Style-CCL: Content-Preserving Style Transfer via Curriculum Continual Learning

• URL: http://arxiv.org/abs/2606.14746v1
• Date: Thu, 04 Jun 2026 13:02:04 GMT
• Title: Style-CCL: Content-Preserving Style Transfer via Curriculum Continual Learning
• Abstract summary: Style transfer, given content and style references, remains challenging for Diffusion Transformers.<n>We propose Style-CCL, a Multi-Stage Curriculum Continual Learning framework that trains SC-DiT from semantic (easy) to texture (hard) styles.<n>Our framework achieves state-of-the-art performance in three core metrics: style similarity, content consistency, and aesthetic quality.
• Score: 52.33837657928263
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Content-Preserving Style transfer, given content and style references, remains challenging for Diffusion Transformers (DiTs) due to entangled content and style features. With a reverse triplet synthesis pipeline to build a million-scale training set and a dual-branch Style-Content DiT (SC-DiT) that decouples style and content via separate ROPE embeddings and causal masking, we observe that such a one-stage training paradigm on mixed style categories causes semantic styles to dominate, hindering texture style learning, and harming content preservation. To address these issues, we propose Style-CCL, a Multi-Stage Curriculum Continual Learning framework that trains SC-DiT from semantic (easy) to texture (hard) styles, and from clean to synthetic data, with Random Memory Rehearsal across stages to avoid catastrophic forgetting. Extensive experiments demonstrate that our Style-CCL achieves state-of-the-art performance in three core metrics: style similarity, content consistency, and aesthetic quality.

Related papers

• FreeStyle: Free Control of Style-Content Dual-Reference Generation from Community LoRA Mining [42.06209840326987]
  Style-content dual-reference generation aims to synthesize an image that preserves the structure and semantics of a content reference while adopting the style of a separate style reference.<n>Key bottleneck is the lack of large-scale triplet data with clean content-style separation and broad long-tail style coverage.<n>We propose FreeStyle, a scalable dual-reference generation framework based on community LoRA mining.
  arXiv  Detail & Related papers  (2026-06-18T17:24:27Z)
• UniCSG: Unified High-Fidelity Content-Constrained Style-Driven Generation via Staged Semantic and Frequency Disentanglement [13.265643008768825]
  DiT-based diffusion models often suffer from content-style entanglement, leading to reference-content leakage and unstable generation.<n>We present UniCSG, a unified framework for content-constrained, style-driven generation in both text-guided and reference-guided settings.
  arXiv  Detail & Related papers  (2026-04-20T05:59:20Z)
• Mixture of Style Experts for Diverse Image Stylization [73.7600741657193]
  StyleExpert is a semantic-aware framework based on the Mixture of Experts (MoE)<n>Our framework employs a unified style encoder, trained on our large-scale dataset of content-style-stylized triplets.<n>Our method adeptly handles diverse styles spanning multiple semantic levels, from shallow textures to deep semantics.
  arXiv  Detail & Related papers  (2026-03-17T15:20:19Z)
• TeleStyle: Content-Preserving Style Transfer in Images and Videos [52.76027947278353]
  We present TeleStyle, a lightweight model for both image and video stylization.<n>We curated a high-quality dataset of distinct specific styles and synthesized triplets using thousands of diverse, in-the-wild style categories.<n>TeleStyle achieves state-of-the-art performance across three core evaluation metrics: style similarity, content consistency, and aesthetic quality.
  arXiv  Detail & Related papers  (2026-01-28T02:16:03Z)
• QwenStyle: Content-Preserving Style Transfer with Qwen-Image-Edit [54.11909509184315]
  We propose the first content-preserving style transfer model trained on Qwen-Image-Edit.<n>QwenStyle V1 achieves state-of-the-art performance in three core metrics: style similarity, content consistency, and aesthetic quality.
  arXiv  Detail & Related papers  (2026-01-08T10:22:51Z)
• USO: Unified Style and Subject-Driven Generation via Disentangled and Reward Learning [17.75580298339048]
  We present U.S.O, a Unified Style-Subject Optimized customization model.<n>First, we construct a large-scale triplet dataset consisting of content images, style images, and their corresponding stylized content images.<n>Second, we introduce a disentangled learning scheme that simultaneously aligns style features and disentangles content from style.<n>Third, we incorporate a style reward-learning paradigm denoted as SRL to further enhance the model's performance.
  arXiv  Detail & Related papers  (2025-08-26T12:10:24Z)
• StyleDistance: Stronger Content-Independent Style Embeddings with Synthetic Parallel Examples [48.44036251656947]
  Style representations aim to embed texts with similar writing styles closely and texts with different styles far apart, regardless of content.<n>We introduce StyleDistance, a novel approach to training stronger content-independent style embeddings.
  arXiv  Detail & Related papers  (2024-10-16T17:25:25Z)
• Arbitrary Video Style Transfer via Multi-Channel Correlation [84.75377967652753]
  We propose Multi-Channel Correction network (MCCNet) to fuse exemplar style features and input content features for efficient style transfer. MCCNet works directly on the feature space of style and content domain where it learns to rearrange and fuse style features based on similarity with content features. The outputs generated by MCC are features containing the desired style patterns which can further be decoded into images with vivid style textures.
  arXiv  Detail & Related papers  (2020-09-17T01:30:46Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.