{
  "schemaVersion": "benmark-task-evidence.v1",
  "task": {
    "task_id": "os_production_static_image_generation",
    "scoring_mode": "hybrid",
    "score_status": "scored",
    "model_results": [
      {
        "model_id": "openai-gpt-image-2.5-sunburst",
        "model_label": "GPT Image 2.5 Sunburst",
        "provider": "OpenAI",
        "model_profile_ids": [
          "default"
        ],
        "quality_score_10": 8.84,
        "selected_prompt_arm": "scaffolded",
        "publication_status": "provisional",
        "evidence_maturity": "provisional",
        "last_tested_at": "2026-09-17T08:08:43.474000+00:00",
        "evaluated_item_count": 10,
        "score_count": 10,
        "scored_run_count": 1,
        "mean_cost_per_scored_output_usd": 0.224483,
        "cost_basis": null,
        "costed_scored_output_count": 10
      },
      {
        "model_id": "runway-seedream-5-pro",
        "model_label": "Seedream 5 Pro",
        "provider": "Runway / ByteDance",
        "model_profile_ids": [
          "default"
        ],
        "quality_score_10": 7.64,
        "selected_prompt_arm": "baseline",
        "publication_status": "provisional",
        "evidence_maturity": "provisional",
        "last_tested_at": "2026-10-16T00:32:26.736006+00:00",
        "evaluated_item_count": 10,
        "score_count": 10,
        "scored_run_count": 1,
        "mean_cost_per_scored_output_usd": null,
        "cost_basis": null,
        "costed_scored_output_count": 0
      },
      {
        "model_id": "openai-gpt-image-2",
        "model_label": "GPT Image 2",
        "provider": "OpenAI",
        "model_profile_ids": [
          "default"
        ],
        "quality_score_10": 7.44,
        "selected_prompt_arm": "scaffolded",
        "publication_status": "provisional",
        "evidence_maturity": "provisional",
        "last_tested_at": "2026-10-16T00:32:26.722899+00:00",
        "evaluated_item_count": 10,
        "score_count": 10,
        "scored_run_count": 1,
        "mean_cost_per_scored_output_usd": null,
        "cost_basis": null,
        "costed_scored_output_count": 0
      },
      {
        "model_id": "google-gemini-3-pro-image",
        "model_label": "Gemini 3 Pro Image",
        "provider": "Google Gemini",
        "model_profile_ids": [
          "default"
        ],
        "quality_score_10": 7.12,
        "selected_prompt_arm": "baseline",
        "publication_status": "provisional",
        "evidence_maturity": "provisional",
        "last_tested_at": "2026-10-16T00:32:26.710159+00:00",
        "evaluated_item_count": 10,
        "score_count": 10,
        "scored_run_count": 1,
        "mean_cost_per_scored_output_usd": null,
        "cost_basis": null,
        "costed_scored_output_count": 0
      }
    ],
    "scaffold_comparison": {
      "model_results": [
        {
          "model_id": "google-gemini-3-pro-image",
          "model_label": "Gemini 3 Pro Image",
          "status": "partial",
          "publication_status": "provisional",
          "comparison_metric_key": "quality_score_10",
          "pairing_basis": "explicit_shared_identity",
          "comparison_unit_count": 5,
          "baseline_score_10": 7.12,
          "scaffolded_score_10": 7.04,
          "unpaired_baseline_count": 20,
          "unpaired_scaffolded_count": 20,
          "unscored_comparison_unit_count": 20,
          "invalid_comparison_unit_count": 0
        },
        {
          "model_id": "openai-gpt-image-2",
          "model_label": "GPT Image 2",
          "status": "complete",
          "publication_status": "provisional",
          "comparison_metric_key": "quality_score_10",
          "pairing_basis": "explicit_shared_identity",
          "comparison_unit_count": 5,
          "baseline_score_10": 7.16,
          "scaffolded_score_10": 7.44,
          "unpaired_baseline_count": 0,
          "unpaired_scaffolded_count": 0,
          "unscored_comparison_unit_count": 0,
          "invalid_comparison_unit_count": 0
        },
        {
          "model_id": "openai-gpt-image-2.5-sunburst",
          "model_label": "GPT Image 2.5 Sunburst",
          "status": "complete",
          "publication_status": "published",
          "comparison_metric_key": "quality_score_10",
          "pairing_basis": "explicit_shared_identity",
          "comparison_unit_count": 5,
          "baseline_score_10": 8.44,
          "scaffolded_score_10": 8.84,
          "unpaired_baseline_count": 0,
          "unpaired_scaffolded_count": 0,
          "unscored_comparison_unit_count": 0,
          "invalid_comparison_unit_count": 0
        },
        {
          "model_id": "runway-seedream-5-pro",
          "model_label": "Seedream 5 Pro",
          "status": "complete",
          "publication_status": "provisional",
          "comparison_metric_key": "quality_score_10",
          "pairing_basis": "explicit_shared_identity",
          "comparison_unit_count": 5,
          "baseline_score_10": 7.64,
          "scaffolded_score_10": 7.48,
          "unpaired_baseline_count": 0,
          "unpaired_scaffolded_count": 0,
          "unscored_comparison_unit_count": 0,
          "invalid_comparison_unit_count": 0
        }
      ]
    }
  },
  "path": "/benchmarks/ai-static-ad-generation/",
  "sourcePublication": "abfb9707d19b695fb05c1694d46222ca",
  "datePublished": "2026-10-16T08:22:38.946Z",
  "reviewedAt": "2026-10-16T08:22:38.946Z",
  "methodology": {
    "label": "AI static ad generation",
    "summary": "Image models create ads from five fixed briefs, using identical product-reference material in both prompt conditions. Ben scores the unedited images on fidelity, copy, claims, compliance and craft.",
    "scoringLabel": "Human-rated image quality / 10",
    "scoreUnit": "score",
    "sections": [
      {
        "heading": "What the model receives",
        "paragraphs": [
          "The image test uses five briefs and one canonical product-reference contact sheet. Each brief produces one image with the simple prompt and one with the scaffold, giving ten outputs per complete run. The reference bytes are shared across models and conditions."
        ]
      },
      {
        "heading": "What changes between prompts",
        "paragraphs": [
          "Provider-neutral prompts keep the task and reference constant. The scaffold adds concise ad-anatomy guidance. Prompt length is bounded rather than silently truncated, so a model is assessed on the brief it actually received."
        ]
      },
      {
        "heading": "How the result is scored",
        "paragraphs": [
          "Raw outputs are human-scored for product fidelity, copy accuracy, claims safety, brief compliance and craft quality. Each dimension must reach 8 for an output to clear the usability threshold. The displayed mean is a comparison measure; it can conceal a weak individual dimension. No editing pass is added before scoring."
        ]
      }
    ]
  },
  "content": {
    "title": "Best AI models for static ad generation",
    "description": "Benmark compares image models on the same five ad briefs and product reference.",
    "intro": "Benmark compares image models on the same five ad briefs and product reference. The images are scored as generated, so the result shows how much useful creative the model delivers before an editing pass.",
    "findingTitle": "Start with fidelity as well as visual quality",
    "resultInterpretation": "The leading model produces the strongest reviewed static-image result. The comparison is useful because every model must work with the same product reference and brief, rather than presenting a hand-picked showcase image.",
    "sections": [
      {
        "heading": "A good-looking image can still be the wrong ad",
        "paragraphs": [
          "Product shape, packaging, copy and claims are part of the job. An image can have polished lighting and composition while getting the product wrong or changing the message. Benmark scores those dimensions separately alongside craft quality.",
          "That makes the input contract important. Give the model the actual reference it needs, specify the copy that must remain exact and make clear which claims are supported. Otherwise, a visually convincing invention can be difficult to spot at a glance."
        ]
      },
      {
        "heading": "The reference is part of the model comparison",
        "paragraphs": [
          "Both prompt conditions receive the same product-reference contact sheet. The scaffold adds concise ad-anatomy guidance. This lets the comparison ask what the model and prompt setup contribute while holding the visible product information constant.",
          "GPT Image 2.5 Sunburst improves from 8.44 to 8.84 in the reviewed paired scores across the five briefs. That makes the structured setup worth trying for this task, with product fidelity and exact copy kept in the review."
        ]
      },
      {
        "heading": "Use the benchmark to choose the first production pass",
        "paragraphs": [
          "Select a promising model, generate against a specific brief and review the result by dimension. If the product is accurate but the composition is weak, the next iteration should address composition. Regenerating everything without identifying the failure makes the process harder to improve.",
          "The published mean helps compare models; a finished asset still needs to clear each required dimension. Benmark's per-output usability rule requires every dimension to reach the threshold, so strong craft cannot cancel out a critical product or claims problem."
        ]
      }
    ],
    "related": [
      {
        "path": "/benchmarks/ai-video-ad-generation/",
        "label": "Comparing video generation"
      },
      {
        "path": "/benchmarks/ad-creative-review/",
        "label": "Reviewing finished creative"
      }
    ]
  }
}