Yandex has introduced Alchemist, a supervised fine-tuning dataset aimed at improving the performance of text-to-image models. With just 3,350 high-quality image-text pairs, the dataset delivers significant performance gains across several benchmarks.
It is available as an open-source release on Hugging Face: huggingface.co/datasets/yandex/alchemist.
Built With Diffusion, Not Heuristics
Alchemist uses a novel curation method that relies on a pre-trained diffusion model to select impactful training samples. Instead of manual labeling or heuristic filtering, Yandex applied the diffusion model to evaluate visual richness among 10 billion web-sourced images. After removing low-quality and duplicate data, the top-scoring samples were selected.
This automated, model-guided method aims to eliminate human bias while focusing on data most likely to improve generative performance.
Stronger Visual Output With Fewer Samples
In testing across five Stable Diffusion variants, models fine-tuned on Alchemist outperformed those trained on both baseline data and LAION-Aesthetics-tuned versions. According to Yandex, aesthetic quality and image complexity improved by 12–20%, while text-image relevance remained stable.
Surprisingly, increasing the dataset beyond 3,350 samples led to a drop in performance. This result underlines the value of targeted curation and highlights the diminishing returns of scaling without strict quality control.
Already In Use, But Questions Remain
Alchemist is already being used to train YandexART v2.5, the company’s latest generative model. Alongside the dataset, Yandex has also released the fine-tuned weights of the models, supporting broader community access and reproducibility.
Despite its promising results, Alchemist raises some open questions. Its small size, while effective, may limit its ability to generalize across all use cases. Potential issues around diversity, representation, and hidden bias are acknowledged in the accompanying research but require further exploration.
A Step Toward Transparent Fine-Tuning
Yandex researchers—including Valerii Startsev, Alexander Ustyuzhanin, and Alexey Kirillov—argue that high-quality, general-purpose fine-tuning datasets are still rare. Many state-of-the-art models use proprietary or undocumented internal data, making progress harder to replicate. Alchemist’s release offers an alternative path—compact, transparent, and measurable.
