CRAFT: Cultural Russian-Oriented Dataset Adaptation for Focused Text-to-Image Generation

TL;DR

CRAFT addresses the Western bias in text-to-image models by constructing a Russian cultural code dataset and using it to fine-tune Kandinsky 3.1 for culturally focused generation. The approach combines entity-centric data collection across 17 Russian cultural categories, manual data processing and captioning, and a two-stage fine-tuning regime, yielding higher Russian-cultural awareness as assessed by human evaluators. The paper discusses ethical considerations, limitations due to data imbalance and translation, and outlines future directions including a RusCode benchmark. The work demonstrates that targeted cultural data can substantially improve generation fidelity for culturally specific prompts, with practical implications for culturally aware AI systems.

Abstract

Despite the fact that popular text-to-image generation models cope well with international and general cultural queries, they have a significant knowledge gap regarding individual cultures. This is due to the content of existing large training datasets collected on the Internet, which are predominantly based on Western European or American popular culture. Meanwhile, the lack of cultural adaptation of the model can lead to incorrect results, a decrease in the generation quality, and the spread of stereotypes and offensive content. In an effort to address this issue, we examine the concept of cultural code and recognize the critical importance of its understanding by modern image generation models, an issue that has not been sufficiently addressed in the research community to date. We propose the methodology for collecting and processing the data necessary to form a dataset based on the cultural code, in particular the Russian one. We explore how the collected data affects the quality of generations in the national domain and analyze the effectiveness of our approach using the Kandinsky 3.1 text-to-image model. Human evaluation results demonstrate an increase in the level of awareness of Russian culture in the model.

Figures (12)

• Figure 1: Comparison of Russian cultural code generations for popular text-to-image models. Reference is an example of a real image with a specific entity. The cultural adaptation procedure we propose helps improve the quality of cultural awareness for the Kandinsky 3.1 model, both in comparison with the previous version Kandinsky 2.2, and for other models.
• Figure 2: 17 main data categories for creating a dataset of the Russian cultural code. Examples for each category are generated by the Kandinsky 3.1 model. For each category, we collect a set of visual entities from which we form our dataset. As a result of additional training, the model increases its level of cultural awareness.
• Figure 3: General pipeline of our CRAFT method for cultural adaptation. We create a list of categories and entities from Russian cultural code based on our own cultural analysis, collect and process data manually, including captioning process. The resulting dataset is used for additional training of the text-to-image model Kandinsky 3.1 arkhipkin2024kandinsky30technicalreport to increase its level of cultural awareness.
• Figure 4: The effect of filtering and custom captioning on the generation quality in Russian culture domain. Reference image is a real image from the dataset that displays a specific entity. Our experiments showed that without additional data processing the model generates unsatisfactory results.
• Figure 5: Examples of images from the instructions that people followed when data filtering. The task was to select only those images that contain Russian and post-Soviet visual features.