Style-Supervised Contrastive Learning (SSCL)

• Style-Supervised Contrastive Learning (SSCL) is an approach that uses contrastive loss on paired image and text data to learn expressive style representations.
• The method integrates large-scale datasets like MegaStyle-1.4M to ensure intra-style consistency and inter-style diversity for robust style transfer.
• SSCL is embedded in a transformer-based diffusion framework, yielding state-of-the-art results in both automatic benchmarks and human evaluations.

MegaStyle-FLUX is a paired-supervised conditional diffusion model for generalizable image style transfer, integrating the FLUX Multi-Modal Diffusion Transformer (MM-DiT) architecture with a large-scale, automatically curated dataset (MegaStyle-1.4M) and style-supervised contrastive pretraining. Developed within the context of advancing text-image-aligned generative models, MegaStyle-FLUX addresses the core challenges of intra-style consistency, inter-style diversity, and scalable coverage of fine-grained artistic styles. The model achieves state-of-the-art results in both automatic and human-evaluated style transfer and style retrieval benchmarks (Gao et al., 9 Apr 2026).

1. Architectural Foundations

MegaStyle-FLUX is constructed atop the FLUX MM-DiT backbone, leveraging a fully transformer-based denoising pipeline with the following main components:

• Frozen VAE Encoder/Decoder: Images, both reference-style and noisy targets, are mapped into (and reconstructed from) a 4× downsampled latent space through the FLUX VAE. Only the DiT blocks and LoRA adapters are trainable.
• Tokenization and Embeddings: The style reference image is encoded to patch tokens $Z_{\text{style}}\in \mathbb{R}^{N_s\times d}$; the noisy target latent is similarly tokenized; the text prompt is processed into tokens $T\in \mathbb{R}^{N_t\times d_t}$ using a frozen pre-trained text encoder.
• Multi-Modal DiT Transformer (LoRA-augmented): Self- and cross-attention layers process $\{Z_{\text{style}}, Z_{\text{noisy}}, T\}$ jointly. A shifted Rotary Positional Encoding (RoPE) is applied to prevent positional collision of style versus target tokens.

The data flow during training sequences style pairs $(x_{\text{ref}}, x_{\text{tgt}})$ (sharing style, differing in content), encodes reference and target images into the latent space, adds forward diffusion noise to the target, and conditions the transformer on all available cues:

$T\in \mathbb{R}^{N_t\times d_t}$1

Integration of FLUX into this pipeline enables paired supervision: the model is trained to reconstruct the target’s content under the style from the reference image (Gao et al., 9 Apr 2026).

2. Diffusion Modeling, Losses, and Training Paradigm

MegaStyle-FLUX implements a standard latent diffusion process:

• Noising Process: For the target image,

$z_0 = \mathrm{VAE}(x_{\text{tgt}})$

$$z_t = \sqrt{\alpha_t} z_0 + \sqrt{1-\alpha_t}\,\epsilon,\quad \epsilon\sim \mathcal{N}(0,I)$$

• Model Prediction: Denoting the transformer output as $\epsilon_\theta(z_t,Z_{\text{style}},T,t)$, the objective is to predict the true noise component in the latent space.
• Diffusion Loss: The loss for the DiT backbone is

$$\mathcal{L}_{\text{diff}} = \mathbb{E}_{x_\mathrm{ref}, x_\mathrm{tgt}, t, \epsilon} \| \epsilon - \epsilon_\theta(z_t, Z_{\text{style}}, T, t) \|_2^2$$

There are no adversarial, perceptual, or explicit style-content reconstruction losses applied to FLUX itself.

For style representation learning, the MegaStyle-Encoder is independently trained using style-supervised contrastive loss:

• Intra-style Supervised Contrastive Loss (SCL): Maximizes similarity between images of the same style.
• Inter-modal Image-Text Contrastive (ITC): Aligns style image encodings and style prompt text descriptions. The sum of both constitutes the overall style-supervised contrastive loss (SSCL), used only to update the image encoder parameters.

No curriculum or staged training is employed; the model is exposed to the full diversity of MegaStyle-1.4M pairings from the onset (Gao et al., 9 Apr 2026).

3. Style Encoder Pretraining and Role

The MegaStyle-Encoder, based on SigLIP’s SoViT image encoder backbone (patch-14, 384 px, 400M parameters), is separately trained over MegaStyle-1.4M for expressive, style-specific representations. After training, the encoder outputs $\ell_2$-normalized embeddings in $\mathbb{R}^{d_z}$ ($T\in \mathbb{R}^{N_t\times d_t}$0), used for:

• Measuring style similarity (cosine distance) in retrieval and evaluation.
• Ablating dataset and encoder contributions to downstream style transfer.
• Enabling rigorous benchmarking of style consistency and diversity in generated results.

Contrastive loss operates as a strict supervised signal, with no bleed-through to the FLUX generator training (Gao et al., 9 Apr 2026).

4. Dataset Construction and Curation: MegaStyle-1.4M

MegaStyle-1.4M is a large-scale, automatically curated dataset designed for style transfer:

• Raw Sources: 2M style images (drawn from JourneyDB, WikiArt, stylized LAION) and 2M non-style LAION images.
• Prompt Extraction: Qwen3-VL captioning yields 2M style and content prompts, subsequently deduplicated with Nemo-Curator.
• Balancing: Hierarchical k-means and top-down cap sampling produce a balanced set of 170K style prompts and 400K content prompts.
• Final Dataset: Via cross-combination, 1.4M pairs enable each style to appear consistently across multiple contents, yielding high intra-style consistency and inter-style diversity. This dataset is critical for disentangling style and content, and for ensuring generalization to both coarse and fine-grained styles.

5. Experimental Results and Benchmarking

MegaStyle-FLUX and MegaStyle-Encoder exhibit strong empirical results:

• Style Retrieval mAP@1/10 and R@1/10: MegaStyle-Encoder outperforms CSD and CLIP by large margins.

Method	mAP@1	mAP@10	R@1	R@10
CLIP (ViT-L)	9.29	6.46	9.29	31.56
CSD (ViT-L)	45.60	37.78	45.60	79.18
MegaStyle-Encoder (ViT-L)	87.26	85.98	87.26	97.61
SigLIP (SoViT)	10.43	7.83	10.43	36.32
MegaStyle-Encoder (SoViT)	88.46	86.77	88.46	97.66

• Style Transfer Main Benchmark:
  • Style similarity (MegaStyle-Encoder cosine): MegaStyle-FLUX = 76.16, surpassing all comparators.
  • Text alignment (CLIP-score): MegaStyle-FLUX = 23.20, highest among all evaluated models.
  • Human evaluation: MegaStyle-FLUX achieves best ranks for both style and text preservation.

Model	Style ↑	Text ↑	Human Style ↑	Human Text ↑
MegaStyle-FLUX	76.16	23.20	31.37	28.72
Next best (StyleShot)	63.42	21.79	15.21	13.69

• Ablation: Training FLUX on smaller or less consistent style datasets (JourneyDB, OmniStyle-150K) results in inferior performance (Style: 34.56, 51.49 vs. 76.16 for MegaStyle-1.4M).
• Qualitative Assessment: MegaStyle-FLUX reproduces brushwork, texture, and color distribution across a spectrum of styles. Failure cases noted include association biases from captioning but no catastrophic style/content bleed.
• Comparative Fine-tuning: StyleShot-FLUX, when retrained on MegaStyle-1.4M, improves but does not reach MegaStyle-FLUX metrics.

6. Analysis of Dataset Ablation and Model Variants

Ablation studies highlight several key factors:

• Intra-style Consistency: Essential for robust stylization; inconsistent pairings lead to poor color and brushwork transfer.
• Inter-style Diversity: Sufficiently granular style coverage is necessary to enable generalization; coarse datasets only yield color transfer.
• Component Robustness: Major gains are attributed to data scale and style-supervised pretraining; individual FLUX architectural modifications (shifted RoPE, LoRA rank) are not explicitly ablated.

A plausible implication is that further performance gains may rely more on advances in dataset curation and style representation than diffusion architecture modifications for this pipeline (Gao et al., 9 Apr 2026).

7. Significance and Research Outlook

MegaStyle-FLUX demonstrates that combining (1) large-scale, intra-style-consistent and inter-style-diverse training data, (2) strict contrastive pretraining of style encoders, and (3) a transformer-based, multi-modal diffusion engine under paired supervision achieves state-of-the-art, generalizable style transfer across thousands of styles. This approach advances beyond U-Net or simple cross-attention paradigms, supporting both text-image alignment and faithful style reproduction. Its evaluation protocol and dataset construction pipeline may become reference standards for subsequent style transfer research, with direct implications for generative art, image retrieval, and foundational studies in style-content disentanglement (Gao et al., 9 Apr 2026).