Key holding or allegation
Reported appellate reasoning distinguished LoRA statistical parameterisation from reproduction of the training image itself while affirming infringement where the trained model was used to generate substantially reproducing outputs; platform liability was rejected on the reported facts.
Why it matters
One of the most important Chinese direct authorities on learning/training versus downstream output reproduction.
Facts
User trained LoRA models using protected Medusa imagery and generated/distributed related outputs through a platform.
Issues before the court
Whether LoRA training/model parameterisation itself reproduced protected expression, whether outputs infringed, and whether the platform was liable.
What the court did not decide
Full official appellate text was not secured in the project snapshot; wording must remain tied to verified reporting until primary text is obtained.
Procedural timeline
Shanghai IP Court
Appellate judgment: Reported appellate reasoning distinguished LoRA statistical parameterisation from reproduction of training images while affirming infringement where outputs substantially reproduced protected Medusa character expression; platform liability was rejected on reported facts.