TIFA: Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering

What is TIFA?

Images synthesized by text-to-image models (e.g. Stable Diffusion) often do not follow the text inputs well. TIFA is a simple tool to evaluate the fine-grained alignment between the text and the image. This repository contains the code and models for our paper TIFA: Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering. This paper is also accepted to ICCV 2023. Please refer to the project page for a quick overview.

Do I need OpenAI API to run TIFA?

No. We have pre-generated the questions with OpenAI APIs, such that users only need to run the VQA models in this repo to benchmark their text-to-image models. In addition, this repo also provide tools to generate your own questions with GPT-3.5

Why TIFA?

TIFA is much more accurate than CLIP, while being fine-grained and interpretable. It is an ideal choice for fine-grained automatic evaluation of image generation. Meanwhile, this repo also provides tools to allow users to customize their own TIFA benchmark.

Leaderboard

Want to submit results on the leaderboard? Please email the authors.

**************************** Updates ****************************

• 04/14/24: We updated the openai completion function to make it compatible with latest openai APIs.
• 08/22: We released the LLaMA 2 fine-tuned for question generation so that users can generate questions without OpenAI API.
• 04/19: We released the human judgment scores on text-to-image faithufulness used in Table 2.
• 03/24: We released the evaluation package, which includes evaluation code, VQA modules, and question generation modules
• 03/22: We released the TIFA v1.0 captions, questions.

Quick Links

• Installation
• Quick Start
• TIFA v1.0 Benchmark
• TIFA v1.0 Human Annotations on Text-to-Image Faithufulness
• VQA modules
• Question Generation modules
  • Question Generation with GPT-3.5
  • Alternative: Question Generation with LLaMA 2
  • Filtering with UnifiedQA
• TIFA on arbitrary image and text

Installation

Clone and build the repo:

git clone https://github.com/Yushi-Hu/tifa.git
cd tifa
pip install -r requirements.txt
pip install -e .

Quick Start

To benchmark your own text-to-image model:

Suppose we are benchmarking on the text inputs in sample/sample_text_inputs.json. The questions are in sample/sample_question_answers.json.

1. Generate the images from sample/sample_text_inputs.json. We have included sample images in sample/
2. Make a JSON file that points out the location of each image. We include an example sample/sample_imgs.json. The format is like:

{
    "coco_301091": "coco_301091.jpg",
    "drawbench_52": "drawbench_52.jpg"
}

3. Run the following code. The output is in sample/sample_evaluation_result.json.

from tifascore import tifa_score_benchmark
import json

# We recommend using mplug-large
results = tifa_score_benchmark("mplug-large", "sample/sample_question_answers.json", "sample/sample_imgs.json")

# save the results
with open("sample/sample_evaluation_result.json", "w") as f:
    json.dump(results, f, indent=4)

Benchmark on TIFA v1.0 benchmark

To evaluate on TIFA v1.0 benchmark, generate images for text inputs in tifa_v1.0/tifa_v1.0_text_inputs.json. and run

results = tifa_score_benchmark("mplug-large", "tifa_v1.0/tifa_v1.0_question_answers.json", "[YOUR IMAGE PATH]")

# save the results
with open("results.json", "w") as f:
    json.dump(results, f, indent=4)

TIFA v1.0 Benchmark

TIFA v1.0 text inputs are in tifa_v1.0/tifa_v1.0_text_inputs.json

You can also Download here

The GPT-3 pre-generated TIFA v1.0 question and answers are in tifa_v1.0/tifa_v1.0_question_answers.json.

You can also Download here

TIFA v1.0 Human Annotations

We release the human annotations on text-to-image faithufulness in human_annotations to facilitate future research in this direction. Refer to the readme in the directory for details.

VQA Modules

We provide easy interface to run VQA inference. The supported image-to-text models include mplug-large, git-base, git-large, blip-base, blip-large, vilt, ofa-large, promptcap-t5large, blip2-flant5xl. You can also customize your own VQA models in tifascore/vqa_models.py.

To peform inference

from tifascore import VQAModel

# automatically on GPU if detect CUDA.
# support all above VQA models. Here we use mplug-large as an example.
model = VQAModel("mplug-large")

# perform free-form VQA
print(model.vqa("sample/drawbench_8.jpg", "What is the fruit?"))
# bananas

# perform multiple-choice VQA
print(model.multiple_choice_vqa("sample/drawbench_8.jpg", "What is the color of the banana?", choices=['black', 'yellow', 'red', 'green']))
# {'free_form_answer': 'yellow', 'multiple_choice_answer': 'yellow'}

In this example, the image is . The answer to what is the fruit? is bananas, and the answer to what is the color of the banana? is yellow.

Question Generation Modules

TIFA allows users to generate questions on their own text inputs. The question generation contains two part. First is question generation with GPT-3. And then we filter the questions with UnifiedQA. We will release a fine-tuned Flan-T5 model so that the users can generate questions locally without OpenAI API. However, the question quality will be lower than GPT-3 generated ones.

Question generation with GPT-3.5

Notice that since code-davinci-002 is deprecated, we rewrite our code with gpt-3.5-turbo. Since the input length limit is 4097, which is smaller than Codex's, we reduce the number of in-context examples in the current codebase. Thus, the questions will be slightly different from the ones in TIFA v1.0 benchmark.

import openai
from tifascore import get_question_and_answers

openai.api_key = "[Your OpenAI Key]"

text = "a black colored banana."
gpt3_questions = get_question_and_answers(text)
# gpt3_questions is a list of question answer pairs with the same format as sample/sample_question_answers.json

Question generation with LLaMA 2

We also provide an alternative way to generate questions using our fine-tuned LLaMA 2. The question quality are similar to the GPT-3 generated ones, and it is fully open and reproducible. The huggingface link is https://huggingface.co/tifa-benchmark/llama2_tifa_question_generation/

from tifascore import get_llama2_pipeline, get_llama2_question_and_answers

pipeline = get_llama2_pipeline("tifa-benchmark/llama2_tifa_question_generation")

llama2_questions = get_llama2_question_and_answers(pipeline, "a black colored banana.")
# llama2_questions has the same format as the gpt3_questions

Filter with UnifiedQA

To double-check the questions, we also run UnifiedQA to filter GPT-3 generated questions.

import openai
from tifascore import UnifiedQAModel, filter_question_and_answers

unifiedqa_model = UnifiedQAModel("allenai/unifiedqa-v2-t5-large-1363200")

# filter gpt3_questions
filtered_questions = filter_question_and_answers(unifiedqa_model, gpt3_questions)

TIFA on arbitrary image and text

Combining above modules, we can compute TIFA on arbitary image-text pair.

from tifascore import get_question_and_answers, filter_question_and_answers, UnifiedQAModel, tifa_score_single, VQAModel
import openai

openai.api_key = "[Your OpenAI Key]"
unifiedqa_model = UnifiedQAModel("allenai/unifiedqa-v2-t5-large-1363200")
vqa_model = VQAModel("mplug-large")
    
    
img_path = "sample/drawbench_8.jpg"
text = "a black colored banana."
    
# Generate questions with GPT-3.5-turbo
gpt3_questions = get_question_and_answers(text)
    
# Filter questions with UnifiedQA
filtered_questions = filter_question_and_answers(unifiedqa_model, gpt3_questions)
    
# See the questions
print(filtered_questions)

# calucluate TIFA score
result = tifa_score_single(vqa_model, filtered_questions, img_path)
print(f"TIFA score is {result['tifa_score']}")   # 0.33
print(result)

In this example, TIFA score between the text a black colored banana and is 0.33.

Citation

If you find our work helpful, please cite us:

@article{hu2023tifa,
  title={TIFA: Accurate and Interpretable Text-to-Image Faithfulness Evaluation with Question Answering},
  author={Hu, Yushi and Liu, Benlin and Kasai, Jungo and Wang, Yizhong and Ostendorf, Mari and Krishna, Ranjay and Smith, Noah A},
  journal={arXiv preprint arXiv:2303.11897},
  year={2023}
}