End-to-End Joint Optimization of MLLM and MMDiT
Fine-tune both the multimodal language model and the video diffusion generator end-to-end using parameter-efficient adapters so the understanding module aligns its hidden states with generative physics.