Want to improve your SOTA image captioning experience for Stable Diffusion? The IMAGE Interrogator is here to make this job easier!

The IMAGE Interrogator is a variant of the original CLIP Interrogator tool that brings all original features and adds other large models like LLaVa and CogVml for SOTA image captioning. Then you can train with fine-tuning on your datasets or use resulting prompts with text-to-image models like Stable Diffusion on DreamStudio to create cool art!

• 2024.03.17: New release v1.0.1 - Added support for moondream1 and moondream2 models; Added support for Qwen VL Chat. Change to keep the model in memory at each execution. 🚨A clean install is required!🚨

Use Python version 3.10.* and have the Python virtual environment installed. For linux users will need to install Tkinter with the follow command:

sudo apt-get install python3-tk

Then run the following commands in the terminal:

git clone https://github.com/DEVAIEXP/image-interrogator.git
cd image-interrogator
(for linux  ) source install_linux.sh
(for windows) install_windows.bat

Assuming T for True and F for False:

• Passing T on first parameter will force deletion of the Venv and repositories folders for a clean installation.

(for linux  ) source install_linux.sh T
(for windows) install_windows.bat T

• Passing T on second parameter will disable use of Venv during installation.

(for linux  ) source install_linux.sh F T
(for windows) install_windows.bat F T

If you ran the installation without Venv, you must edit the script files and change the USE_VENV variable to the value 0 before starting to run the tool.

(for linux  ) source start.sh
(for windows) start.bat

The start.sh and start.bat scripts trigger the image-interrogator.py script via Python. The python script allows you to enter some parameters:

• --cache-dir: Folder to download cache models. If specified, models will be fully downloaded to the specified path instead of the HF cache.
• --share: Create a public access link Edit this files and change last line, for e.g:

(for linux  ) python image-interrogator.py --cache-dir "/mnt/c/models" --share
(for windows) python image-interrogator.py --cache-dir "c:\models" --share

Note: The linux script is configured to run on WSL 2. If you are running on a linux installation you will need to adjust the LD_LIBRARY_PATH variable in the file with the correct path of your CUDA Toolkit.

IMAGE Interrogator support 4-bit quantization and 8-bit quantization (except for CogVLM and CogAgent only 4-bit quantization is enabled) for low memory usage. Precision type parameters have also been added to the interface such as FP16 and BF16. On systems with low VRAM you can try 4-bit quantization or check Optimize settings for low VRAM in Load options from interface. It will reduce the amount of VRAM needed (at the cost of some speed and quality).

• Folder Path: Image folder path for batch processing.
• Single Image Upload: For processing a single image.

The Generate options lets you enable/disable resources that may be generated during or after the prompt is generated.

• Include image features in the prompt: Enable the Features tab where it is possible to select OpenCLIP pretrained CLIP models. It will add image analysis result features such as (artists, flavors, media, movements, trends, negative prompt). Note: for negative prompts, the caption model will be ignored.
• Don't save dataset images: This will disable copying of the image in the caption output path.
• Generate individual caption file: When enabled, this generates individual '.txt' caption files for each image.
• Generate caption in the original path instead of the output folder: When enable will save caption files and datasets files in the image original path. (Only for batch mode).
• Generate dataset: This will compile a dataset into the output path so that it can be loaded into hugging-face datasets or used in model training.
• Dataset name: The name for dataset folder in output path.

In this tab you can choose your preferred model from the list for generating the caption. For some models like LLaVa some additional parameters are available such as: temperature and top p. LLaVa, CogAgent, CogVLM and Kosmos-2 allow you to use question prompts for generation. We automatically suggest question prompts when the template is selected, but you can change the prompt text to ask what you want about an image for the template. Look for more information about how to generate prompts in these models directly on the official page of the chosen model.

For selection of OpenCLIP pretrained CLIP Model. Only one feature mode can selected at a time.

• Generate prompt for single image: The caption/prompt will be generated in the output path for the image selected in the defined settings.
• Batch process for folder: The caption/prompt will be generated for all images in the folder selected for the output path with the defined settings.
• Test all caption models: A csv file will be generated in the output path, delimited by ';' which will contain the caption/prompt generated in each available caption model for comparison purposes. This does not include results from CLIP Models.
• Test all CLIP models: If you chose a winning model in the previous test, you can now test all available CLIP templates for the selected caption template. The result will also be stored in the output path in a csv file delimited by ';'.

It returns a list of words in each feature and their scores for the given image, model.

If you update the version of LLaVa, Gradio and PIL dependencies, this tool will not work correctly. When there is a need to update these dependencies, we will update them in our repository. Whenever there is a new update to this repository, it will be necessary to delete the 'repositories' directory and run the installation script again.

This project is released under the MIT license.

This project is based on CLIP Interrogator, Some codes are brought from LLaVa. Thanks for their awesome works.

If you have any questions, please contact: [email protected]