Mamoda2.5: Enhancing Unified Multimodal Model with DiT-MoE

TL;DR

Mamoda2.5 is a unified multimodal diffusion model with a Mixture-of-Experts backbone that achieves high performance and efficiency in video editing and content moderation tasks.

Contribution

It introduces a scalable MoE-based diffusion transformer architecture and a novel distillation and reinforcement learning framework for fast, high-quality video editing.

Findings

Achieves top-tier performance on VBench 2.0 and OpenVE-Bench.

Surpasses open-source models and matches top proprietary models in video editing quality.

Accelerates inference by up to 95.9 times compared to baselines.

Abstract

We present Mamoda2.5, a unified AR-Diffusion framework that seamlessly integrates multimodal understanding and generation within a single architecture. To efficiently enhance the model's generation capability, we equip the Diffusion Transformer backbone with a fine-grained Mixture-of-Experts (MoE) design (128 experts, Top-8 routing), yielding a 25B-parameter model that activates only 3B parameters, significantly reducing training costs while scaling up the model capacity. Mamoda2.5 achieves top-tier generation performance on VBench 2.0 and sets a new record in video editing quality, surpassing evaluated open-source models and matching the performance of current top-tier proprietary models, including the Kling O1 on OpenVE-Bench. Furthermore, we introduce a joint few-step distillation and reinforcement learning framework that compresses the 30-step editing model into a 4-step model and greatly accelerates model inference. Compared to open-source baselines, Mamoda2.5 achieves up to $95.9\times$ faster video editing inference. In real-world applications, Mamoda2.5 has been successfully deployed for content moderation and creative restoration tasks in advertising scenarios, achieving a 98% success rate in internal advertising video editing scenario.