ConsID-Gen
View-Consistent and Identity-Preserving
Image-to-Video Generation

Mingyang Wu¹ Ashirbad Mishra² Soumik Dey² Shuo Xing¹ Naveen Ravipati² Hansi Wu² Binbin Li² Zhengzhong Tu¹^†

¹Texas A&M University ²eBay Inc.

^†Corresponding Author

Accepted by CVPR 2026

Paper Code

Dataset (Public Set)

Dataset (Proprietary Set)

TL;DR

We propose ConsID-Gen, a view-assisted I2V framework that tackles appearance drift via multi-view geometric priors and unified representation, alongside ConsIDVid, a new benchmark and evaluation suite dedicated to evaluating identity preservation.

Our contribution

A holistic I2V benchmark for identity preservation, with a diverse dataset and a novel multi-view evaluation suite.
ConsID-Gen introduces unified representation before diffusion, with multi-view guidance and improved cross-modal alignment.
Showing that ConsID-Gen outperforms open-source SOTAs in identity consistency and in human evaluation.

Overview

Image-to-Video (I2V) generation often struggles with maintaining fine-grained object identity and geometric stability due to the sparsity of single-view observations. To address this, we approach the problem from both data and model perspectives, introducing a comprehensive framework for identity-preserving video generation.

We curate ConsIDVid, a large-scale object-centric dataset built with a scalable pipeline for high-quality, temporally aligned videos. Alongside this, we establish ConsIDVid-Bench, a novel benchmarking framework designed to evaluate multi-view consistency using metrics sensitive to subtle geometric and appearance deviations.

We further propose ConsID-Gen, a view-assisted I2V generation framework. It augments the first frame with unposed auxiliary views and fuses semantic and structural cues via a dual-stream visual–geometric encoder and a text–visual connector. This unified conditioning for the Diffusion Transformer backbone achieves superior identity fidelity and temporal coherence, consistently outperforming leading models like Wan2.1 and HunyuanVideo.

Method Overview

ConsID-Gen: Model Architecture

ConsID-Gen takes as input the first frame, two uncalibrated images, and a text instruction. Our Dual-Visual Encoder combines a Visual Encoder and a Geometry Encoder to extract visual-appearance and geometric representations. A unified multimodal interaction projector then fuses these features with the prompt to generate conditioning tokens for the DiT backbone.

ConsIDVid & ConsIDVid-Bench

Data statistics and benchmark distribution

Image-to-Video Comparison

Side-by-side comparisons across object categories and aspect ratios highlight stronger identity preservation, geometric stability, and visual faithfulness than the baseline.

Baseline (Wan2.1-I2V) ConsID-Gen

Citation

@InProceedings{Wu_2026_CVPR,
    author    = {Wu, Mingyang and Mishra, Ashirbad and Dey, Soumik and Xing, Shuo and Ravipati, Naveen and Wu, Hansi and Li, Binbin and Tu, Zhengzhong},
    title     = {ConsID-Gen: View-Consistent and Identity-Preserving Image-to-Video Generation},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2026},
    pages     = {1853-1863}
}

ConsID-Gen View-Consistent and Identity-Preserving Image-to-Video Generation