OmniPersona: Holistic Identity-Preserving Video Creation and Editing via Spatiotemporal Decoupled Persona Injection in Unified DiT

Wed, 01 Jul 2026 00:00:00 +0000

Role: Co-first Author (Equal Contribution)

Overview

Generation Example

The static input conditions and the animated output are combined into one synchronized video, following the presentation layout.

OmniPersona introduces Multimodal Holistic Persona Representation and spatiotemporally decoupled persona injection for unified video creation and editing. I co-led the work and contributed to the holistic identity representation module, combining face identity features with DINOv2 and CLIP features for clothing, hairstyle, and body characteristics.

On reference-to-video generation, OmniPersona reaches 79.3% face similarity and 68.7% body similarity, improving over the strongest baseline by 5.2 and 34.2 percentage points, respectively.

Wan-VACE-based Video Object and Shadow Removal

Thu, 01 Jan 2026 00:00:00 +0000

This Huawei collaboration project addresses the shortage of high-quality paired data for video object removal, especially when shadows and reflections must also be removed.

I independently built a data generation and fine-tuning pipeline using Wan2.2-5B, SAM2, and a baseline removal model. From roughly 4,000 generated groups, I selected 1,070 high-quality training pairs and retained 2,500 difficult cases for evaluation. I then fine-tuned the VACE module of Wan2.1-VACE-1.3B with DiffSynth-Studio, Accelerate, and DeepSpeed on six NVIDIA A800 GPUs.

All videos are displayed at 480 × 276, matching the output resolution of our model.

Condition Video vs. Baseline

Condition Video Baseline ↔

Drag the slider to compare both videos frame by frame.

Condition Video vs. Ours

Condition Video Ours ↔

Drag the slider to inspect object and shadow removal.