SkinBench finds false eye-puffiness flags in 95% to 100% of images

By Tom Whitfield

Six leading AI models reported eye puffiness in approximately 95% to 100% of images where puffiness was absent, according to SkinBench, a public, vendor-run benchmark Haut.AI released on October 27, 2026. The study tested six models across 19 tasks involving seven visible skin characteristics, using expert-prepared reference assessments as the comparator.

SkinBench covered wrinkles, sun spots, redness, skin tone, skin type, dark circles and eye puffiness. Eleven tasks tested detection and grading, while eight measured how accurately a model located a feature.

The benchmark assessed Gemini 3.5 Flash from Google, GPT-5.6 Sol from OpenAI, Claude Fable 5 from Anthropic, Grok 4.6 from xAI, Qwen3.8 27B from Alibaba and MedGemma 27B, Google’s medically specialized multimodal model.

Why can one accuracy score mislead beauty brands?

Consumer use is already substantial. NielsenIQ found that 49% of consumers have taken beauty product recommendations from generative AI, while Accenture found that 75% of beauty consumers are open to AI-powered shopping. A Haut.AI survey of 1,238 U.S. women found that 62% use or are interested in AI for skincare or bodycare advice.

In April 2025, Gartner predicted that organizations would use smaller, specialized AI models at least three times as often as general-purpose large language models by 2027. SkinBench tests whether either model class can measure skin rather than merely describe a photograph.

Skin analysis requires separate judgments about concern presence, severity, facial location and skin type across complexions. A blended score can conceal failure in the task a product line depends on, such as detecting puffiness for an eye cream, locating wrinkles for longitudinal measurement or classifying skin for moisturizer selection.

False detections carry a direct commercial cost. Recommendation engines map each reported concern to a product, so an invented concern can produce an unnecessary purchase and undermine trust in the whole service. Progress tracking also breaks when a nonexistent concern cannot improve.

What should an AI skin-analysis benchmark test?\n

Haut.AI’s proposed buyer framework separates six commercially relevant tasks from one averaged result.

• False detection rate: How often does a model report a concern that expert assessors did not find? A high detection rate means little if the model labels almost every image.
• Localization: Can the system trace a wrinkle or spot where an expert marked it? Product-result measurement depends on accurate placement, not detection alone.
• Skin-type classification: Report balanced accuracy for dry, oily and combination skin. Otherwise, a model can score well by always guessing the most common type.
• Weakest skin-tone group: Publish the lowest-performing group’s score. An average across tones can conceal a group where the system fails.
• Agreement with experts: Show system-to-expert agreement beside inter-expert agreement. Expert disagreement sets a realistic ceiling for automated grading.
• Repeatability: Analyze the same skin twice under identical conditions and report consistency. An Institut d’Expertise Clinique study in France reported intraclass correlation coefficients of 0.97 to 0.98 across five facial endpoints against a consensus panel; 1.0 would mean identical results every time.

Published methods matter as much as scores. Buyers need test images, reference assessments, scoring rules and the model list, Haut.AI said. Its Clinical Studies Software provides an example of applying repeat skin measurement in large studies.

How did the models perform?

The false eye-puffiness result exposed a detection failure rather than a minor calibration issue. In separate localization examples, one model marked the nose and mouth when asked to trace frown lines, and another drew no lines on a test image.

When the same computer-generated face was shown five times against human frown-line traces, Haut.AI matched 0.99, compared with 0.29 for Gemini 3.5 Flash, 0.12 for GPT-5.6 Sol and 0.00 for Claude Fable 5. Claude placed the lines on the nose and mouth.

SkinBench reports results at three levels: four headline scores, seven parameter scores and detailed results for all 19 tasks. Every headline score runs from 0 to 100, with 100 indicating perfect agreement with experts on every image.

The Skin Intelligence Score averages a model’s seven parameter scores. Class prediction averages 11 grading tasks, including severity and skin type; segmentation averages eight tracing tasks, including tracing wrinkles or outlining dark circles; and inclusivity reflects the model’s worst-performing skin-tone group. Haut.AI entered SkinFM-powered models, and the leaderboard reports the best performer for each test.

On 350 computer-generated faces, Haut.AI scored 47.8 of 100 for frown-line tracing, against 20.2 for GPT-5.6 Sol, the strongest other model. Haut.AI detected deep forehead lines in 82% of images containing them, versus 15% to 60% for the other models. Its agreement with expert consensus on frown-line severity was 0.82 on a 0-to-1 scale, compared with 0.73 for GPT-5.6 Sol.

“A false detection isn’t just a technical mistake. It can lead to inappropriate product recommendations, unnecessary purchases and a loss of consumer trust,” said Anastasia Georgievskaya, CEO of Haut.AI.

Specialized skin AI did not win every task. Haut.AI ranked below another model on two of 19 tests: Gemini 3.5 Flash scored 71.7 for grading crow’s feet versus Haut.AI’s 61.8, and 90.4 for assessing skin tone versus Haut.AI’s 87.6. Those task-level gaps argue against relying on one average score.

Human experts still set the reference point. Individual graders reached 0.86 agreement with the consensus on frown-line severity, versus Haut.AI’s 0.82. The 350-face, computer-generated test therefore shows specialized AI nearing human consistency on one narrow task, not matching experts across skin analysis generally.

Haut.AI developed and ran SkinBench, and publishes its leaderboard, comparisons and methods at skinbench.ai. The company plans to add more models and test consistency when systems analyze the same concern repeatedly.

What does SkinFM change for beauty brands?

SkinFM is a single image-based model designed to analyze multiple skin characteristics rather than requiring a separate model for each task. Haut.AI launched it alongside SkinBench on October 27, 2026, and says it trained the model on more than 4 million real skin images.

The company says its facial analysis can assess more than 230 measured parameters and 48 biomarkers, including wrinkles, redness and pigmentation. Haut.AI also says the foundation model enables faster measurement development, brand-specific grading systems, more informed recommendations, tracking of visible improvement and support for research and product development.

Those capabilities remain claims that buyers should test task by task. Any new or customized measurement should face the same six criteria, with task-level results available for comparison, before it enters a recommendation engine.