A 4-step LoRA for MiniMax H3, utilizing "QKVM warm-start."
Version 2.0 preserves the style of the original model while reducing image blur and enhancing stability during long-sequence generation.
The design minimizes the addition of elements—other than those facilitating convergence at a low step count—to the model’s inherent characteristics; at this stage, it represents the most well-balanced iteration within our current experiments.
