Minimax H3 Image Subject LoRA: Has anyone achieved good similarity?
Training subject LoRA with Minimax H3 Image yields poor similarity, while the same dataset performs excellently on Qwen 2.1, Krea 2, and Moss. Increasing training epochs or steps showed no improvement; the Minimax H3 video model actually learns subjects better on static images. Motion learning is also difficult, and the poster is seeking advice on training settings for high-fidelity similarity.