RoPECraft: Training-Free Motion Transfer with Trajectory-Guided RoPE Optimization on Diffusion Transformers

TL;DR

RoPECraft introduces a training-free method for video motion transfer using diffusion transformers by modifying rotary positional embeddings with optical flow and trajectory optimization, achieving superior results.

Contribution

It presents a novel training-free approach that modifies rotary positional embeddings for motion transfer in diffusion transformers, incorporating optical flow and phase regularization.

Findings

Outperforms recent methods qualitatively and quantitatively

Effective motion encoding via flow-guided RoPE modification

Maintains fidelity to text prompts and reduces artifacts

Abstract

We propose RoPECraft, a training-free video motion transfer method for diffusion transformers that operates solely by modifying their rotary positional embeddings (RoPE). We first extract dense optical flow from a reference video, and utilize the resulting motion offsets to warp the complex-exponential tensors of RoPE, effectively encoding motion into the generation process. These embeddings are then further optimized during denoising time steps via trajectory alignment between the predicted and target velocities using a flow-matching objective. To keep the output faithful to the text prompt and prevent duplicate generations, we incorporate a regularization term based on the phase components of the reference video's Fourier transform, projecting the phase angles onto a smooth manifold to suppress high-frequency artifacts. Experiments on benchmarks reveal that RoPECraft outperforms all recently published methods, both qualitatively and quantitatively.