Oneig Bench Eval

oneig-bench-eval

Nucleus-Image: Sparse MoE for Image Generation — Akiti et al. (2026) (arXiv:2604.12163, 2026)

What this evaluates

Measures broader image generation capabilities across five axes: alignment, text rendering, reasoning, style, and diversity.

Datasets

• OneIG-Bench — total ?; splits: test (-1)

Metrics

• Overall (primary) — range: [0, 1]
  • Mean score across five axes: alignment, text rendering, reasoning, style, and diversity.

Input / output format

Input: Text prompt designed to test alignment, text rendering, reasoning, style, and diversity.

Output: Generated image at 1024x1024 resolution, 50 inference steps, CFG scale 8.0.

Scoring recipe

dims = ['alignment', 'text', 'reasoning', 'style', 'diversity']
scores = {d: [] for d in dims}
for prompt in prompts:
    img = model.generate(prompt, steps=50, cfg=8.0, res=1024)
    for d in dims:
        scores[d].append(evaluator.score(img, prompt, dimension=d))
overall = mean(mean(v) for v in scores.values())
return overall

Common pitfalls

• Diversity scores tend to be lower across most models, indicating a known limitation in generating varied outputs from similar prompts.
• Text rendering and style evaluation may rely on specialized models or human-like scoring that can vary in strictness.

Evidence (verbatim from paper)

Nucleus-Image achieves an overall score of 0.522, placing it among the top tier of open and proprietary models and ahead of Imagen4 (0.515) and Recraft V3 (0.502).

Citation

@misc{akiti2026nucleusimage,
  title={Nucleus-Image: Sparse MoE for Image Generation},
  author={Akiti et al. (2026)},
  year={2026},
  note={arXiv:2604.12163}
}

• arXiv: 2604.12163