass. No test-time optimization, no per-instance training run. The base model stays untouched, so its general capability survives while the request gets an operator built for it.
The part I find most interesting is that it stays practical at scale. Through structured parameter tokenization it works against an 80B-parameter backbone, with an 8B parameter-generation model on top. Methods that only work on small models are easy; the constraint that matters is whether the approach still holds when the backbone is the size people actually deploy.
Released in March 2026 with inference code and weights. On human preference evaluation it holds up against strong closed-source systems despite the smaller backbone, which is the result the README leads with.