TIFA-160#

Overview#

TIFA-160 is a text-to-image benchmark with 160 carefully curated prompts designed to evaluate the faithfulness and quality of generated images using automated VQA-based evaluation.

Task Description#

  • Task Type: Text-to-Image Generation Evaluation

  • Input: Text prompt for image generation

  • Output: Generated image evaluated using PickScore metric

  • Size: 160 prompts

Key Features#

  • Compact, high-quality prompt set for efficient evaluation

  • Uses PickScore for human preference alignment

  • Tests diverse image generation capabilities

  • Supports both new generation and pre-existing image evaluation

  • Reproducible evaluation pipeline

Evaluation Notes#

  • Default configuration uses 0-shot evaluation

  • Primary metric: PickScore for human preference alignment

  • Evaluates images from the test split

  • Part of the T2V-Eval-Prompts dataset collection

Properties#

Property

Value

Benchmark Name

tifa160

Dataset ID

AI-ModelScope/T2V-Eval-Prompts

Paper

N/A

Tags

TextToImage

Metrics

PickScore

Default Shots

0-shot

Evaluation Split

test

Data Statistics#

Metric

Value

Total Samples

160

Prompt Length (Mean)

56.13 chars

Prompt Length (Min/Max)

13 / 182 chars

Sample Example#

Subset: TIFA-160

{
  "input": [
    {
      "id": "9de3e3b1",
      "content": "A Christmas tree with lights and teddy bear"
    }
  ],
  "id": 0,
  "group_id": 0,
  "metadata": {
    "prompt": "A Christmas tree with lights and teddy bear",
    "category": "",
    "tags": {},
    "id": "TIFA160_0",
    "image_path": ""
  }
}

Prompt Template#

No prompt template defined.

Usage#

Using CLI#

evalscope eval \
    --model YOUR_MODEL \
    --api-url OPENAI_API_COMPAT_URL \
    --api-key EMPTY_TOKEN \
    --datasets tifa160 \
    --limit 10  # Remove this line for formal evaluation

Using Python#

from evalscope import run_task
from evalscope.config import TaskConfig

task_cfg = TaskConfig(
    model='YOUR_MODEL',
    api_url='OPENAI_API_COMPAT_URL',
    api_key='EMPTY_TOKEN',
    datasets=['tifa160'],
    limit=10,  # Remove this line for formal evaluation
)

run_task(task_cfg=task_cfg)