polym-eval-score-image-gemini

Name: polym-eval-score-image-gemini
Availability: InStock
Author: byteplus-sa

Score AI-generated images against text prompts using Gemini models.

8|Updated May 13, 2026

Search Tags:#gemini #json output #reference images #vertex ai #image evaluation #image scoring #a/b comparison

Ready to run

npx skills add https://github.com/byteplus-sa/polym --skill polym-eval-score-image-gemini

💬 Prompt for ChatGPT / Claude Chat▼

Please help me install this Agent Skill.
Skill: polym-eval-score-image-gemini
Source: https://github.com/byteplus-sa/polym/tree/main/skills/polym-eval-score-image-gemini
Command: npx skills add https://github.com/byteplus-sa/polym --skill polym-eval-score-image-gemini

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires google-genai, Pillow, python-dotenv, openai>=2.0.0, and includes scripts (resource) components.

What problem does it solve?

It removes the manual effort of visually grading AI-generated images by producing structured, dimension-level scores and reasons automatically.

Core Features & Use Cases

Pointwise grading: Scores a single generated image across multiple quality dimensions on a 0-5 scale, plus an overall rationale.
Pairwise A/B comparison: Compares two images and returns dimension-wise winners (A, B, or tie) with a short justification.
Backend flexibility: Uses Vertex AI (default) or ByteDance AIDP through OpenAI-compatible APIs.
Optional reference images: Adds context for identity, brand, style, and edit consistency evaluation when reference images are provided.

Quick Start

Run the pointwise evaluation with Vertex AI using prompt and generated image file by executing the command: "python scripts/evaluate.py --mode pointwise --prompt "A golden retriever sitting in a garden" --image generated.png".

Frequently Asked Questions about polym-eval-score-image-gemini

High-intent search queries and answers about installing and using this skill.

FAQPage Schema

How do I score and evaluate AI-generated images against a text prompt?▼

You can score AI-generated images against a text prompt by using Gemini models to evaluate prompt fidelity, structural correctness, and artifacts, returning structured dimension scores from 0-5 in strict JSON format.

Can I use reference images to evaluate brand or style consistency?▼

Yes, you can use reference images to add context for identity, brand, style, and edit consistency evaluation, enabling reference-conditioned consistency checks alongside standard prompt fidelity scoring.

How do I run an A/B comparison between two generated images?▼

You can run an A/B comparison by executing the pairwise evaluation mode, which compares two images and returns dimension-wise winners (A, B, or tie) with a short justification in JSON format.

Does image evaluation work with both Vertex AI and AIDP?▼

Yes, the image evaluation supports backend flexibility, using Vertex AI by default or ByteDance AIDP through OpenAI-compatible APIs, requiring configured environment credentials for the selected backend.

What JSON output format does the pairwise image comparison return?▼

The pairwise image comparison returns strict JSON containing A, B, or tie judgments for each quality dimension, along with a short text justification explaining the winning image's superiority.