A new fix lets style tweaks built for slow AI video generators survive the jump to fast, compressed versions.
LoRA adapters are small add-on files that change how a video-generation model looks or moves, without retraining the whole system. Many video models get "distilled" down from dozens of denoising steps to as few as four, so they run faster and cheaper. That compression often breaks LoRA adapters trained on the original, slower model - the file loads fine, but the effect it's supposed to produce weakens or vanishes. Researchers built a method called DART that recalibrates an adapter's behavior for the shorter schedule, without retraining it or needing the original training videos. On a four-step version of Alibaba's Wan2.2 video model, DART pushed a joint quality score from 0.9029 to 0.9227 and turned a measure of whether the adapter's intended effect actually survived from negative (-0.4644) to positive (+0.1349).
This is the gap between "compatible on paper" and "works in practice," and it's a real cost for anyone building on distilled video models. A studio's library of trained styles can quietly go stale the moment a faster base model ships, with no error message to flag it. DART's own results are uneven by design - some adapters come back strong, others are merely less broken - and the researchers are upfront that not every adapter is worth recovering.
It's a patch, not a promise: useful if you've already sunk time into old-model LoRAs, a lot less interesting if you're just starting out on the newer, faster model anyway.