Your text-to-image system generates biased images. How do you address it?
Computer VisionLevel: IntermediateTroubleshooting Scenarios
Your text-to-image system generates biased images. How do you address it?
Hints
- Identify the prediction unit: image, object, pixel or sequence.
- How do occlusion, lighting or domain changes affect results and evaluation?
Answer
- Bias audit: systematically test prompts like "a photo of a doctor" / "a photo of a nurse" across demographic variations. Measure demographic distribution of generated images.
- Prompt modification: add diversity guidance to default prompts — "diverse group of people," "various ethnicities and genders"
- CLIP-based filter + regeneration: use CLIP to check if generated image matches stereotyped representations; regenerate if bias threshold exceeded
- Fine-tuning on balanced data: fine-tune the model (LoRA + DreamBooth) on a balanced, curated dataset with diverse representation
- User controls: let users specify demographic attributes they want represented rather than relying on model defaults
Source and licence: JustInCache & contributors — Awesome AI Interviews · Apache-2.0