Anchoring and Rescaling Attention for Semantically Coherent Inbetweening

CVPR 2026 ⭐ Highlight

Tae Eun Choi*    Sumin Shim*    Junhyeok Kim    Seong Jae Hwang

Yonsei University

Overview

This repository contains the official implementation of Anchoring and Rescaling Attention for Semantically Coherent Inbetweening, a training-free approach for text-conditioned generative inbetweening that improves semantic fidelity, frame consistency, and pace stability. Given the first frame, last frame, and a text prompt, our method generates semantically coherent intermediate frames while enhancing semantic alignment, temporal consistency, and motion pacing without additional model training. We also introduce TGI-Bench, a benchmark for evaluating text-conditioned generative inbetweening across diverse sequence lengths and motion scenarios.

---

Dataset

The TGI-Bench dataset is available on Hugging Face here.

Installation

We recommend using a conda environment.

Create environment

Python 3.10 or higher is required.

conda create -n tgi python=3.10
conda activate tgi

Install dependencies

pip install -r requirements.txt

Once this is done, the environment setup is complete.

Run Inference

To run inference with the default settings:

python inference.py

Optional Arguments

You can customize inference with additional arguments:

python inference.py \
  --prompt "A freight train moves forward through heavy falling snow." \
  --img_first example/first.jpg \
  --img_last example/last.jpg \
  --seed 0 \
  --num_frames 81 \
  --w_edge 8 \
  --s_edge 1.06 \
  --s_mid 0.94 \
  --beta_end 0.7 \
  --beta_mid 0.3

Argument description

• --prompt: text prompt
• --img_first: path to the first frame
• --img_last: path to the last frame
• --seed: random seed
• --num_frames: number of frames (25, 33, 65, 81)
• --w_edge: width of the fast region near both ends
• --s_edge: scaling parameter near keyframes
• --s_mid: scaling parameter for middle frames
• --beta_end: endpoint weighting parameter
• --beta_mid: middle-region weighting parameter

If not specified, default example values are used.