Test-Time Visual In-Context Tuning (Unofficial)

This repo is an unofficial implementation of Test-Time Visual In-Context Tuning (VICT). Since the original paper leaves some details (e.g., creating noise data and visual prompts) a bit vague, this implementation follows the default settings of Painter.

Warning

I’ve tried my best to match Painter’s results (still a few small gaps).
VICT requires a lot of time for test-time adaptation (TTA) on each sample, and the authors tested on a custom dataset (by adding noise to the original). That part isn’t reproduced here.

Installation

uv sync

# This is important to load the env
source .venv/bin/activate

mim install "mmcv-full==1.7.2"
mim install "mmdet==2.28.2"
mim install "mmpose==0.29.0"

python misc/fix_mm_file.py

Dataset Preparation

Dataset setup is exactly the same as Painter. 👉 See Painter’s DATA.md for instructions.

Evaluation

All inference scripts live in the inference folder.
Run them with:

# ensure the environment is loaded
bash script/inference.sh

Options

Usage: script/inference.sh [--num-gpu N] [--save-dir DIR] [--data-path PATH] [--use-ema true|false] [--resume true|false] [--optimize-steps N] [--mixed-precision MODE]

Arguments:
  --num-gpu N            (Optional) Number of GPUs to use. Default: 1
  --save-dir DIR         (Optional) Directory to save outputs. Default: results
  --data-path PATH       (Optional) Path to data. Default: datasets
  --use-ema true|false   (Optional) Use EMA. Default: false
  --resume true|false    (Optional) Resume from checkpoint. Default: false
  --optimize-steps N     (Optional) Optimize steps. Default: 0
  --mixed-precision MODE (Optional) Mixed precision mode ('no', 'fp16'). Default: no
  -h, --help             Show this help message.

Example

Run with 8 GPUs and EMA updates enabled:

bash script/inference.sh --num-gpu 8 --use-ema true

Note

If you leave --optimize-steps at the default (0), it will fall back to Painter’s behavior.
To align with the VICT paper, set it to 60.

Acknowledgments

• Painter
• pytorch_ema

Citation

Please cite the original paper:

@inproceedings{xie2025test,
  title={Test-Time Visual In-Context Tuning},
  author={Xie, Jiahao and Tonioni, Alessio and Rauschmayr, Nathalie and Tombari, Federico and Schiele, Bernt},
  booktitle={Proceedings of the Computer Vision and Pattern Recognition Conference},
  pages={19996--20005},
  year={2025}
}