Painted and Unpainted: The Hidden AI Bias That Changes Everything

Introduction

In July 2026, a groundbreaking article on Habr exposed a subtle but powerful bias in AI image generation models: the difference between "painted" and "unpainted" subjects. The research, conducted by a team of independent developers, revealed that AI models—particularly those trained on massive datasets like LAION-5B—systematically misinterpret visual cues related to surface coatings, leading to errors in fields ranging from e-commerce to autonomous driving. This article breaks down the findings, the real-world implications, and what practitioners need to know.

The original study, published on July 18, 2026, analyzed over 10,000 generated images across five popular AI models, including Stable Diffusion XL and DALL-E 3. The results were startling: when asked to generate images of "painted" objects (e.g., a painted wooden chair), models often produced unrealistic textures, missing brush strokes, or incorrect reflections. Conversely, "unpainted" objects (e.g., raw metal) were frequently rendered with unintended gloss or patina. Source

The Core Problem: Surface Coating Confusion

The article describes a fundamental issue in how AI models perceive and generate surface properties. The authors conducted a series of controlled experiments:

• Prompt variations: They used identical objects (e.g., "wooden table") but added modifiers like "painted red" vs. "unpainted oak."
• Model comparison: Five models were tested: Stable Diffusion XL, DALL-E 3, Midjourney 6, Adobe Firefly 2, and an open-source model (SDXL Turbo).
• Evaluation metrics: Human evaluators rated realism, texture accuracy, and color consistency on a 1-5 scale.

Key Findings

Source: Habr article, July 2026

The data shows a significant drop in quality when coating is specified. The control group—objects without any coating mention—performed much better. This suggests that the models have learned to associate "painted" or "unpainted" with specific, often erroneous, visual features.

Real-World Case Studies

The article provides two concrete examples of how this bias affects actual applications:

Case 1: E-commerce Product Photography

A mid-sized furniture retailer used AI to generate product images for their online catalog. They uploaded photos of unpainted oak chairs and used prompts like "painted white chair" to create variants. The results were disappointing: chairs appeared plastic-like, with unrealistic shine and missing grain details. After switching to manual retouching, conversion rates on those AI-generated images dropped by 23% compared to real photos.

The developers discovered that the model's training data overrepresented glossy, synthetic surfaces (e.g., car paint, plastic toys) and underrepresented matte, textured paints. This bias caused the model to default to a "plastic" look when "painted" was mentioned.

Case 2: Architectural Visualization

An architectural firm used AI to generate concept images for a client. They requested "unpainted concrete walls" for a modern loft design. The AI consistently produced walls that looked damp or stained, with visible efflorescence (white salt deposits). The client rejected the concepts, assuming poor material quality.

The team traced the issue to the model's training data: many images of "concrete" in the dataset were from construction sites or old buildings, where unpainted concrete often appears weathered. The model learned to associate "unpainted" with "damaged."

Technical Root Cause

The Habr article dives into the technical reasons behind this bias:

1. Dataset imbalance: The LAION-5B dataset contains disproportionately more images of painted objects (cars, furniture, toys) than unpainted ones (raw wood, stone, metal). This skews the model's understanding.
2. Annotation gaps: Captions rarely specify coating types. "Red chair" could mean painted, dyed, or coated, but the model learns from context—often from advertisements where chairs are glossy.
3. Architecture limitations: Current diffusion models struggle with fine-grained texture synthesis. The authors note that even state-of-the-art models like DALL-E 3 have a "texture bottleneck" where high-frequency details (brush strokes, grain) are lost during the denoising process.

Practical Mitigation Strategies

The article offers several actionable solutions for practitioners:

Prompt Engineering

• Be specific: Instead of "painted wall," use "matte latex painted wall with visible roller marks."
• Add negative prompts: Exclude terms like "glossy," "plastic," or "shiny" when generating unpainted surfaces.
• Use reference images: Some models support image-to-image conditioning. Provide a photo of the exact texture you want.

Post-Processing

• Apply texture overlays in tools like Photoshop or GIMP.
• Use dedicated texture synthesis models (e.g., TextureGAN) to refine AI outputs.
• Manually adjust specular highlights and roughness maps if using 3D renderers.

Dataset Curation

• For custom model training, include balanced samples of both painted and unpainted objects with detailed captions.
• Use datasets like Materials in Context (MINC) or Flickr Material Database.
• Annotate coating type explicitly (e.g., "oil-based paint," "water-based paint," "raw wood").

Industry Implications

The bias has broader consequences beyond image generation:

• Autonomous vehicles: Misidentifying road markings (painted) vs. tire marks (unpainted) could affect lane detection.
• Medical imaging: Painted surgical instruments vs. unpainted tissue surfaces in training data could lead to misclassification.
• Augmented reality: AR filters that add "painted" effects to real objects may fail if the model doesn't understand surface interaction.

Conclusion

The painted vs. unpainted bias is a clear example of how AI models inherit limitations from their training data. The Habr article's findings underscore the need for more nuanced datasets and better texture understanding in generative models. For practitioners, the key takeaway is to always validate AI outputs against real-world expectations—especially when surface properties matter.

As AI continues to permeate industries from design to manufacturing, addressing these subtle biases will be critical. The authors call for a community-wide effort to create high-quality, annotated datasets for material properties. Until then, manual oversight remains essential.

Read the full original research: Source