“A content creator in a neon-lit room.”Human identity
Overview
Identity from images, motion from videos.
We fine-tune a pretrained video model with subject-image pairs for identity injection and 4,000 arbitrary videos for motion-awareness preservation. The resulting model handles unseen reference images without subject-video pairs or per-subject optimization.
Results from CogVideoX-5B and Wan 2.2-5B.
Our approach
Factorize identity and motion.
We cast adaptation as two complementary tasks in one shared model. Stochastic switching injects subject identity from images while replaying a small video set to retain the pretrained temporal prior.
Identity from images
Paired views of the same subject teach appearance without requiring a customized video dataset.
Motion from ordinary videos
Random reference frames and image-token dropout discourage first-frame copying and frozen outputs.
Independent update geometry
Identity and motion gradients approach zero cosine similarity while their norms remain non-negligible.
Results
Qualitative results
Zero-shot generations for unseen reference images, without subject-specific fine-tuning.
“A musician in a Hawaiian shirt plays a ukulele.”Clothing reference
“A model walks outside in a red evening dress.”Clothing reference
“A teddy bear is playing guitar.”Character identity
“A dog races through the waves.”Fast motion
“An anime girl takes a selfie beneath cherry blossoms.”Stylized subject
Comparison
Qualitative comparison
All methods use the reference image and prompt shown with each example.
Reference image + prompt
“A musician wearing a Hawaiian shirt strums a ukulele on a boardwalk.”
Reference image + prompt
“The ruffled evening dress is worn by a model outside under a blue sky. The model walks slowly and turns to show its silhouette, detailed folds, and natural movement.”
Reference image + prompt
“A young man wearing a green sleeveless top and red headphones appears as a content creator in a room illuminated with purple and blue neon lights.”
Reference image + prompt
“A man gently clutches a bouquet of vibrant flowers and glances at the camera in a lush garden filled with colorful blooms.”
Citation
BibTeX
@inproceedings{kim2026learning,
author = {Kim, Daneul and Zhang, Jingxu and Jin, Wonjoon and Cho, Sunghyun and Dai, Qi and Park, Jaesik and Luo, Chong},
title = {Learning Zero-Shot Subject-Driven Video Generation Using 1% Compute},
booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
year = {2026}
}