Content-style decomposition from a single image is a fundamental challenge in personalized image generation, aiming to separate subject identity from visual style for flexible recomposition. Existing approaches primarily operate in the spatial domain, without explicitly exploiting the distinct frequency characteristics naturally associated with content and style. In this paper, we propose WaveletLoRA, a novel frequency-aware framework for explicit content-style decomposition through wavelet-domain supervision. Specifically, we first apply a multi-level discrete wavelet transform (DWT) to decompose the reference image into low-frequency and high-frequency subbands. We then assign a dedicated LoRA branch to each subband, where the low-frequency branch models global appearance for style acquisition, while the high-frequency branches capture structural details for content preservation. To further improve disentanglement quality, we introduce a dense Mixture-of-Experts aggregation module together with a frequency-domain regularization objective, which reduces style leakage and enhance content fidelity. In addition, we establish WaveBench and develop a VLM-based evaluation protocol that explicitly measures both attribute similarity and attribute separation for disentanglement assessment. Extensive experiments demonstrate that WaveletLoRA consistently outperforms existing methods and produces results that better align with human preference.
Figures & tables
Figure 1: Given a single input image (middle), WaveletLoRA learns a disentangled content or subject LoRA (left) and style LoRA (right) that can be used to generate new images with the learned concepts.
Figure 2: Overview of WaveletLoRA and WaveBench. (Top) Given a reference image, a two-level DWT decomposes the image into one low-frequency subband and multiple high-frequency subbands. The low-frequency branch learns global appearance attributes for style modeling, while the high-frequency branches preserve multi-scale structural details for content representation. An adaptive aggregation MoE module fuses high-frequency representations into a unified content LoRA, followed by frequency-domain disentanglement supervision and AdaIN-based style alignment. (Bottom) Construction pipeline of WaveBench, including prompt generation, manual verification, and reference image synthesis.
Figure 3: Illustration of two - level DWT. Two-level DWT decomposes an image into one low-frequency and six high-frequency subbands at two scales.
Figure 4: Diversity of WaveBench.
Style category
LL 2 -only style
LL 2 -only subject
High-only style
High-only subject
Cartoon
0.85
0.30
0.15
0.85
Chinese painting
0.80
0.25
0.10
0.80
Futuristic
0.90
0.40
0.25
0.85
Oil painting
0.70
0.30
0.05
0.65
Pixel art
0.85
0.45
0.20
0.70
Pointillism
0.75
0.30
0.05
0.65
Table 1: VLM-based recognition accuracy on frequency-retained reconstructions. LL 2 retains more style information, while high-frequency bands retain more subject information across all 8 categories.
Method
WaveBench
VLM Evaluation
User Study Scores
Content-align.
Style-align.
Content-decom.
Style-decom.
Content.
Style.
Combined.
Qwen-Image-Edit
-
-
0.6404
0.5805
0.5925
0.4025
0.8575
OmniGen2
-
-
0.5333
0.4252
0.3325
0.2425
0.7425
B-LoRA
0.5486
0.6115
0.7541
0.8396
0.5725
0.6250
0.6250
UnZipLoRA
0.5529
0.6286
0.8373
0.8693
0.7250
0.7800
0.7075
USO
0.8031
0.5157
0.6503
0.8707
0.3900
0.8325
0.5825
Table 2: Quantitative comparison with state - of - the - art methods. We report feature space alignment metrics (DINOv2 for content consistency and CLIP-I for style consistency), VLM evaluation scores (content - decom., style - decom.), and user study scores (content, style, combined). Best results are highlighted in bold and second-best results are underlined.
Figure 7
Figure 7: Visualization of dense MoE gating weights.
Figure 8: Qualitative comparison. WaveletLoRA produces cleaner content–style separation than other methods. Content outputs (top rows) retain subject identity without style residues, and style outputs (bottom rows) capture artistic textures without subject leakage.
Configuration
Content-align.
Style-align.
WaveletLoRA
0.5866
0.6330
Decomposition Depth
1-level DWT
0.5904
0.6112
3-level DWT
0.5788
0.6317
[1pt/2pt] High-frequency Fusion
Equal averaging
0.5752
0.6329
Table 3: Ablation on decomposition depth and fusion strategy.
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
Style category
Count
Painting
426
Cartoon
97
Futuristic
171
Art
128
Other
178
Total
1,000
Appendix
Table 5: Top-level style and content category distributions in WaveBench.
Qwen2.5-VL-72B
Gemini-3.5-Flash
GPT-5.1
Method
Content
Style
Content
Style
Content
Style
Qwen-Image-Edit
0.6404
0.5805
0.6103
0.5724
0.6152
0.5920
OmniGen2
0.5333
0.4252
0.5169
0.3985
0.5087
0.4003
B-LoRA
0.7541
0.8396
0.7346
0.8450
0.7326
0.8433
UnZipLoRA
0.8373
0.8693
0.8113
0.8584
0.7945
0.8608
USO
0.6503
0.8707
0.5752
0.8601
0.6032
0.8607
Appendix
Table 6: Evaluation scores from three VLMs under the same protocol.
Component
Training time
Inference time
Additional parameters
Base Z-Image (without LoRA)
–
∼ 5 s
0
Base + style and content LoRA
∼ 230 min
∼ 6 s
2×(2×64×d)
WaveletLoRA (full)
∼ 320 min
∼ 10 s
7×(2×64×d)+6×d+AdaIN
Appendix
Table 7: Computational cost comparison.
Table 14
Figure 9: Additional qualitative comparisons across diverse content–style pairs. We show decomposition results on other set of subjects and artistic styles. For each example, the top row displays content - decoupled outputs and the bottom row shows style - decoupled outputs. WaveletLoRA consistently preserves subject identity without ghosting from the original style and extracts coherent artistic textures without silhouette leakage from the source content.
Personalized image generation requires effectively balancing content fidelity with stylistic consistency when synthesizing images based on text and reference examples. Low-Rank Adaptation (LoRA) offers an efficient personalization approach, with potential for precise control through combining LoRA weights on different concepts. However, existing combination techniques face persistent challenges: entanglement between content and style representations, insufficient guidance for controlling elements' influence, and unstable weight fusion that often require additional training. We address these limitations through CRAFT-LoRA, with complementary components: (1) rank-constrained backbone fine-tuning that injects low-rank projection residuals to encourage learning decoupled content and style subspaces; (2) a prompt-guided approach featuring an expert encoder with specialized branches that enables semantic extension and precise control through selective adapter aggregation; and (3) a training-free, timestep-dependent classifier-free guidance scheme that enhances generation stability by strategically adjusting noise predictions across diffusion steps. Our method significantly improves content-style disentanglement, enables flexible semantic control over LoRA module combinations, and achieves high-fidelity generation without additional retraining overhead.
Yu Li, Yujun Cai, Chi Zhang
AGI Lab, Westlake University · George Washington University · The University of Queensland
Learning representations that separate content and style is crucial for controllable generation and compositional generalization. However, diffusion and flow-based models trained primarily with generative objectives often produce entangled or misaligned factors. To address this gap, we introduce Contrastive Augmented Flow Matching (CAtFM), a framework that integrates contrastive regularization into an invertible flow matching formulation to promote structured content-style representations. Rather than constraining intermediate latents or velocity fields, we apply contrastive supervision to predicted endpoints during training, enforcing semantic consistency across transported distributions while allowing disentanglement to emerge implicitly, without assuming strictly pure or fully factorized content and style representations. Our main experiments operate in the CLIP embedding space, with additional validation using frozen DINO and ALIGN encoders. Across synthetic data, in-domain styles, and real-world benchmarks (ImageNet, WikiArt, DomainNet, and DTD), CAtFM improves content and style retrieval, enhances embedding cluster separation, and achieves stronger open-set robustness compared to generative and discriminative baselines. Overall, CAtFM provides a simple way to couple discriminative constraints with deterministic transport, improving disentanglement and robustness under distribution shift.
Yusong Li, Pingchuan Ma, Ming Gui +2
Computer Vision & Learning Group at the University of Munich
Diffusion-based style transfer must balance inference efficiency with stylization fidelity. Adapter-based methods are efficient, but they inject style as an external condition and can either weaken reference-specific appearance or copy reference semantics into the generated image. Optimization-based personalization methods such as LoRA internalize style more effectively, but require a separate training process for every new style. We introduce i2L (image-to-LoRA), a framework that amortizes style LoRA training into a single forward pass. Given one or more reference images, i2L predicts LoRA weights for a text-to-image model, enabling immediate style instantiation without per-style optimization. The architecture combines an image encoder, learnable LoRA queries, and compressed decoding heads that generate adapted matrices. Training on semantically diverse style pairs encourages the predictor to preserve appearance cues while suppressing reference-content copying. Experiments on Z-Image, FLUX.2, and Hidream-O1 show that i2L improves style fidelity, prompt alignment, and perceptual quality over existing baselines. Because i2L produces explicit LoRA weights, it also supports asymmetric classifier-free guidance, multi-reference style fusion, and composition with controllable-generation modules.