Despite the substantial progress in text-to-image (T2I) generation brought about by models such as DALL-E 3, Imagen 3, and Stable Diffusion 3, achieving consistent output quality — both in aesthetic and alignment terms — remains a persistent challenge. Although large-scale training provides general information, it does not provide high quality aesthetics and alignment. While supervised fine tuning (SFT), a post-training process, is critical to the success of the training program, its efficiency depends heavily on the quality and quantity of data that was used in the fine-tuning.
Public datasets that are currently being used by SFT target specific visual domains, such as anime and art. They also rely heavily on basic heuristic filtering of web-scale data. Curation by humans is time-consuming, expensive and does not scale. It also fails to produce the most effective samples. Recent T2I model use proprietary data with little transparency. This limits reproducibility and slows collective progress.
Approach: Model-Guided dataset curation
Yandex released an update to help mitigate the issues. Alchemist, A publicly-available, general-purpose SFT data set composed of carefully selected 3,350 image-text pair. Alchemist uses a novel method that relies on a diffusion model pre-trained to serve as an estimator of sample quality. The approach allows the choice of high-impact training data without depending on subjectivity or simple aesthetics.
Alchemist The release is intended to enhance the output quality by focusing on fine tuning. This release includes five fine-tuned Stable Diffusion Models that are publicly available. Both the dataset and model are available on Hugging Face Under an open licence. More about the methodology and experiments — in the preprint .
Filtering pipeline and dataset characteristics: Technical design
Alchemist’s construction involves a filtering pipeline with multiple stages, starting at a base of 10 Billion web images. It is organized as follows.
- The initial filtering: Removal of NSFW content and low-resolution images (threshold >1024×1024 pixels).
- Fine Quality FilteringClassifiers are used to eliminate images that have compression artifacts (such as motion blur), watermarks and other defects. The classifiers have been trained using standard datasets for image quality assessments such as KonIQ-10k or PIPAL.
- Deduplication and IQA Based PruningImages are further scored using the TOPIQ model, which ensures that only clean samples will be retained. Images are scored further using the TOPIQ Model, to ensure that clean samples remain.
- Selecting by DiffusionThe cross-attention activations of a diffusion pre-trained model are used to help rank the images. The scoring function is used to identify samples with features that are strongly associated with aesthetic appeal and style richness. It is possible to select samples which are likely enhance the performance of downstream models.
- Caption RewritingThis is done using an optimized vision-language model to create prompt style textual descriptions. The final selected images are re-captioned using a vision-language model that is finely tuned to produce prompt style textual descriptions.
The authors found that the quality of models is reduced when the data set size exceeds 3,350 samples (e.g. 7k or 19.k). This reinforces the importance of high-quality, targeted data.
Results Across Multiple T2I Models
Alchemist’s effectiveness was assessed across five Stable Diffusion variations: SD1.5 (Stable Diffusion 1.5), SD2.1 (Stable Diffusion 2.1), SDXL (Stable Diffusion XL), SD3.5 Medium and SD3.5 large. Three datasets were used to fine-tune each model: the Alchemist dataset (i), a size-matched portion of LAION-Aesthetics V2, (ii), and their respective baselines.
Human Evaluation: Expert annotators performed side-by-side assessments across four criteria — text-image relevance, aesthetic quality, image complexity, and fidelity. Alchemist-tuned models showed statistically significant improvements in aesthetic and complexity scores, often outperforming both baselines and LAION-Aesthetics-tuned versions by margins of 12–20%. Text-image relevancy remained constant, indicating that the prompt alignment wasn’t affected.
Automatic MetricsAlchemist-tuned model scored generally higher on metrics, such as FD DINOv2, CLIP score, ImageReward or HPS-v2, than the baseline models. Noteworthy, the improvements in comparison to LAION models based on size-matched data were consistent.
The Dataset Size AblationThe performance of Alchemist was lower when fine-tuning the algorithm with large datasets (7k or 19k samples). This shows that a stricter filtering process and a higher quality per sample are more important than dataset sizes.
Yandex plans to keep using the dataset in its future model upgrades.
The conclusion of the article is:
Alchemist The approach is based on a high-quality sample over a large dataset and uses a reproducible methodology to build the dataset without using proprietary software.
The framework highlights that while improvements in aesthetics, image complexity and other perceptual attributes are notable, there can be trade-offs in terms of fidelity. This is especially true for the newer models which have already been optimized by internal SFT. Alchemist, however, establishes an industry standard for SFT datasets that are suitable for general use. This is valuable for both researchers and developers who want to improve the output of generative model.
Click here to find out more Paper here The following are some examples of how to get started: Alchemist Dataset on Hugging Face. Thanks to the Yandex team for the thought leadership/ Resources for this article.
Asif Razzaq, CEO of Marktechpost Media Inc. is a visionary engineer and entrepreneur who is dedicated to harnessing Artificial Intelligence’s potential for the social good. Marktechpost was his most recent venture. This platform, a Media Platform for Artificial Intelligence, is renowned for its comprehensive coverage of news on machine learning and deep-learning that is both technical and understandable to a broad audience. This platform has over 2,000,000 monthly views which shows its popularity.


