Search papers, labs, and topics across Lattice.
Dept. of Artificial Intelligence, Hanyang University, South Korea
7
0
7
Achieving fairness in image generation without retraining, TES optimizes text embeddings in real-time to combat demographic bias effectively.
Conditioning entanglement in text-to-image models can be effectively resolved by decoupling subject and context pathways, leading to superior personalization outcomes.
Recovering over 95% of individual-expert performance from a single merged model could revolutionize multi-task learning efficiency.
Anisotropic gradient scaling in LoRA can severely hinder model adaptation, but SDS-LoRA effectively mitigates this issue, leading to superior performance.
Forget training: this method detects anomalies by simply measuring how much a query patch *needs* to change to fit in with normal data, using a clever graph Laplacian energy minimization.
Forget hand-labeled bias attributes: BiasEdit automatically detects and edits biases in image datasets, paving the way for fairer visual classifiers without extra training.
Forget retraining: BiCo lets you transfer task-specific knowledge between different model architectures with just a single forward-backward pass.
