OmniStyle2: Data-Driven Artistic Style Transfer

• OmniStyle2 is an artistic style transfer framework that uses destylization to extract style-free images from artworks, enabling authentic data generation.
• It constructs DST-100K, a 100K-triplet dataset, by filtering and pairing style images with their destylized counterparts to overcome the lack of ground-truth stylizations.
• The approach employs a feed-forward model based on FLUX-Dev with LoRA fine-tuning, achieving competitive performance in style fidelity and visual quality.

to=arxiv_search ՞ւjson {"query":"OmniStyle2 scalable and high quality artistic style transfer data generation via destylization", "max_results": 5} OmniStyle2 is an artistic style transfer framework that reframes stylization as a data generation problem through destylization, the inverse operation of removing stylistic elements from artworks to recover natural, style-free counterparts. In the formulation introduced in "OmniStyle2: Scalable and High Quality Artistic Style Transfer Data Generation via Destylization" (Wang et al., 7 Sep 2025), destylization is used to construct DST-100K, a large-scale triplet dataset that aligns real artistic styles with their underlying content, and the resulting dataset is then used to train a simple feed-forward model based on FLUX.1-dev. The central claim of the framework is that scalable, authentic supervision obtained from destylization addresses the longstanding absence of ground-truth paired data in artistic style transfer.

1. Problem setting and conceptual reformulation

Artistic style transfer seeks to re-render a content image in the style of a reference artwork. A persistent difficulty in this area is that there is no ground-truth stylization for a given content–style pair. Prior approaches therefore rely on hand-crafted feature or statistical losses, unstable optimization or inversion procedures, or synthetic pseudo-targets produced by existing stylization systems and filtered with multimodal LLMs. In the OmniStyle2 formulation, these pseudo-targets are treated as inherently capped by the quality of the stylization model that generated them.

The key reformulation is to invert the problem. Instead of attempting to learn from an unavailable forward target, OmniStyle2 performs destylization: given a real artistic image, it reconstructs a plausible natural, style-free image representing the same underlying content. This makes the original artwork usable as an authentic style target for supervised forward training. A plausible implication is that the method shifts the difficult question from “what should the stylized target look like?” to the more tractable question “what would this artwork depict if rendered naturally?”, which can be supervised synthetically and filtered automatically.

This perspective is explicitly data-centric rather than loss-centric. The system is organized around four named components: DST, DST-Filter, DST-100K, and OmniStyle2 itself (Wang et al., 7 Sep 2025).

2. Destylization and the DST model

Destylization is defined as mapping a stylized image $x_{\text{sty}}$ and a content caption $t$ to a style-free natural image $x_{\text{desty}}$ that approximates a natural counterpart $x_{\text{nat}}$. In DST, this is trained on synthetic pairs for which the original natural content image is known. Given a content image $x_{\text{cnt}}$, a stylized version $x_{\text{sty}}$, and a caption $t$ describing the content of $x_{\text{cnt}}$, the model learns

$$f_\theta(x_{\text{sty}}, t) \approx x_{\text{cnt}}.$$

DST is built on FLUX-Dev. Images are encoded by a VAE, captions are encoded by a text encoder, and the DiT denoiser is trained with a diffusion reconstruction objective. The training setup uses

$$z_{\text{cnt}} = \text{VAE}_{\text{enc}}(x_{\text{cnt}}), \quad z_{\text{sty}} = \text{VAE}_{\text{enc}}(x_{\text{sty}}), \quad c = e(t),$$

with noisy target latent

$t$0

The denoiser is trained by

$t$1

The DST architecture uses a VAE image encoder/decoder, a FLUX DiT backbone, a text encoder, and token concatenation of stylized image tokens, text tokens, and noisy target tokens. During inference, only the style image and its content caption are given; the model then samples a style-free output consistent with the caption.

The DST training corpus is itself synthetic and separate from DST-100K. It uses content images from HQ-50K and FFHQ, covering six categories—humans, objects, animals, plants, scenes, and architectures—with 200 representative images per category, plus 200 classical paintings (NGA) and 200 Style30K images as style references. Stylized images are produced using STROTSS, StyleID, CSGO, and Attention Distillation, while InternVL2.5-7B generates detailed content captions. The result is approximately 60K triplets of the form $t$2. DST training uses full DiT fine-tuning, 8 × A800 GPUs, learning rate $t$3, batch size 8, and horizontal and vertical flipping (Wang et al., 7 Sep 2025).

3. DST-Filter and the construction of DST-100K

DST-100K is a 100K-triplet dataset

$t$4

constructed from destylization outputs. Its purpose is to provide authentic supervision signals by pairing real or synthesized style images with corresponding style-free content counterparts.

The construction process begins with a large style image pool assembled from two sources. The first source is real artistic imagery from WikiArt and the National Gallery of Art, filtered to remove low-resolution, non-artistic, duplicate, or otherwise unsuitable images. InternVL2.5-7B is used to categorize images into Human, Animal, Plant, Object, Scene, Architecture, and abstract or stylistically ambiguous images unsuitable for destylization are discarded. This yields approximately 10K curated, high-quality artworks, covering 669 artists and 117 art movements, all resized to 1024×1024.

The second source is synthetic style imagery generated with FLUX-Dev. The generation uses a taxonomy of 65 style categories such as Pixel Style, Cyberpunk, Low Poly, Line Art, Watercolor, and Cartoon, together with a hierarchical content tree comprising 6 major classes and 10 subtypes per major class. For each style category, 300 $t$5 combinations are randomly sampled, GPT-4o generates prompts describing style and content, and FLUX-Dev produces the images at 1024×1024. This contributes approximately 150K synthetic style images, producing a combined style pool of roughly 160K images.

For each style image $t$6, GPT-4o is asked to imagine the underlying content without style and to produce a content prompt $t$7 focusing on object categories, pose, scene type, and spatial layout while ignoring stylistic descriptors. DST then generates a destylized image $t$8, yielding raw pairs $t$9.

DST-Filter is a multi-stage evaluation pipeline that scores each pair on content preservation and style discrepancy using Chain-of-Thought reasoning. In the content stage, GPT-4o first enumerates meaningful regions or objects in the style image and then checks whether each is present in the destylized image with preserved structure, pose, and layout. A conservative score $x_{\text{desty}}$0 is assigned, and only pairs with $x_{\text{desty}}$1 are retained. In the style discrepancy stage, GPT-4o extracts stylistic attributes such as color palette, brush strokes, line patterns, texture structure, lighting and shading style, and rendering effects, then evaluates whether those attributes have been removed or attenuated in the destylized image. A score $x_{\text{desty}}$2 is assigned, and only pairs with $x_{\text{desty}}$3 are kept.

For each surviving style image, a reference image is chosen from the same style category by maximizing CSD style similarity. This produces the final triplet $x_{\text{desty}}$4, where $x_{\text{desty}}$5 is the target style image, $x_{\text{desty}}$6 is a style exemplar, and $x_{\text{desty}}$7 is the destylized content image. The final result is DST-100K, a dataset of 100K high-quality triplets (Wang et al., 7 Sep 2025).

Component	Role	Key mechanism
DST	Destylization model	FLUX-Dev, text-guided diffusion reconstruction
DST-Filter	Automatic pair filtering	GPT-4o CoT scoring for content preservation and style discrepancy
DST-100K	Supervised training dataset	100K triplets $x_{\text{desty}}$8
OmniStyle2	Forward style transfer model	FLUX-Dev with LoRA fine-tuning on DST-100K

4. OmniStyle2 model architecture and training procedure

OmniStyle2 is a feed-forward style transfer model built on FLUX-Dev and trained on DST-100K triplets. Given $x_{\text{desty}}$9, the target is the style image $x_{\text{nat}}$0, while the conditioning inputs are the reference image $x_{\text{nat}}$1 and the destylized image $x_{\text{nat}}$2.

The three images are encoded by the FLUX VAE: $x_{\text{nat}}$3 Noise is added to the target latent: $x_{\text{nat}}$4 The DiT then consumes the token sequence

$x_{\text{nat}}$5

using sequential positional encoding so that style tokens, reference tokens, and destylized tokens occupy non-overlapping continuous position ranges. The model uses empty text input, so conditioning is purely image-to-image. Training minimizes

$x_{\text{nat}}$6

No explicit separate content loss or style loss is introduced; supervision is entirely derived from reconstructing the real style image $x_{\text{nat}}$7 from $x_{\text{nat}}$8 within the diffusion denoising framework. Fine-tuning uses LoRA only on the DiT rather than full fine-tuning, with the stated motivation of reducing memory while preserving pretrained knowledge. Training uses 8 × A800 GPUs, learning rate $x_{\text{nat}}$9, batch size 48, and horizontal and vertical flipping.

At inference time, the model receives a style reference and a content or destylized image, encodes them to $x_{\text{cnt}}$0 and $x_{\text{cnt}}$1, initializes a noisy latent $x_{\text{cnt}}$2, conditions on the paired image tokens with the same positional encoding scheme, runs standard diffusion denoising, and decodes the resulting latent with the VAE. The conditioning mechanism is explicitly purely image-image and does not use prompts or text. This distinguishes OmniStyle2 from prompt-mediated editing systems that translate style images into textual surrogates.

5. Empirical evaluation and reported performance

The evaluation benchmark contains 55 content images and 56 style references, yielding 3080 stylized outputs per method through full pairwise combination. Baselines are divided into style transfer systems—OmniStyle, Attention Distillation, StyleID, StyleShot, CSGO, AesPA-Net, and STROTSS—and image editing systems including GPT-4o, FLUX-Kontext, Qwen-Image-Edit, Bagel, and Bagel-Thinking. For the editing models, the style image is converted to a text description because these systems do not support direct multi-image style and content conditioning.

The reported metrics include DINO-Score and CLIP-Score for content preservation, CSD-Score and Style Loss for style similarity, and three Qwen-VL-Max scores in $x_{\text{cnt}}$3: Qwen-Content-Score, Qwen-Style-Score, and Qwen-Aesthetic-Score.

Against style transfer baselines, OmniStyle2 reports:

• DINO-Score: 0.8203
• CLIP-Score: 0.2702
• CSD-Score: 0.5606 (best)
• Style Loss: 0.1170 (best, lower is better)
• Qwen-Content-Score: 8.1385
• Qwen-Style-Score: 7.5763 (best)
• Qwen-Aesthetic-Score: 8.7326 (best by a clear margin)

The interpretation given in the paper is that OmniStyle2 achieves the strongest style fidelity and visual quality while maintaining strong content preservation. OmniStyle and StyleID sometimes obtain slightly higher content scores, but the paper associates that behavior with weaker stylization, described as only subtle color change.

Against editing models, GPT-4o is typically strongest across several metrics, but OmniStyle2 remains highly competitive among open systems. Reported values for OmniStyle2 in that comparison include second-best DINO-Score (0.8203) and CLIP-Score (0.2702), a CSD-Score (0.5606) slightly higher than GPT-4o’s 0.5536, second-best Style Loss (0.1170, versus GPT-4o’s 0.0380), and second-best Qwen-Style-Score (7.5763, behind GPT-4o’s 8.1156). The paper characterizes OmniStyle2 as extremely competitive among open-source, multi-image style transfer models (Wang et al., 7 Sep 2025).

A user study with 30 participants and 810 votes evaluates style preservation, content preservation, and aesthetic appeal. OmniStyle2 receives 28.21% of first-place votes, compared with 18.82% for the next best method, OmniStyle, and appears in the top 3 in 58.95% of judgments, compared with 56.40% for OmniStyle.

Qualitatively, OmniStyle2 is reported to avoid content leakage, maintain consistent style across regions, and produce more coherent cartoonized and artistic renderings than earlier style transfer baselines. Against prompt-based editing models, its advantage is attributed to direct visual style conditioning rather than text-only style surrogates.

6. Position within stylization research, limitations, and scope

OmniStyle2 sits within a broader movement toward FLUX-based and DiT-based stylization systems, but its contribution is primarily data-centric. An adjacent line of work is "OmniConsistency: Learning Style-Agnostic Consistency from Paired Stylization Data" (Song et al., 24 May 2025), which introduces a universal consistency plugin for diffusion transformers under the Flux framework and targets identity, composition, and fine-detail preservation with arbitrary style LoRAs. The contrast is instructive: OmniConsistency addresses stylization fidelity through a plug-in consistency mechanism, whereas OmniStyle2 addresses stylization fidelity through large-scale supervision built from destylization. This suggests two complementary design axes in recent open stylization research: consistency-preserving conditioning and supervision-quality improvement.

Several misconceptions are explicitly corrected by the OmniStyle2 design. First, it is not presented as merely a new loss function; its stated novelty is the reframing of artistic style transfer as a data problem. Second, it is not a prompt-driven editor at inference time; OmniStyle2 uses no text input and relies only on image-based conditioning. Third, although DST-100K provides authentic supervision signals, the supervision is not exact in a strict photorealistic sense, because the dataset depends on approximate destylization, GPT-4o-generated content prompts, and GPT-4o-based filtering.

The paper identifies several limitations. OmniStyle2 can exhibit identity changes, attributed to noise in destylization and approximate captions. Noisy data remains after filtering, and this can propagate into training. The distribution of styles is biased by what is available in WikiArt, NGA, and the selected synthetic taxonomy, despite coverage of 669 artists and 65 digital styles. Out-of-distribution content, especially highly unusual or abstract content, may not destylize well. Finally, DST-Filter relies on GPT-4o, so errors or biases in GPT-4o’s reasoning can affect sample selection.

Future directions named in the paper include more robust filtering, caption-free destylization, larger and more diverse datasets including more non-Western art and niche digital styles, integration with stronger foundation models than FLUX-Dev, and extension of the destylization paradigm to other inverse-style tasks and broader content/style disentanglement problems. Within that trajectory, OmniStyle2 is best understood as a concrete demonstration that authentic supervision via destylization can compete with or surpass methods based on handcrafted feature matching, synthetic pseudo-targets, or prompt-mediated editing (Wang et al., 7 Sep 2025).