Time Adapter for Kandinsky-5 I2V

Overview

This repository contains a lightweight TA-only temporal adapter checkpoint for Kandinsky-5 image-to-video generation.

Lower target FPS typically leads to slower perceived motion, while higher target FPS leads to faster perceived motion.

The released checkpoint keeps the base DiT weights unchanged and adds only the time-adapter parameters for global FPS conditioning and latent-time control.

Use this checkpoint with the `time_adapter` branch of kandinsky-5 and the config configs/k5_i2v_time_adapter_step5750_ta_only.yaml.

Download

Download the complete repository with the Hugging Face CLI:

hf download kandinskylab/Time-Adapter --local-dir Time-Adapter

Or use huggingface_hub from Python:

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="kandinskylab/Time-Adapter",
    local_dir="Time-Adapter",
)

Examples

The gallery below shows TA-only examples for pose motion control and forest motion control at different target FPS values.

The config follows the standard kandinsky-5 layout and expects model files under weights/model, weights/vae, weights/text_encoder, and weights/text_encoder2.