The LoRA file wasn't handed in, so there's nothing to download for this project yet.
The question
If you fine-tune a model on one culture's interiors, does it stop defaulting to its own assumptions? Frank trained a LoRA on 1970s Soviet living rooms to test how much Western bias would remain.
What they did
- Collected photographs of Soviet apartment interiors from the period and captioned them with a trigger phrase.
- Wrote twelve detailed room prompts (living room, kitchen, hallway, child's room, study, a room in winter) and generated one image for each.
- Reviewed the results with the review group, including someone who had lived in Russia.
What they found
- The images still showed strong Western influence. The model kept adding American-style objects, especially televisions, some of them not even from the 1970s.
- It leaned into stereotypes, often producing cramped, inefficient rooms that reflected general assumptions rather than real homes.
- The base model's original training kept showing through after fine-tuning. AI doesn't fully unlearn its biases, which matters wherever accuracy and cultural authenticity count.
























