DST-100K: A 100K Triplet Dataset for Artistic Style Transfer

• DST-100K is a dataset comprising 100K triplets that reverse style transfer by using destylization to recover natural, style-free images from artworks.
• It employs a diffusion model and FLUX-Dev’s architecture to generate high-quality de-stylized images, ensuring content fidelity and style accuracy.
• A multi-stage DST-Filter refines pairs by enforcing strict content and style thresholds, enabling its effective use in training OmniStyle2.

DST-100K is a 100 K–triplet dataset for artistic style transfer introduced in "OmniStyle2: Scalable and High Quality Artistic Style Transfer Data Generation via Destylization" (Wang et al., 7 Sep 2025). It is constructed by reframing style transfer as a data problem through destylization: reversing style transfer by removing stylistic elements from artworks to recover natural, style-free counterparts. The resulting supervision signal aligns real artistic styles with their underlying content, yielding triplets of the form $$\langle \text{style},\,\text{reference},\,\text{de-stylized}\rangle$$. Within the OmniStyle2 framework, DST-100K is presented as a response to the lack of ground-truth data in artistic style transfer and as the training substrate for a simple feed-forward model based on FLUX.1-dev.

1. Destylization as the organizing principle

The central idea behind DST-100K is destylization, defined as reversing style transfer by removing stylistic elements from artworks to recover natural, style-free counterparts (Wang et al., 7 Sep 2025). In the source formulation, this is not treated as a post hoc cleanup step, but as the mechanism by which supervision is created. A stylized image is paired with a reconstructed style-free image, and the pair is retained only if content fidelity and style removal satisfy explicit thresholds.

This framing is significant because the paper identifies a fundamental challenge in artistic style transfer: the lack of ground-truth data. DST-100K addresses that challenge by constructing supervision rather than assuming it. A plausible implication is that the dataset is intended to substitute for unavailable paired data between natural scenes and their artistic renditions.

DST-100K is not merely a collection of stylized images. Its defining unit is the triplet $\langle I_s, I_{j^*}, \hat I_c\rangle$, where $I_s$ is the style image, $\hat I_c$ is the de-stylized image produced by the destylization model, and $I_{j^*}$ is a reference image selected from the same style category according to the CSD metric. This triplet structure is central both to dataset construction and to downstream training.

2. Construction pipeline: DST

DST-100K is built using DST, a text-guided destylization model whose architecture is based on FLUX-Dev’s diffusion backbone (DiT) (Wang et al., 7 Sep 2025). DST takes as input a stylized image $I_s$ and a content caption $c$, where the caption is obtained from InternVL2.5-7B. The model uses a pretrained VAE encoder $\mathrm{Enc}_{\mathrm{VAE}}$ to extract feature maps

$z_s=\mathrm{Enc}_{\mathrm{VAE}}(I_s),$

and a text encoder $\mathrm{Enc}_{\mathrm{txt}}$ to extract

$\langle I_s, I_{j^*}, \hat I_c\rangle$0

The architecture further includes Gaussian noise injection on latent content features $\langle I_s, I_{j^*}, \hat I_c\rangle$1, spatial concatenation $\langle I_s, I_{j^*}, \hat I_c\rangle$2 as conditioning tokens, a diffusion Transformer that predicts denoised latents, and a VAE decoder $\langle I_s, I_{j^*}, \hat I_c\rangle$3 to reconstruct the de-stylized image $\langle I_s, I_{j^*}, \hat I_c\rangle$4. In this formulation, the model explicitly combines visual evidence from the stylized input with textual evidence describing the “unstyled” scene.

DST is trained as a conditional diffusion model with additional reconstruction, perceptual, and adversarial terms. The overall loss is given as

$\langle I_s, I_{j^*}, \hat I_c\rangle$5

The diffusion loss is

$\langle I_s, I_{j^*}, \hat I_c\rangle$6

The reconstruction loss is

$\langle I_s, I_{j^*}, \hat I_c\rangle$7

The perceptual loss is

$\langle I_s, I_{j^*}, \hat I_c\rangle$8

where $\langle I_s, I_{j^*}, \hat I_c\rangle$9 are intermediate layers of a pretrained VGG network.

The adversarial loss is

$I_s$0

These objectives indicate that DST is optimized not only for denoising consistency but also for pixel-level reconstruction, feature-level similarity, and realism. This suggests that the de-stylized outputs are intended to be usable as training targets rather than merely approximate inversions.

3. Quality control: DST-Filter and triplet formation

After destylization, DST-100K is refined by DST-Filter, a multi-stage evaluation model that automatically discards low-quality pairs while ensuring content fidelity and style accuracy (Wang et al., 7 Sep 2025). The judgment is decomposed into two parts, each implemented via Chain-of-Thought prompts on GPT-4o: content preservation and style discrepancy.

For content preservation, the procedure is: (1) region/object identification in the style image $I_s$1; (2) for each region, checking structure, visibility, and detail fidelity in $I_s$2; and (3) assigning a content score

$I_s$3

equal to the minimum region score. The use of the minimum region score means that a local failure can determine the final content judgment.

For style discrepancy, the procedure is: (1) decomposing style attributes such as color palette, texture, lighting, and rendering effects from $I_s$4; (2) checking which of those attributes survive in $I_s$5; and (3) assigning a style-difference score

$I_s$6

Pairs are retained only when both thresholds are satisfied:

$I_s$7

Once filtering is completed, the dataset forms final triplets by reference image selection within the same style category. For each style image $I_s$8, a second image is chosen that is most similar under the CSD metric:

$I_s$9

This yields the final triplet

$\hat I_c$0

A common misconception in paired-data construction for style transfer is that any inverse rendering or caption-based reconstruction is sufficient. The DST-100K pipeline does not adopt that assumption. Instead, it explicitly requires both content preservation and style discrepancy, which suggests that inverse quality alone is not treated as adequate unless stylistic residue has also been removed.

4. Composition and scale

DST-100K contains 100 000 triplets of the form $\hat I_c$1 (Wang et al., 7 Sep 2025). Its style pool combines real artistic images and synthetic style images, and its captions are generated to describe the unstyled scene.

Component	Specification
Total triplets	100 000
Real artistic images	10 000 images from WikiArt & National Gallery
Real-art coverage	669 artists, 117 movements
Synthetic style images	150 000 images across 65 digital style categories
Caption source	GPT-4o
Final reference alignment	Peer with highest CSD within the same style category

The real artistic portion consists of 10 000 real artistic images from WikiArt and National Gallery, covering 669 artists and 117 movements. The synthetic portion consists of 150 000 synthetic style images across 65 digital style categories, including line-art, watercolor, and low-poly. Content captions are generated by GPT-4o to describe the “unstyled” scene. After DST destylization and DST-Filter, 100 K high-quality style–de-stylized pairs remain. Reference alignment then selects, within each style category, the peer with highest CSD to serve as the multi-reference input.

The mixture of real and synthetic styles is notable because it combines historically grounded artworks with digitally defined categories. This suggests that DST-100K is intended to span both traditional art-historical variation and contemporary stylization regimes used in image editing systems.

5. Evaluation metrics and integrity criteria

DST-100K data integrity and OmniStyle2 performance are evaluated along content fidelity, style accuracy, aesthetic quality, and human evaluation axes (Wang et al., 7 Sep 2025). The metrics listed in the source are used both to characterize the dataset and to assess the model trained on it.

For content fidelity, the paper reports DINO-Score, CLIP-Score, and Qwen-Content-Score. DINO-Score is defined as

$\hat I_c$2

CLIP-Score is defined as

$\hat I_c$3

Qwen-Content-Score is a 0–10 image–image content similarity score via Qwen-VL-Max.

For style accuracy, the metrics are CSD-Score, Style Loss, and Qwen-Style-Score. CSD-Score is described as Cross-Similarity Distance, where higher means more style fidelity. Style Loss is the Gram-matrix distance

$\hat I_c$4

Qwen-Style-Score is a 0–10 score via Qwen-VL-Max. For aesthetic quality, the metric is Qwen-Aesthetic-Score, also on a 0–10 scale.

Human evaluation is reported as rank-based preference over 30 participants and 810 votes. The reported metrics are Rank-1 preference and Top-3 preference. In the user study, OmniStyle2 achieves 28.21% Rank-1 and 58.95% Top-3.

These criteria show that dataset quality is not framed solely as low-level reconstruction fidelity. Instead, the evaluation space jointly measures semantic preservation, stylistic faithfulness, and perceived overall quality.

6. Function in OmniStyle2 and reported outcomes

DST-100K serves as the training data for OmniStyle2, a simple feed-forward model based on FLUX.1-dev (Wang et al., 7 Sep 2025). During training, each batch contains

$\hat I_c$5

and no text input is used, represented as an empty string. The conditioning uses $\hat I_c$6 and $\hat I_c$7 as conditional tokens, while $\hat I_c$8 is the denoising target.

The learning objective for OmniStyle2 is pure diffusion denoising, using $\hat I_c$9, described as the same as DST’s diffusion loss but with the style image as target. Architectural additions include sequential positional encoding to disambiguate “reference” versus “content” tokens, and LoRA fine-tuning on the DiT with no full-model updates. The training schedule uses 8×A800 GPUs, learning rate $I_{j^*}$0, batch size 48, and data augmentation by random horizontal and vertical flips.

Quantitatively, Table 1 compares OmniStyle2 with seven state-of-the-art style transfer models. The reported scores are best CSD-Score $I_{j^*}$1, best style loss $I_{j^*}$2, best Qwen-Style $I_{j^*}$3, and best Qwen-Aesthetic $I_{j^*}$4. Reported content scores are DINO $I_{j^*}$5, CLIP $I_{j^*}$6, and Qwen-Content $I_{j^*}$7. Table 2 compares against closed- and open-source editing models and states that OmniStyle2 is second only to GPT-4o on many metrics, but without content leakage.

Qualitative comparisons report that OmniStyle2 produces clean cartoon abstractions, preserves facial regions, avoids background bleed, and replicates 3D origami and watercolor styles more faithfully than all baselines. In the paper’s interpretation, DST-100K underpins this balance of content fidelity and style accuracy.

The connection between dataset design and model behavior is direct: DST-100K supplies de-stylized content targets and within-category style references, while OmniStyle2 consumes exactly those signals. This suggests that the dataset is not an auxiliary resource but the core supervisory mechanism of the overall system.