ContextBias: Controlled Evaluation of Bias Persistence Under Context Shift in Text-to-Image Models
Shaghayegh Kolli, Sina Emami, Moreno D'Incà, Pouyan Nejadi, Nicu Sebe, Massimiliano Mancini, Jana Diesner
EMNLP 2026
cs.CV, cs.CL
2026-08-30
TUM and Trento test four T2I models on 66,240 images across 92 jobs: an unrelated scene does not suppress role-linked attributes; pooled BI rises from 0.452 to 0.499.
Text-to-image models carry visual templates for jobs: doctors in white coats, mechanics drawn as men, dancers drawn as women. Most bias audits prompt the role in isolation ("a photo of a doctor") and tally gender, clothing, and props. Real prompts are messier. A doctor might be examining a patient, or jogging in a park. When the scene changes, do the role-linked cues move with it, or stay glued to the person?
That is a compositional-generalization question. The model should bind occupation and scene as separable parts, not stamp the strongest training prototype onto every mention of the job. Prior bias work rarely varies location and activity while holding the role fixed, so persistence under context shift was mostly unmeasured.
Teams at TUM, the University of Trento, and Orreco built ContextBias, a controlled evaluation, and ContextBench, the matching benchmark. Roles come from the U.S. Bureau of Labor Statistics SOC taxonomy, manually filtered to 92 canonical titles. For each role, GPT-4o-mini proposed related and unrelated location/activity banks; humans then dropped rare, implausible, or stereotype-baiting entries, plus contexts that lean on culture-specific symbols. The resulting bank is English and Western-leaning, not a culture-neutral sample.
Prompts sit in three bins, changing only place and activity:
Each base prompt gets two semantic paraphrases; related and unrelated bins also swap locations and activities. That yields 18 configurations per role and 1,656 prompts in total. Four generators: FLUX.1, Stable Diffusion XL, Stable Diffusion 3.5, Qwen-Image. Ten images per prompt per model, 16,560 images per generator, 66,240 in all.
A 30-dimension schema covers four cohorts: Scene, Camera, Objects, People. Most dimensions are closed-set. Items, clothing garment, and activities are open-vocabulary. GPT-5-mini fills the schema per image; open labels are clustered with sentence embeddings and canonicalized by GPT-4o-mini. Insufficient evidence becomes unknown and drops out of frequency counts.
Two scores. Bias Intensity (BI) is how peaked a label distribution is: one minus normalized Shannon entropy, 0 for uniform, 1 for a single label. It does not name the winner, only the squeeze. They report a pooled version (BIpool) and a per-role average (BIrole). Context Consistency Score (CCS) tracks a specific label that is both common and stable across the three bins: prevalence divided by (1 + cross-condition percentage-point range). A chi-squared homogeneity test sits beside it; a label counts as context-invariant only when p>0.05 and the range is at most 5 percentage points.
If models composed roles with scenes, unrelated context should suppress role-linked cues and lower concentration. The pooled numbers go the other way.
| Aggregation | CF | CA-U | Delta |
| BIpool overall | 0.452 | 0.499 | +0.047 |
| BIrole overall | 0.724 | 0.694 | −0.031 |
| Scene (pool) | 0.430 | 0.524 | +0.093 |
| Objects (pool) | 0.251 | 0.309 | +0.058 |
| People (pool) | 0.457 | 0.499 | +0.042 |
| Camera (pool) | 0.668 | 0.665 | −0.003 |
A role-level cluster bootstrap (B=1,000) puts the overall BIpool rise at +0.045 (95% CI [0.030, 0.050]). The two aggregations disagree on purpose: unrelated scenes do not sharpen a given role's prototype, and they do not wash it out. They pull different roles toward a shared default, so the benchmark-wide distribution tightens while each role's own distribution does not.
Person-level labels barely move. Dancer is female in 98%–100% of images across all four models (CCS 79.9); flight attendant 99%–100% female; nurse 99% female; comedian 99%–100% male; mechanic 94%–100% male. Garments and tools are rarer but still sticky: pharmacist/scientist coats at 31%–50%, baker aprons 31%–33%, welder gloves 29%–32%. Expression and accessories score low CCS, so the framework can tell locked cues from movable ones.
Paraphrase and substitution do not break the bindings. Across models, 93.3% of role–label tuples stay invariant under leave-one-out tests (People 92.1%, Objects 95.0%). Camera is the soft spot at 64.9% invariance; framing and depth of field are what actually track the prompt.
On a 2,200-image-per-model check set where the attribute is written into the prompt, extraction averages 90.2% accuracy and 94.8 F1 (FLUX.1 lowest at 86.4%/92.7, SDXL highest at 92.1%/95.9). Three annotators labeled 1,200 images: pairwise Cohen's κ 0.912, agreement with the pipeline 0.822.
Fairness audits that only prompt "a photo of a doctor" miss the layer this paper measures. Changing the scene mostly restyles the background and the camera. The person stays on the occupational template. If a product team hopes that dropping people into everyday settings will dilute stereotypes, these numbers do not back that hope.
This is an evaluation kit, not a new debiasing method. ContextBench and the code are to be released. For teams inspecting occupational depiction in their own generators, the three-bin contrast is more informative than another round of context-free gender ratios. Four architecturally different models land in the same direction, which points to shared training-data prototypes more than a sampler quirk at one lab.
High BI is not the same as harm. The paper says so in the limitations: a white coat on a doctor can be a visual regularity of the job, and BI cannot split regularity from stereotype. CCS measures whether a cue survives a scene change. It does not decide whether the cue is objectionable.
Only location and activity vary. Lighting, cultural setting, and interpersonal interaction are out of scope. The 92 roles are U.S. SOC titles in English, so the regularities are Western-benchmark patterns, not universal ones. Attribute extraction runs through GPT-5-mini, which may leak its own occupational priors into the labels. Human κ of 0.822 bounds some error; it does not bound systematic bias. Open-vocab canonicalization can dump edge labels into coarse buckets such as "coat" or "shirt", so garment prevalences around 30% partly reflect merge granularity, not only what the generator drew.
The extraction check set uses explicit prompts ("a photo of a male bartender"), which is easier than the main study, where the attribute is implied. Main-study reliability leans on 1,200 human-labeled images covering 10 roles, not all 92. Unrelated contexts were LLM-proposed then hand-filtered, and that filter itself defines what "unrelated" means. There is no before/after comparison against a debiasing method, so the paper cannot say whether existing interventions would cut CCS.