timmzak/ReText-LLaVA

ReText: Fine-tuned VLM Weights

LLaVA fine-tuned on UFine3C for person image captioning

GitHub · Paper (arXiv) · Textual annotations

Accepted by WACV 2027

Overview

This directory contains the fine-tuned VLM weights used to generate textual descriptions for ReText and a script for captioning your own images using HAM.

Release contents

File	Description
ham-release/ham/non_lora_trainables.safetensors	Fine-tuned weights loaded on top of the base model.
ham-release/ham/config.json	Model configuration.
ham-release/caption_own_images.py	Batch caption generation script.
ham-release/requirements.txt	Python dependencies.

The base models are not bundled. The script uses liuhaotian/llava-v1.6-vicuna-7b and openai/clip-vit-large-patch14-336 from Hugging Face unless local copies are present under MODELS_DIR in directories named llava-v1.6-vicuna-7b and clip-vit-large-patch14-336.

Caption your own images

Requirements: Linux, Python 3.10, and an NVIDIA GPU with approximately 16 GB of memory.

1. Download the release

Install the Hugging Face Hub client and download the files:

pip install huggingface_hub
hf download timmzak/ReText-LLaVA --local-dir ReText-LLaVA
cd ReText-LLaVA

2. Get HAM and the captioning script

Run from the downloaded release directory:

git clone https://github.com/MPI-Lab/HAM.git
cp ham-release/caption_own_images.py ham-release/requirements.txt HAM/
cd HAM

3. Create the environment

python3.10 -m venv ham-env
source ham-env/bin/activate
pip install -r requirements.txt

Do not run pip install -e . for HAM itself.

4. Configure paths

Edit the settings at the top of caption_own_images.py:

MODELS_DIR = "/absolute/path/to/ReText-LLaVA/ham-release"
IMAGE_DIR = "/absolute/path/to/images"
OUTPUT_PATH = "captions.jsonl"

MODELS_DIR must contain the ham directory with the weights and configuration. The script also exposes GPU, BATCH_SIZE, and NUM_WORKERS settings.

5. Generate captions

python caption_own_images.py

Output

The script writes a JSONL file with one record per image. Each record has an image field containing the path relative to IMAGE_DIR and a caption field containing the generated description.

Restarting the script with the same output file skips images already recorded there. The captioning prompt can be changed through PROMPT; each caption uses a randomly chosen annotator style, as in HAM.

Citation

@article{mamedov2026retext,
  title={ReText: Text boosts generalization in image-based person re-identification},
  author={Mamedov, Timur and Kvanchiani, Karina and Konushin, Anton and Konushin, Vadim},
  journal={arXiv preprint arXiv:2602.05785},
  year={2026}
}