Evaluating Demographic Misrepresentation in Image-to-Image Portrait Editing

This paper investigates and quantifies demographic bias in instruction-guided image-to-image portrait editing, revealing that identical prompts cause systematic identity preservation failures and stereotype reinforcement across different demographic groups, which can be mitigated through prompt-level identity constraints without model retraining.

The Gist

Imagine you have a magical photo-editing app. You upload a picture of yourself, type in "Make me look like a CEO," and hit enter. You expect the app to put you in a suit and give you a confident pose, but keep you—your face, your skin tone, your features—exactly the same.

This paper is a report card on how well these "magic apps" actually do that job. The researchers found that while the apps are great at following instructions, they have a hidden, unfair habit: they treat people differently based on who they look like.

Here is the breakdown of what they discovered, using some everyday analogies.

1. The Two Ways the Magic Goes Wrong

The researchers found that when you ask the app to edit a photo, it fails in two specific ways, especially for people of color, women, or older adults.

• The "Ghost Edit" (Soft Erasure):
  Imagine you ask a chef to add extra salt to your soup. The chef brings the soup back, but it tastes exactly the same as before. They didn't refuse to do it; they just quietly ignored your request.

  • In the app: You ask to "show the person in a wheelchair," but the app just gives you the same standing photo. It pretends to edit, but the requested change is silently erased.

• The "Hollywood Remake" (Stereotype Replacement):
  Imagine you ask a director to cast a specific actor in a movie. The director agrees, but then swaps the actor for a different one who fits a "typical" stereotype of that role.

  • In the app: You ask to make a Black woman look like a "CEO." The app puts her in a suit, but it also lightens her skin, straightens her hair, and changes her facial features to look more like a white person. It didn't just edit the clothes; it edited her identity to fit a stereotype.

2. The "Default Setting" Problem

The study tested three different popular photo-editing models. They used a massive library of 84 different faces (covering all races, genders, and ages) and asked them to do 20 different things (like "look like a doctor" or "look older").

The Big Discovery:
The apps have a "Default Setting" that assumes the "standard" human looks white, young, and male.

• If you are a white person, the app mostly keeps you looking like you.
• If you are a Black, Indian, or Latino person, the app often tries to "fix" you by making your skin lighter and your features more "European."

It's like a photocopier that has a "Whiteness" filter turned on by default. If you put a dark photo in, the machine automatically tries to make it lighter, even if you didn't ask for that.

3. The "Occupation Trap"

The researchers also tested how the apps handle jobs.

• The Test: They asked the app to show a "CEO" or a "Nurse" without specifying a gender.
• The Result: Even if they started with a photo of a woman, the app often turned the "CEO" into a man. Even if they started with a man, the app turned the "Nurse" into a woman.
• The Metaphor: The app is like a person who has only ever seen movies where CEOs are men and nurses are women. When you ask for a CEO, their brain automatically pictures a man, regardless of the photo you gave them.

4. The "Magic Spell" Fix (Prompt-Level Control)

The researchers didn't just point out the problem; they tried a clever workaround. They realized they couldn't easily re-train the massive AI models (which is like trying to rebuild the engine of a car while driving it).

Instead, they tried adding a "Magic Spell" to the instructions.

• Normal Instruction: "Make this person look like a CEO."
• Magic Spell Instruction: "Make this person look like a CEO, BUT keep their deep brown skin, round face, and braided hair exactly the same."

Did it work?
Yes! For people of color, adding this detailed description of their physical features stopped the app from changing their race or skin tone.

• The Catch: It worked much better for non-white people than for white people. This proved that the app's "Default Setting" was indeed biased toward whiteness. When you explicitly told the app "Don't change the skin tone," it stopped trying to make everyone look white.

5. Why This Matters

The authors used a "Vision-Language Model" (a super-smart AI that can look at pictures and grade them) to check thousands of images. They found that these AI editors are currently unfair.

• Trust: If you use these tools for your profile picture or advertising, you can't trust that the result will actually look like you.
• Harm: These tools reinforce harmful stereotypes (e.g., "Doctors look like white men," "Black people shouldn't be in wheelchairs").

The Bottom Line

Current AI photo editors are like a well-meaning but biased assistant. They are eager to help, but they have a hidden bias that tries to make everyone look like a "standard" white person.

The good news is that we can fix this right now without rebuilding the AI. We just need to be very specific in our instructions, telling the AI exactly what features to keep. However, the researchers argue that the real responsibility lies with the companies building these tools: they need to build the "fairness" into the engine itself, so users don't have to constantly fight the machine to keep their own identity.

Technical Summary

1. Problem Statement

While demographic bias in Text-to-Image (T2I) generation is well-documented, Instruction-guided Image-to-Image (I2I) editing remains underexplored. Users expect I2I editors to modify specific attributes (e.g., clothing, background) while strictly preserving the subject's identity. However, the authors hypothesize that identical edit instructions yield systematically different outcomes based on the subject's demographics (race, gender, age).

The paper identifies that current open-weight I2I editors suffer from two distinct failure modes when processing portraits of minority groups:

1. Soft Erasure: The requested edit is silently ignored or weakly realized, resulting in an output that fails to comply with the instruction without explicitly refusing.
2. Stereotype Replacement: The model introduces unrequested, stereotype-consistent demographic attributes (e.g., lightening skin tone, changing race, or altering gender presentation) that override the source image's identity.

2. Methodology

A. Controlled Benchmark Construction

The authors created a rigorous benchmark to isolate demographic-conditioned failures:

• Source Data: 84 portraits sampled factorially from the FairFace dataset, covering 7 races, 2 genders, and 6 age groups.
• Diagnostic Prompts: A set of 20 prompts designed to expose specific biases, categorized into:
  • Occupational Stereotypes: Probing role-induced bias (e.g., "CEO," "Housekeeper," "Nurse").
  • Vulnerability Attributes: Probing edit suppression in sensitive contexts (e.g., "Wheelchair user," "Aging," "Disability").
• Models Evaluated: Three state-of-the-art open-weight I2I editors: FLUX.2-dev, Step1X-Edit-v1p2, and Qwen-Image-Edit-2511.
• Scale: 5,040 edited images generated (84 images × 20 prompts × 3 models).

B. Evaluation Protocol

The study employs a dual-evaluation strategy to ensure reliability:

1. VLM-Based Scoring: Two independent Vision-Language Models (Gemini 3.0 Flash and GPT-5-mini) score outputs on a 1–5 Likert scale across five axes:
   • Edit Success (detecting Soft Erasure).
   • Skin Tone, Race, Gender, and Age Change (detecting Stereotype Replacement).
2. Human Evaluation: 1,000 samples were annotated by 30 human raters on Prolific to validate VLM scores. The study found strong alignment between VLM and human judgments, validating VLMs as a scalable alternative for bias detection.

C. Mitigation Strategy: Feature Prompts

To test if failures can be mitigated without retraining models, the authors introduced a prompt-level identity constraint.

• Mechanism: A "Feature Prompt" ($p_{feat}$) is generated by a VLM describing observable physical attributes of the source image (skin tone, facial structure, hair, etc.) using descriptive language rather than categorical labels.
• Application: This prompt is prepended to the original edit instruction ($p_{feat} + p$).
• Goal: To act as a regularizer that forces the model to preserve specific visual features, thereby reducing demographic drift.

D. Supplementary Experiment

A WinoBias-based study was conducted to isolate gender-occupation stereotyping. By applying occupation prompts (e.g., "CEO" vs. "Nurse") to both male and female source images without specifying gender, the authors measured the rate at which models forced the subject's gender to align with occupational stereotypes.

3. Key Contributions

1. Formalization of Failure Modes: The paper explicitly defines and measures Soft Erasure (silent non-compliance) and Stereotype Replacement (unrequested identity drift), distinguishing them from general generation failures.
2. Demographic Benchmark: Introduction of a controlled benchmark yielding 5,040 edited images to systematically probe I2I behavior across race, gender, and age.
3. Prompt-Level Mitigation: Demonstration that adding observable appearance constraints to prompts can significantly reduce identity drift for minority groups without modifying model weights.
4. Validation of Automated Evaluation: Evidence of strong alignment between VLM-based scoring and human evaluation, offering a cost-effective method for future bias auditing.

4. Key Results

A. Pervasive Soft Erasure

• Step1X-Edit showed the lowest edit success rates, frequently ignoring requests entirely.
• FLUX.2-dev achieved high edit success but exhibited the most severe identity drift.

B. Systematic Stereotype Replacement (Racial Disparity)

• Skin Lightening: 62–71% of all edited outputs exhibited lighter skin tones than the source.
  • Indian and Black subjects experienced 72–75% skin lightening.
  • White and East Asian subjects experienced 44% and 54% respectively.
• Race Drift: Indian subjects experienced a 14% race change rate, compared to only 1% for White subjects.
• Conclusion: Models possess a "Default to White" prior; when identity constraints are underspecified, outputs regress toward White-presenting features.

C. Asymmetric Mitigation

• The Feature Prompt intervention significantly reduced race change for non-White groups (e.g., a 1.48-point reduction for Black subjects) but had negligible effect on White subjects (0.06-point reduction).
• This confirms that the baseline failure is asymmetric: non-White subjects suffer from larger deviations that are correctable via prompt constraints, whereas White subjects are already close to the model's default output space.

D. Gender-Occupation Bias

• In the WinoBias experiment, models followed occupational stereotypes in 84–86% of cases.
• Regardless of the source image's gender, models consistently altered the subject's presentation to match the gender stereotypical of the requested occupation (e.g., making a female subject appear male when prompted as a "CEO").

5. Significance and Implications

• Identity Preservation is Uneven: The study proves that "identity preservation" is not a uniform capability in current I2I editors; it fails disproportionately for minority groups.
• Architectural Priors: The findings suggest that demographic biases are deeply embedded in the latent priors of diffusion architectures, not just prompt engineering issues.
• User Burden vs. Developer Responsibility: While prompt-level constraints can mitigate some bias, the paper argues this places an unfair burden on users to manually specify attributes that should be preserved by default.
• Future Directions: The work calls for:
  1. Auditing training data for demographic balance.
  2. Developing evaluation protocols that prioritize identity stability.
  3. Architectural changes to ensure demographic attributes are preserved by default, rather than relying on user-side prompt engineering.

In summary, this paper provides the first systematic evidence that instruction-guided portrait editing is not neutral. It reveals that current models systematically erase or alter the identities of minority subjects, often replacing them with stereotypical or majority-group features, and offers a preliminary, prompt-based method to partially mitigate these harms.