{"id":28,"date":"2025-05-09T22:53:01","date_gmt":"2025-05-09T22:53:01","guid":{"rendered":"https:\/\/mscvprojects.ri.cmu.edu\/2025team8-2\/?page_id=28"},"modified":"2025-12-12T17:53:27","modified_gmt":"2025-12-12T17:53:27","slug":"method","status":"publish","type":"page","link":"https:\/\/mscvprojects.ri.cmu.edu\/2025team8-2\/method\/","title":{"rendered":"Method"},"content":{"rendered":"\n

Our approach for high-fidelity video editing is built upon a two-pronged method:\u00a0Dual-Path Diffusion Sampling<\/strong>\u00a0for consistent generation, and\u00a0VLM-Guided Semantic Refinement<\/strong>\u00a0for ensuring fidelity to the edit instruction.<\/p>\n\n\n\n

1. Dual-Path Diffusion Sampling<\/h3>\n\n\n\n

We address the challenge of maintaining semantic consistency between the edited initial image and the resulting video sequence by employing a\u00a0Dual-Path Diffusion Sampling<\/strong>\u00a0scheme (Algorithm 2). Inspired by recent work Coupled Diffusion Sampling (Alzayer et al), which introduces a coupling function that forces two independent diffusion sampling trajectories to be “closer” to one another, we extend this idea to bridge image editing and video generation models.<\/p>\n\n\n\n

Concretely, during the reverse diffusion process of two concurrent diffusion models, an Image Editing model (\u03b8I\u200b) and a Text-to-Video model (\u03b8V\u200b), we introduce a\u00a0Dual-Path Guidance Step<\/strong>. This step forces the two sampling paths to align by calculating a cross-guidance term based on the discrepancy of their respective\u00a0x0\u200b\u00a0predictions (\u03bb(x^0,I\u200b\u2212x^0,V\u200b)). This coupling term is applied iteratively to the intermediate latent states. This mechanism ensures that the generated video frames are not only temporally coherent but also semantically consistent with the final edited image, leading to a unified, high-quality output.<\/p>\n\n\n\n

\"\"<\/figure>\n\n\n\n

2. VLM-Guided Semantic Refinement<\/h3>\n\n\n\n

To guarantee that the generated video faithfully executes the desired edit while preserving unedited content, we incorporate a Vision-Language Model (VLM) into a refinement loop.<\/p>\n\n\n\n

The VLM serves as a sophisticated semantic critic, evaluating the generated output against two key criteria:<\/p>\n\n\n\n