AINovice2005/pixel-art-bench

🎨 Pixel Art Bench

Pixel Art Bench is a structured-output benchmark designed to evaluate language models on their ability to generate valid, interpretable, and semantically meaningful JSON outputs under strict constraints.

The benchmark focuses on pixel art generation over a fixed 24×24 grid, requiring models to produce outputs that are not only syntactically correct but also structurally and visually coherent.

While many benchmarks evaluate free-form text generation, real-world applications increasingly require:

Strict JSON compliance Structured reasoning Format adherence under constraints

Pixel Art Bench targets this gap by evaluating:

Whether a model can follow rigid output schemas Whether outputs are machine-interpretable Whether generated structures are usable downstream (e.g., rendering)

🧪 Task Definition

Each sample consists of:

A prompt describing a subject A requirement to generate a 24×24 pixel grid

Input

A natural language instruction:

Draw pixel art on a 24x24 grid. Return a JSON object with:

• "grid": 24x24 list of integers (0–9)
• "palette": list of colors used

Subject: the planet Saturn with its rings

Expected Output Format

{ "grid": [[0,1,1,...],[...],...], "palette": ["black", "yellow", "orange"] }

Constraints:

• grid must be 24×24
• values must be integers (0–9)
• output must be valid JSON

📊 Evaluation Protocol

The benchmark evaluates model outputs across three dimensions, using robust parsing and graded scoring functions.

1. JSON Validity

Evaluates whether the model produces a parsable JSON object, using a multi-stage extraction pipeline:

• Direct JSON parsing
• Regex-based extraction of JSON blocks
• Cleanup of formatting artifacts (e.g., markdown code fences)

Outputs are scored as:

1.0 (correct) → valid JSON extracted 0.0 (incorrect) → parsing failed

2. Render Success (Structural Correctness)

Measures whether the output can be interpreted as a valid 24×24 pixel grid.

This is a graded score ∈ [0,1], composed of:

Height score — how close grid height is to 24 Width score — consistency of row lengths Type score — proportion of valid integer cells

Final score: Render = (height + width + type) / 3

This ensures partial credit for structurally plausible outputs.

3. Pixel Art Quality

Evaluates structural richness and completeness of the generated grid.

This is also a graded score ∈ [0,1], combining:

Color diversity

• number of unique integer values
• scaled with a soft cap at ~8 colors

Grid density

• proportion of valid filled cells in the 24×24 grid

Final score: Quality = 0.7 × diversity + 0.3 × density

Aggregate Score

The final benchmark score is computed as:

Score = 0.4 × JSON Validity + 0.3 × Render Success + 0.3 × Pixel Art Quality