Paper deep dive
Disentangling Dual Image References in Frequency Aware Diffusion Models for Personalized Generation
Haipeng Liu, Yang Wang, Meng Wang
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 92%
Last extracted: 10/13/2026, 4:06:28 AM
Summary
The paper introduces Dual-FDM, a framework for personalized image generation that disentangles dual image references (customization and style/color) within frequency-aware diffusion models. It addresses text misalignment issues in previous methods by using Discrete Cosine Transform (DCT) and Singular Value Decomposition (SVD) to separate foreground and background information into different frequency bands. Specifically, it replaces mid-frequency bands for customization style transfer and low-frequency bands for color style transfer, using adaptive entropy-based modulation to balance denoising stages.
Entities (9)
Relation Signals (8)
Dual-FDM → performstask → Customization Style Transfer
confidence 95% · tackle two crucial personalized image generation tasks: customization style transfer and color style transfer
Dual-FDM → performstask → Color Style Transfer
confidence 95% · tackle two crucial personalized image generation tasks: customization style transfer and color style transfer
Dual-FDM → usestechnique → Discrete Cosine Transform
confidence 95% · map the foreground or background of image references into the frequency domain via Discrete Cosine Transform (DCT)
Dual-FDM → usestechnique → Singular Value Decomposition
confidence 95% · reformulated through singular value decomposition (SVD) to derive the corresponding background matrix
Customization Style Transfer → modifies → Mid-frequency band
confidence 90% · replace the mid-frequency band of the background in the style reference with that from the foreground
Color Style Transfer → modifies → Low-frequency band
confidence 90% · substitute the low-frequency band of the background in the style reference with that from both the foreground and background
Dual-FDM → outperforms → IP-Adapter
confidence 90% · Experiments validate the superiority of Dual-FDM over the state-of-the-art diffusion models
→ →
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Personalized image generation aims to synthesize text-driven images conditioned on reference images, while mainly casting the generation as image customization for foreground and style transfer for background. Previous arts of diffusion models suffers from the text misalignment with background for image customization and foreground for style transfer during the denoising process. Such facts, as we observed, rooted from the entanglement among hybrid frequency bands during the denoising process. To address such salient limitation, in this paper, we study personalized generation based on dual references - customization and color and style reference - and propose a paradigm to disentangle these Dual image references within Frequency-aware Diffusion Models, dubbed Dual-FDM, to simultaneously tackle two crucial personalized image generation tasks: customization style transfer and color style transfer, by disentangling different frequency bands via mask strategy within frequency domain. For customization style transfer, we replace the mid-frequency band of the background in the style reference with that from the foreground of the customized reference. For color style transfer, we substitute the low-frequency band of the background in the style reference with that from both the foreground and background of the color reference. Both the substituted frequency bands are used as the key and value to reconstruct the query foreground and background of the denoised personalized this http URL experiments validate the superiority of Dual-FDM over the state-of-the-art diffusion models for personalized image generation. Our code can be accessed from this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2610.07684v1
- Canonical: https://arxiv.org/abs/2610.07684v1
Trouble viewing inline? Open PDF directly →
Full Text
105,303 characters extracted from source content.
Expand or collapse full text
Disentangling Dual Image References in Frequency Aware Diffusion Models for Personalized Generation Haipeng Liu Yang Wang †thanks: Yang Wang is the corresponding author Meng Wang Affiliation: School of Computer Science and Information Engineering Affiliation: Hefei University of Technology, China Email: hpliu_hfut@hotmail.com yangwang@hfut.edu.cn eric.mengwang@gmail.com Abstract Personalized image generation aims to synthesize text-driven images conditioned on reference images, while mainly casting the generation as image customization for foreground and style transfer for background. Previous arts of diffusion models suffers from the text misalignment with background for image customization and foreground for style transfer during the denoising process. Such facts, as we observed, rooted from the entanglement among hybrid frequency bands during the denoising process. In particular, the low-frequency band primarily captures color information, mid-frequency band encodes global structure, while high-frequency band corresponds to local texture. Besides, the early (high-noise) denoising stage mainly focuses on the low-and-mid frequency bands, while the late (low-noise) denoising stage progressively exhibits the high-frequency band, resulting in the dominance of low-and-mid over high frequency bands, especially when entangled within a single image reference, tends to reinforce each other at the early stage, hence the pitfall of overwhelming suppression of text prompt guidance. To address such salient limitation, in this paper, we study personalized generation based on dual references - customization and color and style reference - and propose a paradigm to disentangle these Dual image references within Frequency-aware Diffusion Models, dubbed Dual-FDM, to simultaneously tackle two crucial personalized image generation tasks: customization style transfer and color style transfer, by disentangling different frequency bands via mask strategy within frequency domain. For customization style transfer, we replace the mid-frequency band of the background in the style reference with that from the foreground of the customized reference. For color style transfer, we substitute the low-frequency band of the background in the style reference with that from both the foreground and background of the color reference. Both the substituted frequency bands are used as the key and value to reconstruct the query foreground and background of the denoised personalized image. We further compute the ratio of information entropy of the substituted frequency bands to adaptively modulate the denoising timesteps between the early and late stages. Extensive experiments validate the superiority of Dual-FDM over the state-of-the-art diffusion models for personalized image generation. Our code can be accessed from https://github.com/htyjers/Dual-FDM. Figure 1: Personalized images generated by our Dual-FDM. (a) The generated examples of Dual-FDMcust for customized references (left) and style references (above); (b) The generated examples of Dual-FDMcost for color references (left) and style references (above). 1 Introduction Figure 2: (a) Comparison of the generated results, as well as the attention score map of text prompts and image references, between IP-Adapter Ye et al. (2023) and MS-Diffusion Wang et al. (2025) for image customization. (b) Comparison of the generated results between InstantStyle Wang et al. (2024a) and StyleStudioLei et al. (2025) for style transfer. (c) Comparison of generated results guided by different frequency band of the image reference. Personalized image generation focuses on synthesizing text-driven images conditioned on reference images, which enjoys remarkable progress, benefiting from the powerful denosing capabilities of Denoising Diffusion Probabilistic Models (DDPMs) Ho et al. (2020); Liu et al. (2024); Liu et al. (2025a), e.g., Stable Diffusion Rombach et al. (2022); Esser et al. (2024); Podell et al. (2023); Qian et al. (2022); Qian et al. (2023); Wang et al. (2024c). Typically, the personalized image generation involves two key subtasks, i.e., image customization and style transfer. Given text prompts, image customization Xiao et al. (2025); Zeng et al. (2024); Pang et al. (2024); Gal et al. (2022); Ruiz et al. (2023); Kumari et al. (2023); Liu et al. (2025b); Ma et al. (2024); Li et al. (2024); Li et al. (2023); Liu et al. (2022); Cao et al. (2026); Jiang et al. (2026) aims to generate images by substituting the foreground semantics in the text prompts with the foreground extracted from the reference images, while preserving the background of the text prompts. In contrast, style transfer Qin et al. (2025); Rout et al. (2025); Liu et al. (2023); Sohn et al. (2023); Zhu et al. (2025); Qi et al. (2024); Hertz et al. (2024); Lei et al. (2025); Gao et al. (2024); Xing et al. (2024); Wang et al. (2024b); Wang et al. (2024a); Jeong et al. (2024); Wang et al. (2022) focuses on disentangling style (background) from reference images and aligning the foreground with the text prompt. However, prior arts have struggled to align the generated image with the text prompts within each individual subtask. Figure 3: Illustration of our proposed Dual-FDM pipeline (a); For customization style transfer (b), we replace the mid-frequency band of the background in the style reference with that from the foreground of the customized reference, named Dual-FDMcust (Sec. 2.3). For the color style transfer (c), we substitute the low-frequency band of the background in the style reference with that from both the foreground and background of the color reference, named Dual-FDMcost (Sec. 2.4). Each of the two personalized image generation tasks are conducted in the early and late stage of the whole denoising process guided by text prompts as illustrated in Eq. (1). For image customization, substantial efforts Ye et al. (2023); Aiello et al. (2025); Huang et al. (2025b) aim to emphasize dense image references of the foreground while overwhelming sparse text prompts of the background, especially for complex ones, e.g., a degraded “flower wreath” and the missing “coffee” in the IP-Adapter Ye et al. (2023) of Fig. 2(a). Building on this, layout-based methods Wang et al. (2025); Huang et al. (2025b); Zhang et al. (2024) exploit spatial layouts to explicitly constrain reference concepts to the foreground, while assigning background regions to text prompts. This preserves the foreground identity, but the explicit separation of foreground and background in the layout disrupts overall consistency, leading to the misalignment with (complex / long) text prompts. As illustrated in the MS-Diffusion Wang et al. (2025) of Fig. 2(a), the generated spatial arrangement of “flower wreath” and “coffee” diverges from the prompt and instead resembles “A flower wreath around a dog near a coffee”. When coming to style transfer, during the early stage of the denoising process with high-level noise, the foreground is reconstructed from sparse text prompts, the dense style information intended for the background often leaks into the foreground, resulting in misalignment with the text prompts, e.g., the incorrect number of “woman" in the InstantStyle Wang et al. (2024a) of Fig. 2(b). To resolve such issue, StyleStudio Lei et al. (2025) applies style via instance normalization Huang and Belongie (2017) in the cross-attention layers and leverages layout to preserve foreground in the self-attention layers. Nevertheless, self-attention score maps highlight only the most salient regions, implying that the extracted layout can align only with simple text prompts (i.e., “a dog") rather than the multi-objects (“two pillows and one lemon" in the StyleStudio Lei et al. (2025) of Fig. 2(b)) text prompts, thereby preventing it from handling the complex text prompts. So far, neither image customization nor style transfer achieves satisfactory alignment between the generated image and the text prompt for diffusion models. Such facts, as we observed, rooted from the entanglement among hybrid frequency bands. Intuitively, as shown in Fig. 2(c), the low-frequency and mid-frequency bands capture color and global structure, while high-frequency band focuses on local texture, where different frequency bands are mainly denoised during different stages of the denoising process. As we observed, the early (high-noise) denoising stage mainly focuses on the low-and-mid frequency bands, while the late (low-noise) denoising stage progressively exhibits the high-frequency band (See Sec. A of the Appendix), the dominance of low-and-mid over high frequency bands, especially when entangled within a single image reference, suggesting reinforce each other at the early stages, resulting in an overwhelming suppression of text prompt guidance. To resolve the text misalignment for personalization generation, we propose to solve the problem upon dual references and develop a paradigm to disentangle these Dual image references within Frequency-aware Diffusion Models, dubbed Dual-FDM, for two classical personalized image generation tasks in a row: customization style transfer and color style transfer, by disentangling different frequency bands via mask strategy within frequency domain. Specifically: • We propose to map the foreground or background of image references into the frequency domain via Discrete Cosine Transform (DCT), where a mask is applied from left to right to progressively disentangle the low-, mid-, and high-frequency bands. Notably, for customization style transfer, we replace the mid-frequency band of the background in the style reference with that from the foreground of the customized reference, named Dual-FDMcust (Fig. 1(a)). The style reference with the mid-frequency band substituted is used as the key and value to reconstruct the query foreground and background of the denoised personalized image in the decoupled cross-attention layer. For the color style transfer, we substitute the low-frequency band of the background in the style reference with that from both the foreground and background of the color reference, named Dual-FDMcost (Fig. 1(b)). The style reference with the substituted low-frequency band is further utilized as the key and value to reconstruct the query foreground and background of the denoised personalized image in the decoupled cross-attention layer. • To pop up the foreground or background of the image reference, we utilize the key projection matrices in the decoupled cross-attention layer to represent the foreground, which are then reformulated through singular value decomposition (SVD) to derive the corresponding background matrix for the background. • For each task, we further compute the ratio of information entropy of the substituted frequency bands to adaptively modulate the denoising timesteps between the early and late stages, so as to achieve a well-balanced performance between text alignment and dual-image similarity. Extensive experiments validate the superiority of Dual-FDM over the state-of-the-art diffusion models for personalized image generation. 2 Methodology Central to our proposed Dual-FDM cover the following aspects: 1) how to pop up the background of the image references (Sec. 2.2); 2) how to disentangle the mid-frequency band in customization style transfer (Sec. 2.3); 3) how to disentangle the low-frequency band in color style transfer (Sec. 2.4); 4) how to adaptively modulate the timesteps within the denoising process between early and late stages (Sec. 2.5). Before shedding light on Dual-FDM, we elaborate the preliminary regarding decoupled cross-attention mechanism, which plays a pivotal role of personalized image generation. 2.1 Preliminary Often, personalized image generation methods adopt a decoupled cross-attention mechanism to integrate text prompt Cte∈ℝste×dteC_te ^s_te× d_te, the reference image Cim∈ℝsim×dimC_im ^s_im× d_im into every cross-attention layer of the denoising network where stes_te and sims_im denote the token number of the text prompts and image patches, such that dted_te and dimd_im are the dimension for token and patch representation. The denoised latent image feature F∈ℝ(h×w)×dF ^(h× w)× d serves as the query to be reconstructed by the key of the text prompts via the l-th cross-attention layers for the UNet architecture ϵθ(F,t,Cte,Cim) _θ(F,t,C_te,C_im), as formulated below: F¯te=Atten(Q,Kte,Vte)=Softmax(QKteTdte)Vte, split F_te=Atten(Q,K_te,V_te)=Softmax ( QK_te^T d_te )V_te,\\ split (1) where Q∈ℝ(h×w)×dQ ^(h× w)× d denotes the query matrix, such that Q=FWqQ=FW_q with Wq∈ℝd×dW_q ^d× d being as trainable linear projection matrix. Likewise, Kte∈ℝste×dK_te ^s_te× d and Vte∈ℝste×dV_te ^s_te× d denote the key and value matrices, such that K=CteWkteK=C_teW_k^te and V=CteWvteV=C_teW_v^te with Wkte∈ℝdte×dW_k^te ^d_te× d and Wvte∈ℝdte×dW_v^te ^d_te× d being as trainable linear projection matrices. Meanwhile, F is fed into another cross-attention module to interact with the reference image embeddings below: F¯im=Atten(Q,Kim,Vim)=Softmax(QKimTdim)Vim, split F_im=Atten(Q,K_im,V_im)=Softmax ( QK_im^T d_im )V_im,\\ split (2) where Kim∈ℝsim×dK_im ^s_im× d and Vim∈ℝsim×dV_im ^s_im× d denote the key and value matrices, such that K=CimWkimK=C_imW_k^im and V=CimWvimV=C_imW_v^im with Wkim∈ℝdim×dW_k^im ^d_im× d and Wvim∈ℝdim×dW_v^im ^d_im× d being as trainable linear projection matrices. The predicted noise, i.e., output of ϵθ(F,t,Cte,Cim) _θ(F,t,C_te,C_im), is then subtracted from F to update the latent feature, such denoising process iterates until t=0t=0, yielding the final image. 2.2 Popping up the Background of Image Reference Given text prompts, the major bottleneck for disentangling the dual image references lies in mapping the foreground or background of image references into the frequency domain. To this end, our basic idea is to pop up the background of the image reference, which paves the way to achieve the text alignment with two classical personalized image generation tasks in a row. Following Eq. (2), to obtain the background of the image reference within the decoupled cross-attention layer, we visualize WkimW_k^im to analyze the regions where its contributions are predominantly concentrated, as illustrated in the last column of Fig. 5(a), where WkimW_k^im indicates that the attention scores among different patches effectively encode their foreground. To mitigate such issue, we perform singular value decomposition (SVD) on WkimW_k^im: Wkim=UΣVT,U∈ℝdim×dim,V∈ℝd×dim,Σ=diag(σ1,σ2,…,σdim)∈ℝdim×dim, W_k^im=U V^T,U ^d_im× d_im,V ^d× d_im, =diag\! ( _1, _2,…, _d_im ) ^d_im× d_im, (3) where all singular values σi _i in Σ indicate the relative importance of each component. We leverage distinct singular components to disentangle the background and foreground from the reference image, following the observations in Remark 1. Figure 4: We visualize image components reconstructed from different singular value intervals. The reference image from first two columns correspond to 3D spatial relationships ("in front of"), and the last two to 2D cases ("on right of" and "on top of"). Larger σi _i (0–10%) mainly capture foreground, while smaller ones (90–100%) correspond to background. In 3D scenes, varying saliency leads singular values in 20–30% to represent partial objects. In contrast, 2D scenes exhibit more uniform saliency across targets. Figure 5: The style transfer generated results using (a) WkimW_k^im (Eq. (3)), and (b) background matrix W¯kim W_k^im (Eq. (4)), which demonstrates that the background from the image reference predominantly contributes to reconstructing the denoised image. Remark 1. Larger σi _i tend to correspond to the foreground, whereas smaller σi _i are more likely associated with the background. As illustrated in Fig. 5, SVD decomposes the image feature into components corresponding to different saliency levels, where components associated with larger singular values represent more salient structures in the image. ■ To suppress the foreground components and enhance the background components, we invert the singular values and construct the background matrix W¯kim W_k^im as follows: W¯kim=Udiag(1σ1,1σ2,…,1σdim)VT,σi≠0. W_k^im=Udiag\! ( 1 _1, 1 _2,…, 1 _d_im )V^T, _i≠ 0. (4) We remark the intuition of Eq. (4) as follows: Remark 2. Inverting the singular values suppresses dominant foreground components while enhancing background-related information. Eq. (4) provides an intuitive means of disentangling the background from the foreground in the cross-attention layer. As illustrated in the fourth column of Fig. 5(b), W¯kim W_k^im suppresses the foreground and retains background-related information. Due to page limitation, we provide more intuitions for the rationale between WkimW_k^im and W¯kim W_k^im in Sec. B of the Appendix. ■ 2.3 Customization Style Transfer: Replacing the Mid-frequency of Image Reference Following WkimW_k^im (Eq. (3)) and W¯kim W_k^im (Eq. (4)), we initialize the denoising process with random Gaussian noise zTz_T, together with the text prompts CteC_te to describe both foreground and background, the customized reference Ccu∈ℝsim×dimC_cu ^s_im× d_im, and the style reference Cst∈ℝsim×dimC_st ^s_im× d_im, as inputs to the denoising network ϵθ(F,t,Cte,Ccu,Cst) _θ(F,t,C_te,C_cu,C_st) to generate the denoised results for both the image customization and style transfer, then we reformulate Eq. (2) as follows: F¯imcust=Atten(Q,Kimcustmmid,Vimcustmmid)=Softmax(Q(Kimcustmmid)Tdim)Vimcustmmid, split F_im_cust&=Atten(Q,K^m_mid_im_cust,V^m_mid_im_cust)=Softmax ( Q(K^m_mid_im_cust)^T d_im )V^m_mid_im_cust,\\ split (5) where Kimcustmmid∈ℝsim×dK^m_mid_im_cust ^s_im× d and Vimcustmmid∈ℝsim×dV^m_mid_im_cust ^s_im× d denote the key and value matrices obtained by replacing the mid-frequency bands of the background in the style reference with those from the foreground of the customized reference via mid-frequency replacing layer (see Fig. 6(a)), as formulated below: Kimcustmmid K^m_mid_im_cust =IDCT(DCT(CcuWkim)⊙mid+DCT(CstW¯kim)⊙(1−mid)), =IDCT (DCT(C_cuW_k^im) m_mid+DCT(C_st W_k^im) (1-m_mid) ), (6) Vimcustmmid V^m_mid_im_cust =IDCT(DCT(CcuWvim)⊙mid+DCT(CstWvim)⊙(1−mid)), =IDCT (DCT(C_cuW_v^im) m_mid+DCT(C_stW_v^im) (1-m_mid) ), where DCT(⋅·) represents Discrete Cosine Transform, IDCT(⋅·) represents Inverse DCT, ⊙ is the element-wise multiplication and midm_mid denotes the mid-pass mask to isolate mid-frequency band, the denoised mid-frequency band layer employs a binary mask defined by the sum of the 2D coordinates as a threshold below: mid(x,y)=1,if thmp1<x+y≤thmp20,otherwisem_mid(x,y)= cases1,&if th_mp1<x+y≤ th_mp2\\ 0,&otherwise cases (7) where thmp1th_mp1 and thmp2th_mp2 are the thresholds for mid-pass filtering to extract mid-frequency band. Following the above intuition, the final output of the decoupled cross-attention is F¯cust=F¯te+λF¯imcust F_cust= F_te+λ F_im_cust, such that λ∈[0,1]λ∈[0,1] is a balancing weight. Figure 6: Illustration of (a) mid-frequency replacing layer and (b) low-frequency replacing layer. 2.4 Color Style Transfer: Replacing the Low-frequency of Image Reference Based on WkimW_k^im (Eq. (3)) and W¯kim W_k^im (Eq. (4)), we initialize the denoising process with random Gaussian noise zTz_T, together with the text prompts CteC_te to describe both foreground and background, the color reference Cco∈ℝsim×dimC_co ^s_im× d_im, and the style reference Cst∈ℝsim×dimC_st ^s_im× d_im, as inputs to the denoising network ϵθ(F,t,Cte,Cco,Cst) _θ(F,t,C_te,C_co,C_st) to generate the denoised results for both the color transfer and style transfer, then we reformulate Eq. (2) as follows: F¯imcost=Atten(Q,Kimcostmlow,Vimcostmlow)=Softmax(Q(Kimcostmlow)Tdim)Vimcostmlow, split F_im_cost&=Atten(Q,K^m_low_im_cost,V^m_low_im_cost)=Softmax ( Q(K^m_low_im_cost)^T d_im )V^m_low_im_cost,\\ split (8) where Kimcostmlow∈ℝsim×dK^m_low_im_cost ^s_im× d and Vimcostmlow∈ℝsim×dV^m_low_im_cost ^s_im× d denote the key and value matrices obtained by replacing the low-frequency bands of the background in the style reference with those from the foreground and the background of the color reference via low-frequency replacing layer (see Fig. 6(b)), and is then formulated as: Kimcostmlow K^m_low_im_cost =IDCT(DCT(Cco(Wkim+W¯kim2))⊙mlow+DCT(CstW¯kim)⊙(1−mlow)), =IDCT (DCT\! (C_co\! ( W_k^im+ W_k^im2 ) ) m_low+DCT(C_st W_k^im) (1-m_low) ), (9) Vimcostmlow V^m_low_im_cost =IDCT(DCT(CcoWvim)⊙mlow+DCT(CstWvim)⊙(1−mlow)), =IDCT (DCT(C_coW_v^im) m_low+DCT(C_stW_v^im) (1-m_low) ), where mlowm_low is defined as the low-pass mask below: mlow(x,y)=1,if x+y≤thlp0,otherwisem_low(x,y)= cases1,&if x+y≤ th_lp\\ 0,&otherwise cases (10) where thlpth_lp is the threshold for low-pass filtering. Following the above, the final output of the decoupled cross-attention is F¯cost=F¯te+λF¯imcost F_cost= F_te+λ F_im_cost. 2.5 Adaptively Modulating the Early and Late Denoising Stage To adaptively modulate the early and late stages, we compute the threshold Δcust _cust for customization style transfer and Δcost _cost for color style transfer as follows: Δcust _cust =H(Kimcustmmide)H(Kimcustmmidl),Δcost=H(Kimcostmlowe)H(Kimcostmlowl),H(x)=−∑i=1dSoftmax(x)ilog(Softmax(x)i+ϵ), = H\! (K^m^e_mid_im_cust )H\! (K^m^l_mid_im_cust ), 9.24994pt _cost= H\! (K^m^e_low_im_cost )H\! (K^m^l_low_im_cost ), 9.24994ptH(x)=- _i=1^dSoftmax(x)_i \! (Softmax(x)_i+ε ), (11) where H(⋅)H(·) denotes the information entropy of the mixed image references, and ϵε is a small constant to ensure numerical stability. midem^e_mid and midlm^l_mid represent the mid-frequency masks in the early and late denoising stages, respectively, where midem^e_mid covers a wider range of mid-frequency bands than midlm^l_mid. To be analogous, mlowem^e_low and mlowlm^l_low denote the low-frequency masks in the early and late denoising stages, respectively, with mlowem^e_low encompassing a broader range of low-frequency bands than mlowlm^l_low. Notably, a larger ratio Δcust _cust or Δcost _cost indicates that the mixed reference of early stage contains more meaningful semantics than that within the late stage, hence more timesteps in the early stage and fewer timesteps in the late stage are required. For the timesteps of the denoising process from T to 0, as per the threshold Δ∈Δcust,Δcost ∈\ _cust, _cost\ in Eq. (11), we adaptively define the timesteps interval within t∈[11+ΔT,T]t∈[ 11+ T,T] to cover the early stage of the denoising process, while the late stage as t∈[0,11+ΔT)t∈[0, 11+ T). More details about the algorithm can be accessed from Sec.C of the Appendix. 3 Experiments 3.1 Implementation Details We evaluate Dual-FDM on combined datasets covering image customization, style transfer, and text-to-image (T2I) generation: for small-scale customization-style transfer, we use DreamBench Ruiz et al. (2023) and TI benchmark Gal et al. (2022) as customization references with IS-Plus Wang et al. (2024b) as style references; for large-scale settings, we adopt DreamBench++ Peng et al. (2025) and StyleBench Gao et al. (2024); and for color style transfer, we combine IS-Plus and StyleBench with prompts from T2I-CompBench++ Huang et al. (2025a). We provide 9 evaluation metrics: (i) Text and Image Alignment via CLIP11 1 Since CLIP is trained on real image–text (T) pairs, it is particularly sensitive to image style, which may distort text–image similarity. To alleviate this, we propose CLIP-TS by measuring similarity with style reference (S), thereby reducing the style influence . In addition, the generated background aligned with text prompts can interfere with the customized reference (C); hence we propose CLIP-TC, which incorporates text prompts to suppress such interference. We also provide CLIP-T (image–text similarity) and CLIP-CS (image–image reference similarity) for comparison. Radford et al. (2021) and DINO Caron et al. (2021); (i) Image Quality via Image Reward (IR) Xu et al. (2023) and HPS Wu et al. (2023). All experiments are implemented in PyTorch on a RTX 4090 GPU using IP-Adapter Ye et al. (2023) with SDXL Podell et al. (2023), 30 steps, and official hyper-parameters for fair comparison. More implementation details can be accessed in the Sec.D of the Appendix. Table 1: Quantitative comparisons between Dual-FDMcust and other diffusion-based personalization models on several datasets. ■ and ■ denote the best and second-best results, respectively. Dual-FDMcust consistently achieves the best performance. Baseline IP-Adapter Arxiv’23 SSR-Encoder CVPR’24 MIP-Adapter AAAI’25 Dreamcache CVPR’25 PatchDPO CVPR’25 MS-Diffusion ICLR’25 Dual-FDMcust CSGO (Arxiv’24) CLIP-T ↑ 0.2851 0.2829 0.2809 0.2803 0.2793 0.3123 0.3194 CLIP-TC ↑ 0.4802 0.4689 0.4744 0.4518 0.4738 0.4680 0.5031 CLIP-TS ↑ 0.4415 0.4439 0.4420 0.4525 0.4440 0.4577 0.4516 CLIP-CS ↑ 0.6658 0.6587 0.6626 0.6465 0.6638 0.6419 0.6687 IR×10 ↑ -5.5733 -7.4241 -6.5255 -6.7782 -7.1456 4.2023 5.3487 HPS×10 ↑ 2.3401 2.2692 2.2624 2.1794 2.2493 2.5436 2.6007 StyleID (CVPR’24) CLIP-T ↑ 0.2731 0.2768 0.2848 0.2647 0.2650 0.3012 0.3194 CLIP-TC ↑ 0.4828 0.4773 0.4699 0.4424 0.4744 0.4614 0.5031 CLIP-TS ↑ 0.4349 0.4370 0.4458 0.4488 0.4370 0.4556 0.4516 CLIP-CS ↑ 0.6731 0.6665 0.6584 0.6460 0.6710 0.6423 0.6687 IR×10 ↑ -8.6015 -8.7602 -9.5626 -9.9445 -10.4733 2.6485 5.3487 HPS×10 ↑ 2.0861 2.0979 2.0073 1.9126 1.9978 2.3322 2.6007 StyleSSP (CVPR’25) CLIP-T ↑ 0.2642 0.2656 0.2584 0.2561 0.2591 0.2907 0.3194 CLIP-TC ↑ 0.4664 0.4729 0.4616 0.4355 0.4638 0.4540 0.5031 CLIP-TS ↑ 0.4397 0.4407 0.4409 0.4468 0.4413 0.4583 0.4516 CLIP-CS ↑ 0.6680 0.6592 0.6678 0.6433 0.6700 0.6462 0.6687 IR×10 ↑ -10.8440 -10.4537 -12.0747 -11.0862 -12.1339 0.5184 5.3487 HPS×10 ↑ 2.0087 2.0153 1.9075 1.8562 1.9204 2.2367 2.6007 Figure 7: Comparison of (a) customization style transfer and (b) color style transfer with state-of-the-arts. Dual-FDM delivers superior results compared with existing approaches. 3.2 Comparison with State-of-the-arts Quantitative and Qualitative Analysis for Customization Style Transfer To validate the superiority of Dual-FDMcust, we perform a thorough comparison against state-of-the-art diffusion models for customization-style transfer. For fair comparison, we also compare with both image customization and style transfer. For image customization, we consider two groups of representative methods. The first group includes IP-Adapter Ye et al. (2023), Dreamcache Aiello et al. (2025), and PatchDPO Huang et al. (2025b), while these approaches emphasize dense image references of the foreground while overwhelming sparse text prompts of the background, especially for complex ones. The second group consists of SSR-Encoder Zhang et al. (2024), MIP-Adapter Huang et al. (2025b), and MS-Diffusion Wang et al. (2025), resorting to layout constraints to integrate customized references. This preserves the foreground identity, but the explicit separation of foreground and background in the layout disrupts overall consistency, leading to misalignment with (complex / long) text prompts. Building on the above, we adopt the latest image-driven style transfer methods for the second stage of the style transfer, including CSGO Xing et al. (2024), StyleID Chung et al. (2024), and StyleSSP Xu et al. (2025). The quantitative results in Table. 1 summarize our findings below: Dual-FDMcust enjoys larger CLIP-T, CLIP-TC, CLIP-TS, CLIP-CS, IR and HPS than the competitors. Notably, [MS-Diffusion Wang et al. (2025) + CSGO Xing et al. (2024)] remains the large performance margins at most 2.27% for CLIP-T, 7.5% for CLIP-TC, 4.17% for CLIP-CS 27.28% for DINO-CS and 2.24% for HPS) compared to Dual-FDMcust in Table. 1, the results verifies the intuition in Sec. 1 – the dominance of low-and-mid over high frequency bands—especially when entangled within a single image reference—tends to reinforce each other at the early stages, resulting in an overwhelming suppression of text prompt guidance. We also illustrate more generation results in Fig. 7(a). More analysis of computational efficiency can be seen in Sec.E of the Appendix. Table 2: Quantitative comparisons for color style transfer between Dual-FDMcost and other diffusion-based personalization models on T2I-CompBench++. ■ and ■ denote the best and second-best results. Dual-FDMcost achieves the best performance. InstantStyle CSGO StyleShot StyleAligned DEADiff StyleCrafter MaskST StyleStudio Dual-FDMcost Arxiv’24 NeurIPS’25 Arxiv’24 CVPR’24 CVPR’24 TOG’24 ICLR’25 CVPR’25 Color CLIP-TS ↑ 0.4587 0.4539 0.4644 0.4832 0.4448 0.4771 0.4641 0.4497 0.5015 CLIP-S ↑ 0.5967 0.6392 0.6782 0.6904 0.6177 0.6304 0.6763 0.5859 0.7178 DINO-S ↑ 0.2146 0.3924 0.4769 0.4437 0.3318 0.3541 0.4764 0.2370 0.4970 Shape CLIP-TS ↑ 0.4507 0.4561 0.4595 0.4797 0.4392 0.4656 0.4597 0.4504 0.4956 CLIP-S ↑ 0.5864 0.6357 0.6602 0.7051 0.5986 0.6274 0.6602 0.6162 0.7271 DINO-S ↑ 0.2087 0.3711 0.4374 0.4658 0.2944 0.3843 0.4368 0.3087 0.4982 Texture CLIP-TS ↑ 0.4524 0.4561 0.4543 0.4819 0.4304 0.4712 0.4534 0.4443 0.4988 CLIP-S ↑ 0.5879 0.6406 0.6553 0.7246 0.5835 0.6348 0.6538 0.5889 0.7358 DINO-S ↑ 0.2131 0.3760 0.4476 0.4844 0.2707 0.3729 0.4465 0.2616 0.4991 Non Spatial CLIP-TS ↑ 0.4424 0.4553 0.4504 0.4839 0.4402 0.4553 0.4504 0.4314 0.4958 CLIP-S ↑ 0.5723 0.6279 0.6387 0.7251 0.5972 0.6064 0.6387 0.5566 0.7261 DINO-S ↑ 0.2336 0.4278 0.4853 0.5171 0.3533 0.3795 0.4854 0.2646 0.5406 2D Spatial CLIP-TS ↑ 0.4568 0.4670 0.4661 0.4912 0.4563 0.4829 0.4668 0.4565 0.5122 CLIP-S ↑ 0.5835 0.6309 0.6592 0.7129 0.6362 0.6421 0.6597 0.6060 0.7314 DINO-S ↑ 0.2262 0.4113 0.4615 0.4998 0.3824 0.3977 0.4647 0.3253 0.5187 3D Spatial CLIP-TS ↑ 0.4573 0.4624 0.4663 0.4875 0.4497 0.4773 0.4663 0.4429 0.5078 CLIP-S ↑ 0.5918 0.6333 0.6680 0.7100 0.6333 0.6357 0.6680 0.5732 0.7334 DINO-S ↑ 0.2166 0.3819 0.4654 0.4760 0.3655 0.3774 0.4696 0.2249 0.5094 Number CLIP-TS ↑ 0.4490 0.4509 0.4597 0.4902 0.4465 0.4783 0.4597 0.4436 0.5029 CLIP-S ↑ 0.5845 0.6182 0.6562 0.7241 0.6226 0.6499 0.6558 0.6011 0.7334 DINO-S ↑ 0.2289 0.3882 0.4779 0.5116 0.3541 0.4346 0.4738 0.3160 0.5469 Complex CLIP-TS ↑ 0.4404 0.4500 0.4556 0.4836 0.4390 0.4636 0.4558 0.4351 0.4954 CLIP-S ↑ 0.5864 0.6445 0.6582 0.7256 0.6021 0.6387 0.6582 0.5791 0.7329 DINO-S ↑ 0.2134 0.3911 0.4533 0.4961 0.2938 0.3956 0.4536 0.2274 0.5051 Quantitative and Qualitative Analysis for Color Style Transfer we conduct additional evaluations against state-of-the-art diffusion models for color style transfer. These include InstantStyle Wang et al. (2024a), CSGO Xing et al. (2024), StyleShot Gao et al. (2024), StyleAligned Hertz et al. (2024), DEADiff Qi et al. (2024), StyleCrafter Liu et al. (2023), MaskST Zhu et al. (2025), IP-Adapter Ye et al. (2023), and SADis Qin et al. (2025). However, these methods often suffer from style leakage, leading to misalignment with text prompts. StyleStudio Lei et al. (2025) employs layout to preserve the foreground, but it still struggles to handle text prompts involving multiple objects, SADis Qin et al. (2025) mainly focuses on color style transfer, but it often fails to capture enough stylistic details from the reference due to entanglement across frequency bands. We employ more complex text prompts in T2I-CompBench++ to generate the style transfer result in Table.2, which indicates that Dual-FDMcost enjoys larger CLIP-TS, CLIP-S and DINO-S than the competitors in different kinds of compositional text prompts. Notably, Our Dual-FDMcost maintains achieves the large performance gain over StyleAligned, which indicates that Dual-FDMcost outperforms by substitute the low-frequency band of the background in the style reference with that from both the foreground and background of the color reference (See Sec. 2.4). To shed more light on the advantages of our method, we further perform the generation analysis on the color style transfer results as illustrated in Fig. 7(b). More generation results for different T2I models can be seen in Sec.F of the Appendix. 3.3 Ablation Studies Discussion on Different Modules of Dual-FDM. To validate of our Dual-FDMcust, we perform the ablation study on DreamBooth/TI benchmark and DreamBooth++ with two variants: Case A: removing W¯kim W_k^im in Eq. (4); Case B: setting the same mid-frequency masks for the early and late denoising stages. Table. 3.3 reports that our Dual-FDMcust exhibits the higher similarity to the style reference (CLIP-TS and CLIP-S) than Case A, confirming that W¯kim W_k^im can enhance the background of the style reference, which in line with Sec. 2.2. Case B exhibits a suboptimal performance degradation, confirming that the mid-frequency band converge more rapidly during the early stages of the denoising process under high-level noise, while the high-frequency band evolves more gradually, confirming the intuition of Sec. 1, along with the intuition of Fig. 8. Discussion on Denoising Timesteps Between the Early and Late Stages. We conduct ablation studies by dividing the denoising process into early and late stages according to different masks of mid-frequency bands, i.e., midem^e_mid and midlm^l_mid, and perform experiments by varying the proportion of the early stage, i.e., 0.1; 0.2; 0.3; 0.4; 0.5, We also provide the result of adaptively modulating the early and late denoising stage, name Ada.. Fig. 9 reports the results for all cases, When the proportion is set to 0.1, the generated images show the highest similarity to the style reference while being least similar to the customized reference; conversely, at 0.5, the opposite trend is observed. We can also conclude that Ada. lies between the results of the 0.4 and 0.5 proportions, demonstrating that our dynamic adjustment strategy can effectively regulate the ratio of early to late stages, leading to a well-balanced performance between text alignment and image similarity, confirming the intuition of Sec. 2.5. More generation results for dual references can be seen in Sec.G of the Appendix. Figure 8: Ablation studies about the influence of different modules on generating images. Table 3: Ablation studies on the influence of different modules on image generation. Case A: removing W¯kim W_k^im in Eq. (4); Case B: using the same mid-frequency masks for early and late denoising stages. ■ denotes the best results. Method DreamBooth/TI benchmark DreamBooth++ CLIP-TC↑ CLIP-TS↑ CLIP-C↑ CLIP-S↑ CLIP-TC↑ CLIP-TS↑ CLIP-C↑ CLIP-S↑ Case A 0.5171 0.4346 0.7817 0.5215 0.5562 0.4395 0.8438 0.5459 Case B 0.5039 0.4387 0.7510 0.5254 0.5444 0.4512 0.8042 0.5547 Dual-FDMcust 0.5054 0.4468 0.7676 0.5571 0.5459 0.4526 0.8208 0.5737 Figure 9: Ablation studies about the influence of different denoising timesteps between the early and late stages on generating the images. Table 4: Ablation studies on the effect of masks filtering mid- and low-frequency bands for image generation. ■ denotes the best results. Frequency Band Case CLIP-T↑ CLIP-C↑ DINO-C↑ CLIP-S↑ DINO-S↑ Mid-Frequency thmp1=0.1,thmp2=0.9th_mp1=0.1,th_mp2=0.9 0.3313 0.8628 0.7082 0.5249 0.1239 thmp1=0.2,thmp2=0.8th_mp1=0.2,th_mp2=0.8 0.3425 0.8247 0.6168 0.5420 0.1564 thmp1=0.3,thmp2=0.7th_mp1=0.3,th_mp2=0.7 0.3467 0.7798 0.5050 0.5723 0.2177 thmp1=0.4,thmp2=0.6th_mp1=0.4,th_mp2=0.6 0.3311 0.7236 0.3909 0.6348 0.3206 Frequency Band Case CLIP-T↑ CLIP-S↑ DINO-S↑ Colorl2Color_l_2↓ Colorl1Color_l_1↓ Low-Frequency thlp=0.8th_lp=0.8 0.4304 0.5420 0.1069 5.4735 1.3168 thlp=0.6th_lp=0.6 0.4419 0.5645 0.1550 5.6028 1.3453 thlp=0.4th_lp=0.4 0.4519 0.5928 0.2200 6.0283 1.3775 thlp=0.2th_lp=0.2 0.4692 0.6372 0.3070 6.2901 1.3969 Discussion on Different Masks to filter Mid-and-Low Frequency Bands. We conduct ablation studies on Dual-FDMcust and Dual-FDMcost to investigate the effects of different frequency bands. Specifically, for the mid-frequency band, we define different thresholds (thmp1,thmp2)(th_mp1,th_mp2), where (thmp1=0.1,thmp2=0.9)(th_mp1=0.1,th_mp2=0.9) represents the widest mid-frequency range and ([0.4,0.6]CLOSE([0.4,0.6] represents a narrower range. For the low-frequency band, we vary thlp∈0.9,0.8,0.7,0.6th_lp∈\0.9,0.8,0.7,0.6\, where thlp=0.9th_lp=0.9 indicates the largest proportion of low-frequency components. Table. 3.3 reports the results under different configurations. When the mid-frequency range is the widest (i.e., [0.1,0.9][0.1,0.9]), the generated images show the highest similarity to the customized reference (as reflected by CLIP-C and DINO-C) and the lowest similarity to the style reference (CLIP-S and DINO-S). As the mid-frequency range becomes narrower (i.e., thmp1=0.4,thmp2=0.6th_mp1=0.4,th_mp2=0.6), the opposite trend is observed, suggesting that a narrower mid-frequency band enhances the representation of style features. For the low-frequency band, where lower Colorl2Color_l_2 and Colorl1Color_l_1 values indicate better color consistency, decreasing thlpth_lp from 0.8 to 0.2 (thus reducing the proportion of low-frequency bands) leads to higher similarity between the generated images and the style reference (CLIP-S, CLIP-TS, DINO-S), but worsens color consistency — the color error metrics (Colorl2Color_l_2, Colorl1Color_l_1) increase, which is consistent with Sec. 2.4. More generation results for different masks can be accessed from Sec.H of the Appendix. 4 Conclusion In this paper, we study personalized generation based on dual references - customization and color and style reference - and propose a paradigm to disentangle dual image references within frequency-aware diffusion models, dubbed Dual-FDM, for two classical personalized image generation tasks in a row: customization style transfer and color style transfer, by disentangling different frequency bands via mask strategy within frequency domain. Extensive experiments validate the superiority of Dual-FDM over the state-of-the-art diffusion models. Acknowledgments This research is sponsored by CCF-NetEase ThunderFire Innovation Research Funding (NO. CCF-Netease 202513) and supported by Institute of Advanced Medicine and Frontier Technology (2023IHM01080); The computation is completed on the HPC Platform of Hefei University of Technology. References [1] E. Aiello, U. Michieli, D. Valsesia, M. Ozay, and E. Magli (2025) DreamCache: finetuning-free lightweight personalized image generation via feature caching. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 12480–12489. Cited by: NeurIPS Paper Checklist, §1, §3.2. [2] S. Cao, B. Qian, H. Liu, Y. Wang, and M. Wang (2026) AI-t2i: aggregating-and-isolating cross-attention to diffusion models for text-to-image synthesis. IEEE Transactions on Multimedia (), p. 1–13. External Links: Document Cited by: §1. [3] M. Caron, H. Touvron, I. Misra, H. Jégou, J. Mairal, P. Bojanowski, and A. Joulin (2021) Emerging properties in self-supervised vision transformers. In Proceedings of the International Conference on Computer Vision (ICCV), Cited by: 1st item, §3.1. [4] J. Chung, S. Hyun, and J. Heo (2024) Style injection in diffusion: a training-free approach for adapting large-scale diffusion models for style transfer. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 8795–8805. Cited by: §3.2. [5] P. Esser, S. Kulal, A. Blattmann, R. Entezari, J. Müller, H. Saini, Y. Levi, D. Lorenz, A. Sauer, F. Boesel, et al. (2024) Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first international conference on machine learning, Cited by: §1. [6] R. Gal, Y. Alaluf, Y. Atzmon, O. Patashnik, A. H. Bermano, G. Chechik, and D. Cohen-Or (2022) An image is worth one word: personalizing text-to-image generation using textual inversion. arXiv preprint arXiv:2208.01618. Cited by: 1st item, NeurIPS Paper Checklist, NeurIPS Paper Checklist, §1, §3.1. [7] J. Gao, Y. Liu, Y. Sun, Y. Tang, Y. Zeng, K. Chen, and C. Zhao (2024) Styleshot: a snapshot on any style. arXiv preprint arXiv:2407.01414. Cited by: 2nd item, NeurIPS Paper Checklist, NeurIPS Paper Checklist, §1, §3.1, §3.2. [8] A. Hertz, A. Voynov, S. Fruchter, and D. Cohen-Or (2024) Style aligned image generation via shared attention. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 4775–4785. Cited by: §1, §3.2. [9] J. Ho, A. Jain, and P. Abbeel (2020) Denoising diffusion probabilistic models. Advances in Neural Information Processing Systems (NIPS) 33, p. 6840–6851. Cited by: §1. [10] K. Huang, C. Duan, K. Sun, E. Xie, Z. Li, and X. Liu (2025) T2I-CompBench++: An Enhanced and Comprehensive Benchmark for Compositional Text-to-Image Generation . IEEE Transactions on Pattern Analysis Machine Intelligence (01), p. 1–17. External Links: ISSN 1939-3539, Link Cited by: 3rd item, NeurIPS Paper Checklist, NeurIPS Paper Checklist, §3.1. [11] Q. Huang, S. Fu, J. Liu, H. Jiang, Y. Yu, and J. Song (2025) Resolving multi-condition confusion for finetuning-free personalized image generation. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 39, p. 3707–3714. Cited by: NeurIPS Paper Checklist, §1, §3.2. [12] X. Huang and S. Belongie (2017) Arbitrary style transfer in real-time with adaptive instance normalization. In Proceedings of the IEEE international conference on computer vision, p. 1501–1510. Cited by: §1. [13] J. Jeong, M. Kwon, and Y. Uh (2024) Training-free content injection using h-space in diffusion models. In Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, p. 5151–5161. Cited by: §1. [14] Z. Jiang, H. Liu, M. Wang, and Y. Wang (2026) Rare concept generation via counterfactual inference in diffusion models. In Proceedings of the 34th ACM International Conference on Multimedia, Cited by: §1. [15] N. Kumari, B. Zhang, R. Zhang, E. Shechtman, and J. Zhu (2023) Multi-concept customization of text-to-image diffusion. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 1931–1941. Cited by: §1. [16] M. Lei, X. Song, B. Zhu, H. Wang, and C. Zhang (2025) StyleStudio: text-driven style transfer with selective control of style elements. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 23443–23452. Cited by: Figure 2, Figure 2, §1, §1, §3.2. [17] D. Li, J. Li, and S. Hoi (2023) Blip-diffusion: pre-trained subject representation for controllable text-to-image generation and editing. Advances in Neural Information Processing Systems 36, p. 30146–30166. Cited by: §1. [18] P. Li, Q. Nie, Y. Chen, X. Jiang, K. Wu, Y. Lin, Y. Liu, J. Peng, C. Wang, and F. Zheng (2024) Tuning-free image customization with image and text guidance. In European Conference on Computer Vision, p. 233–250. Cited by: §1. [19] G. Liu, M. Xia, Y. Zhang, H. Chen, J. Xing, Y. Wang, X. Wang, Y. Yang, and Y. Shan (2023) Stylecrafter: enhancing stylized text-to-video generation with style adapter. arXiv preprint arXiv:2312.00330. Cited by: §1, §3.2. [20] H. Liu, Y. Wang, B. Qian, M. Wang, and Y. Rui (2024) Structure matters: tackling the semantic discrepancy in diffusion models for image inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 8038–8047. Cited by: §1. [21] H. Liu, Y. Wang, M. Wang, and Y. Rui (2022) Delving globally into texture and structure for image inpainting. In Proceedings of the 30th ACM International Conference on Multimedia, p. 1270–1278. Cited by: §1. [22] H. Liu, Y. Wang, and M. Wang (2025) One stone with two birds: a null-text-null frequency-aware diffusion models for text-guided image inpainting. Advances in Neural Information Processing Systems 38, p. 10833–10859. Cited by: §1. [23] M. Liu, D. She, J. Pang, Q. Huang, J. Ying, W. He, Y. Hou, and S. Fu (2025) TFCustom: customized image generation with time-aware frequency feature guidance. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 2714–2723. Cited by: §1. [24] J. Ma, J. Liang, C. Chen, and H. Lu (2024) Subject-diffusion: open domain personalized text-to-image generation without test-time fine-tuning. In ACM SIGGRAPH 2024 Conference Papers, p. 1–12. Cited by: §1. [25] L. Pang, J. Yin, B. Zhao, F. Wu, F. L. Wang, Q. Li, and X. Mao (2024) Attndreambooth: towards text-aligned personalized text-to-image generation. Advances in Neural Information Processing Systems 37, p. 39869–39900. Cited by: §1. [26] Y. Peng, Y. Cui, H. Tang, Z. Qi, R. Dong, J. Bai, C. Han, Z. Ge, X. Zhang, and S. Xia (2025) DreamBench++: a human-aligned benchmark for personalized image generation. In The Thirteenth International Conference on Learning Representations, Cited by: 1st item, NeurIPS Paper Checklist, NeurIPS Paper Checklist, §3.1. [27] D. Podell, Z. English, K. Lacey, A. Blattmann, T. Dockhorn, J. Müller, J. Penna, and R. Rombach (2023) Sdxl: improving latent diffusion models for high-resolution image synthesis. arXiv preprint arXiv:2307.01952. Cited by: Appendix D, §1, §3.1. [28] T. Qi, S. Fang, Y. Wu, H. Xie, J. Liu, L. Chen, Q. He, and Y. Zhang (2024) Deadiff: an efficient stylization diffusion model with disentangled representations. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 8693–8702. Cited by: §1, §3.2. [29] B. Qian, Y. Wang, R. Hong, and M. Wang (2023) Adaptive data-free quantization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 7960–7968. Cited by: §1. [30] B. Qian, Y. Wang, H. Yin, R. Hong, and M. Wang (2022) Switchable online knowledge distillation. In European Conference on Computer Vision, p. 449–466. Cited by: §1. [31] J. Qin, S. Li, A. Gomez-Villa, S. Yang, Y. Wang, K. Wang, and J. van de Weijer (2025) Free-lunch color-texture disentanglement for stylized image generation. Advances in Neural Information Processing Systems. Cited by: §1, §3.2. [32] A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, et al. (2021) Learning transferable visual models from natural language supervision. In International conference on machine learning, p. 8748–8763. Cited by: §3.1. [33] R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer (2022) High-resolution image synthesis with latent diffusion models. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Vol. , p. 10674–10685. External Links: Document Cited by: §1. [34] L. Rout, Y. Chen, N. Ruiz, A. Kumar, C. Caramanis, S. Shakkottai, and W. Chu (2025) RB-modulation: training-free stylization using reference-based modulation. In The Thirteenth International Conference on Learning Representations, Cited by: §1. [35] N. Ruiz, Y. Li, V. Jampani, Y. Pritch, M. Rubinstein, and K. Aberman (2023) Dreambooth: fine tuning text-to-image diffusion models for subject-driven generation. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 22500–22510. Cited by: 1st item, NeurIPS Paper Checklist, NeurIPS Paper Checklist, §1, §3.1. [36] K. Sohn, N. Ruiz, K. Lee, D. C. Chin, I. Blok, H. Chang, J. Barber, L. Jiang, G. Entis, Y. Li, et al. (2023) Styledrop: text-to-image generation in any style. arXiv preprint arXiv:2306.00983. Cited by: §1. [37] H. Wang, M. Spinelli, Q. Wang, X. Bai, Z. Qin, and A. Chen (2024) Instantstyle: free lunch towards style-preserving in text-to-image generation. arXiv preprint arXiv:2404.02733. Cited by: Figure 2, Figure 2, §1, §1, §3.2. [38] H. Wang, P. Xing, R. Huang, H. Ai, Q. Wang, and X. Bai (2024) Instantstyle-plus: style transfer with content-preserving in text-to-image generation. arXiv preprint arXiv:2407.00788. Cited by: 2nd item, NeurIPS Paper Checklist, NeurIPS Paper Checklist, §1, §3.1. [39] X. Wang, S. Fu, Q. Huang, W. He, and H. Jiang (2025) MS-diffusion: multi-subject zero-shot image personalization with layout guidance. In The Thirteenth International Conference on Learning Representations, Cited by: Figure 2, Figure 2, §1, §3.2, §3.2. [40] Y. Wang, J. Peng, H. Wang, and M. Wang (2022) Progressive learning with multi-scale attention network for cross-domain vehicle re-identification. Science China Information Sciences 65 (6), p. 160103. Cited by: §1. [41] Y. Wang, B. Qian, H. Liu, Y. Rui, and M. Wang (2024) Unpacking the gap box against data-free knowledge distillation. IEEE Transactions on Pattern Analysis and Machine Intelligence 46 (9), p. 6280–6291. Cited by: §1. [42] X. Wu, Y. Hao, K. Sun, Y. Chen, F. Zhu, R. Zhao, and H. Li (2023) Human preference score v2: a solid benchmark for evaluating human preferences of text-to-image synthesis. arXiv preprint arXiv:2306.09341. Cited by: 2nd item, §3.1. [43] G. Xiao, T. Yin, W. T. Freeman, F. Durand, and S. Han (2025) Fastcomposer: tuning-free multi-subject image generation with localized attention. International Journal of Computer Vision 133 (3), p. 1175–1194. Cited by: §1. [44] P. Xing, H. Wang, Y. Sun, Q. Wang, X. Bai, H. Ai, R. Huang, and Z. Li (2024) Csgo: content-style composition in text-to-image generation. arXiv preprint arXiv:2408.16766. Cited by: §1, §3.2, §3.2, §3.2. [45] J. Xu, X. Liu, Y. Wu, Y. Tong, Q. Li, M. Ding, J. Tang, and Y. Dong (2023) ImageReward: learning and evaluating human preferences for text-to-image generation. In Proceedings of the 37th International Conference on Neural Information Processing Systems, p. 15903–15935. Cited by: 2nd item, §3.1. [46] R. Xu, W. Xi, X. Wang, Y. Mao, and Z. Cheng (2025) StyleSSP: sampling startpoint enhancement for training-free diffusion-based method for style transfer. In Proceedings of the Computer Vision and Pattern Recognition Conference, p. 18260–18269. Cited by: §3.2. [47] H. Ye, J. Zhang, S. Liu, X. Han, and W. Yang (2023) Ip-adapter: text compatible image prompt adapter for text-to-image diffusion models. arXiv preprint arXiv:2308.06721. Cited by: Appendix D, NeurIPS Paper Checklist, Figure 2, Figure 2, §1, §3.1, §3.2, §3.2. [48] Y. Zeng, V. M. Patel, H. Wang, X. Huang, T. Wang, M. Liu, and Y. Balaji (2024) Jedi: joint-image diffusion models for finetuning-free personalized text-to-image generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 6786–6795. Cited by: §1. [49] Y. Zhang, Y. Song, J. Liu, R. Wang, J. Yu, H. Tang, H. Li, X. Tang, Y. Hu, H. Pan, et al. (2024) Ssr-encoder: encoding selective subject representation for subject-driven generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 8069–8078. Cited by: NeurIPS Paper Checklist, §1, §3.2. [50] L. Zhu, X. Wang, C. Zhou, Q. Gu, and N. Ye (2025) Less is more: masking elements in image condition features avoids content leakages in style transfer diffusion models. In The Thirteenth International Conference on Learning Representations, Cited by: §1, §3.2. Due to page limitation of the mainbody, the supplementary material offers further more quantitative results and qualitative results with higher resolution, which are summarized below: • The frequency variation during the denoising process, as mentioned in Sec. 1 of the mainbody (Sec. A) • More intuitions for the rationale between WkimW_k^im and W¯kim W_k^im, as mentioned in Sec.2.2 of the mainbody (Sec. B) • The algorithm of customization style transfer and color style transfer, as mentioned in Sec.2 of the mainbody (Sec. C) • More details about the implementation details, as mentioned in Sec.3.1 of the mainbody. (Sec. D) • More Analysis of Computational Efficiency in the Sec.3.2 of the Mainbody. (Sec. E) • More generation results of different T2I models, as mentioned in Sec.3.2 of the mainbody. (Sec. F) • Additional generation results for the ablation study about the denoising timesteps between the early and late stages, as mentioned in Sec.3.3 of the mainbody. (Sec. G) • Additional generation results for the ablation study about the different masks to filter mid-and-low frequency bands, as mentioned in Sec.3.3 of the mainbody. (Sec. H) • We list the limitations and broader impacts (Sec.I). • We provide the code for reproducibility. (Sec. J) Appendix A Frequency Variation During the Denoising Process As mentioned in Sec. 1 of the mainbody, We claim that during the denoising process, images recover low-and-mid frequency bands first, followed by high-frequency details. To validate the claims via frequency analysis, we experimentally measure the changes in both low- and high-frequency bands at different timesteps during the denoising process. The results are reported in Tab. 5: • Low-frequency information. We adopt Mean Gray, Mean HSV-V, and Mean Luma, which primarily capture global brightness and luminance structure. These metrics quantify the stability of the low-frequency bands throughout the denoising process. • High-frequency information. We use Average Gradient, Variance, and LBP Variance, which are sensitive to edge sharpness, local contrast, and texture complexity. These metrics reflect the recovery of high-frequency details. The low-frequency metrics remain nearly unchanged from the beginning (e.g., Mean Gray varies only slightly from 119.5 to 120.9). In contrast, the high-frequency metrics increase steadily throughout the reverse process (e.g., Avg Gradient rises from 59.01 to 233.88), confirming that high-frequency details are refined gradually during denoising. This is consistent with Sec. 1 of the mainbody. Table 5: The high-and-low frequency bands across different timesteps during the denoising process. ■ denotes the best result. Timesteps Mean Gray Mean HSV-V Mean Luma Avg Gradient Variance LBP Variance 0 120.9 136.0 120.9 59.01 3955.0 4.45 5 120.8 134.6 120.8 75.07 4098.3 3.71 10 120.7 133.5 120.7 101.58 4378.9 3.25 15 120.7 133.7 120.7 133.26 4790.8 3.07 20 120.9 135.1 120.9 164.23 5274.8 3.07 25 121.2 137.3 121.2 190.49 5743.9 3.15 30 121.4 139.2 121.4 209.84 6099.5 3.25 35 121.1 140.0 121.1 222.03 6291.1 3.32 40 120.5 140.1 120.5 228.75 6363.9 3.35 45 119.9 139.8 119.9 232.19 6394.8 3.38 50 119.5 139.5 119.5 233.88 6419.5 3.40 Appendix B More intuitions for the rationale between WkimW_k^im and W¯kim W_k^im As indicated in Sec.2.2 of the mainbody, we provide more intuitions for the rationale between WkimW_k^im and W¯kim W_k^im. To verify the intuition that larger σi _i tends to correspond to the foreground, whereas smaller σi _i is more likely associated with the background, as shown in Tab. 6, we utilize the text prompts for both the foreground and background to generate their corresponding images using a text-to-image model. These generated images are then used as image references in our Dual-FDM framework, where we disentangle the foreground using (WkimW_k^im) and the background using (W¯kim W_k^im). For quantitative evaluation, we compute the similarity between the disentangled images and the text prompts for foreground or background in Tab. 6 using the CLIP model. In addition, we generate images from text prompts for the foreground or background using a text-to-image model, and then measure the similarity between the disentangled images and these generated images. Both CLIP and DINO models are employed to assess the image-level similarity, providing a comprehensive evaluation of the consistency of the foreground–background disentanglement. As illustrated in the left of Tab. 7, WkimW_k^im is able to disentangle more foreground information than W¯kim W_k^im, as evidenced by the higher scores in CLIP-T, CLIP-I, and DINO-I, whereas W¯kim W_k^im is more effective at disentangling background information, as shown in the right of Tab. 7. We also conduct additional ablation studies to investigate different strategies for suppressing foreground components and enhancing background components in the singular values. Specifically, we consider the hard suppression approach σtop-n%=0 _top-n\%=0, which denotes setting the top n%n\% of singular values to zero, and the soft suppression approach σe−σ e^-σ, which represents a reweighting scheme that attenuates larger singular values while preserving smaller ones. Tab. 8 reports the results for all cases: overall, all the adjustment strategies show no significant difference in their ability to extract background information (measured by CLIP-T and CLIP-I mean scores); the main variations are observed in the fluctuations across different images (measured by CLIP-T and CLIP-I std scores). Our method demonstrates stable background disentanglement across diverse images. Setting different proportions of the top 10–50% of singular values to zero does not yield a clear improvement in performance. This is because the hard partitioning of foreground and background components makes it difficult to adaptively determine the proportion of foreground and background for each image, leading to limited performance gains and high variability across images. Similarly, the soft suppression approach σe−σ e^-σ fails to effectively disentangle background for each image. While it adaptively suppresses foreground components, it does not explicitly enhance background components. As a result, the contribution of background components in the cross-attention remains limited, leading to incomplete background disentanglement. In contrast, our reciprocal operation 1/σ1/σ not only weakens the dominant foreground components but also actively amplifies the smaller singular values corresponding to background. This balanced adjustment ensures both suppression of foreground and enhancement of background, which results in more stable and consistent background disentanglement across images. We further analysis the singular values in the different cross-attention layers based on the singular value decomposition (SVD) on WkimW_k^im (Eq.(3) of the mainbody) and W¯kim W_k^im (Eq.(4) of the mainbody). For the singular value matrices obtained from different layers, we compute the proportion of the top n% singular values relative to the sum of all singular values, named ∑σtop-n%Σ _top-n\%. We also take the reciprocal of each singular value and perform the same calculation. In both cases, the proportions are evaluated for n∈2,5,10,20n∈\2,5,10,20\. Tab. 9 indicates that: under the same spatial resolution, the proportions of the top n% singular values exhibit highly consistent patterns across different layers. This consistency supports our claim that larger σi _i tends to correspond to foreground-dominant components, whereas smaller σi _i is more closely associated with background information (Sec. 2.2 of the mainbody), which confirms universally across layers. Furthermore, after inverting the singular values, the resulting distributions across layers also display similar behaviors, indicating that our style-disentangled matrices W¯kim W_k^im effectively amplify the relatively weak background components. Table 6: The examples of text prompts for both foreground and background, foreground, and background to discussion intuitions for the rationale between WkimW_k^im and W¯kim W_k^im. ID Both foreground and background Foreground Background 1 A snow leopard is leaping across a fractured frozen expanse. a snow leopard the fractured frozen expanse 2 A giant squid is drifting in a radiant deep chasm. a giant squid the radiant deep chasm 3 A phoenix is glowing over an age-worn conflict field. the phoenix the age-worn conflict field 4 A red-crowned crane is gliding through a crimson-lit horizon. a red-crowned crane the crimson horizon 5 A polar bear is sitting beside a fractured icy dome-like form. a polar bear the fractured icy dome 6 A chameleon is blending into a glare-soaked luminous corridor. a chameleon the luminous corridor 7 A mountain gorilla is climbing an overrun observational platform. a mountain gorilla the overrun platform 8 A humpback whale is singing beneath drifting frozen masses. a humpback whale the drifting frozen masses 9 A komodo dragon is sunbathing on dark volcanic granules. a komodo dragon the volcanic granules 10 A peregrine falcon is landing on a ruined high outpost. a peregrine falcon the ruined high outpost Table 7: Quantitative comparisons of foreground–background disentanglement across different images via WkimW_k^im and W¯kim W_k^im. Method Foreground Background CLIP-T↑ CLIP-I↑ DINO-I↑ CLIP-T↑ CLIP-I↑ DINO-I↑ WkimW_k^im 0.3115 0.8228 0.5774 0.2303 0.6157 0.1353 W¯kim W_k^im 0.2629 0.7090 0.2591 0.2486 0.6421 0.1966 Table 8: Ablation studies are conducted to evaluate three strategies for manipulating singular values: σtop-n%=0 _top-n\%=0, which removes the largest n%n\% of singular values by setting them to zero; σe−σ e^-σ, a reweighting scheme that suppresses larger singular values while retaining smaller ones; and ours (1σ 1σ), which inversely scales singular values. These strategies are examined to assess their effectiveness in suppressing foreground components and enhancing background components. Method CLIP-T-Mean↑ CLIP-T-Std↓ CLIP-I-Mean↑ CLIP-I-Std↓ σtop-10%=0 _top-10\%=0 0.2469 (-0.68%) 0.0352 (+13.5%) 0.6450 (+0.45%) 0.0867 (+8.36%) σtop-20%=0 _top-20\%=0 0.2496 (+0.40%) 0.0344 (+11.0%) 0.6433 (+0.19%) 0.0806 (+0.62%) σtop-30%=0 _top-30\%=0 0.2500 (+0.56%) 0.0332 (+7.10%) 0.6401 (-0.31%) 0.0842 (+5.11%) σtop-40%=0 _top-40\%=0 0.2495 (+0.36%) 0.0334 (+7.74%) 0.6382 (-0.30%) 0.0826 (+3.11%) σtop-50%=0 _top-50\%=0 0.2499 (+0.52%) 0.0322 (+3.87%) 0.6416 (-0.08%) 0.0868 (+8.36%) σe−σ e^-σ 0.2498 (+0.49%) 0.0330 (+6.45%) 0.6401 (-0.31%) 0.0883 (+10.2%) Ours (1σ 1σ) 0.2486 0.0310 0.6421 0.0801 Table 9: Summation of the top-k%k\% singular values and their reciprocal counterparts across different cross-attention layers. Layer Name Res. ∑σtop-2%Σ _top-2\% ∑σtop-5%Σ _top-5\% ∑σtop-10%Σ _top-10\% ∑σtop-20%Σ _top-20\% ∑1σtop-2%Σ 1σ_top-2\% ∑1σtop-5%Σ 1σ_top-5\% ∑1σtop-10%Σ 1σ_top-10\% ∑1σtop-20%Σ 1σ_top-20\% down_blocks.1.0.0 64×64 0.127 0.259 0.408 0.606 0.520 0.724 0.849 0.934 down_blocks.1.1.1 64×64 0.147 0.279 0.431 0.635 0.562 0.756 0.869 0.944 down_blocks.2.0.0 32×32 0.197 0.360 0.544 0.763 0.525 0.715 0.838 0.927 down_blocks.2.1.9 32×32 0.228 0.392 0.562 0.751 0.552 0.733 0.849 0.932 mid_block.0.0 32×32 0.208 0.384 0.573 0.790 0.533 0.721 0.842 0.929 mid_block.0.9 32×32 0.234 0.381 0.536 0.719 0.563 0.739 0.853 0.934 up_blocks.0.0.0 32×32 0.179 0.344 0.532 0.755 0.518 0.710 0.835 0.926 up_blocks.0.2.9 32×32 0.285 0.489 0.652 0.783 0.556 0.735 0.850 0.933 up_blocks.1.0.0 64×64 0.130 0.275 0.437 0.641 0.537 0.739 0.859 0.939 up_blocks.1.2.1 64×64 0.150 0.292 0.448 0.647 0.527 0.733 0.856 0.938 Appendix C The algorithm of customization style transfer and color style transfer Due to the page limitation, we summarize the process of customization style transfer in Algorithm 1 and color style transfer in Algorithm 2. To tackle the text-alignment issue in personalized image generation, in this paper, we study personalized generation based on dual references - customization and color and style reference - and propose a paradigm to disentangle these Dual image references within Frequency-aware Diffusion Models, dubbed Dual-FDM, for two classical personalized image generation tasks in a row: customization style transfer and color style transfer, by disentangling different frequency bands via mask strategy within frequency domain. For customization style transfer, we replace the mid-frequency band of the background in the style reference with that from the foreground of the customized reference. For color style transfer, we substitute the low-frequency band of the background in the style reference with that from both the foreground and background of the color reference. Both the substituted frequency bands are used as the key and value to reconstruct the query foreground and background of the denoised personalized image. For each task, we further compute the ratio of information entropy of the substituted frequency bands to adaptively modulate the denoising timesteps between the early and late stages. Algorithm 1 Dual-FDMcust: Customization Style Transfer Input: the random Gaussian noise zTz_T, the text prompts CteC_te, the customized reference CcuC_cu, the style reference CstC_st, the denoising network ϵθ _θ, the mid-frequency masks in the early denoising stages midem^e_mid, the mid-frequency masks in the late denoising stages midlm^l_mid, the timestep T and the decoder D Output: The personalized generated image IcustI_cust 1 Calculating KimcustmmideK^m^e_mid_im_cust and VimcustmmideV^m^e_mid_im_cust conditioned on midem^e_mid via Eq.(7) of the mainbody for the early stage; 2 Calculating KimcustmmidlK^m^l_mid_im_cust and VimcustmmidlV^m^l_mid_im_cust conditioned on midlm^l_mid via Eq.(7) of the mainbody for the late stage; 3 Calculating the threshold Δcust _cust conditioned on KimcustmmideK^m^e_mid_im_cust and KimcustmmidlK^m^l_mid_im_cust via Eq.(12) of the mainbody; 4 for t = T, …, 11+Δcust 11+ _custT do 5 Calculating zt−1z_t-1 from ztz_t is performed in the decoupled cross-attention (Eq.(1) and Eq.(6) of the mainbody) of ϵθ(Cte,Ccu,Cst) _θ(C_te,C_cu,C_st), the substituted frequency bands are used as the key KimcustmmideK^m^e_mid_im_cust and value VimcustmmideV^m^e_mid_im_cust to reconstruct the query foreground and background of the denoised personalized image; 6 end for 7 for t = 11+Δcust 11+ _custT, …, 0 do 8 Calculating zt−1z_t-1 from ztz_t is performed in the decoupled cross-attention (Eq.(1) and Eq.(6) of the mainbody) of ϵθ(Cte,Ccu,Cst) _θ(C_te,C_cu,C_st), the substituted frequency bands are used as the key KimcustmmidlK^m^l_mid_im_cust and value VimcustmmidlV^m^l_mid_im_cust to reconstruct the query foreground and background of the denoised personalized image; 9 end for 10 return Output the personalized generated image Icust=(z0)I_cust=D(z_0) Algorithm 2 Dual-FDMcost: Color Style Transfer Input: the random Gaussian noise zTz_T, the text prompts CteC_te, the color reference CcoC_co, the style reference CstC_st, the denoising network ϵθ _θ, the low-frequency masks in the early denoising stages mlowem^e_low, the low-frequency masks in the late denoising stages mlowlm^l_low, the timestep T and the decoder D Output: The personalized generated image IcostI_cost 1 Calculating KimcostmloweK^m^e_low_im_cost and VimcostmloweV^m^e_low_im_cost conditioned on mlowem^e_low via Eq.(10) of the mainbody for the early stage; 2 Calculating KimcostmlowlK^m^l_low_im_cost and VimcostmlowlV^m^l_low_im_cost conditioned on mlowlm^l_low via Eq.(10) of the mainbody for the late stage; 3 Calculating the threshold Δcost _cost conditioned on KimcostmloweK^m^e_low_im_cost and KimcostmlowlK^m^l_low_im_cost via Eq.(12) of the mainbody; 4 for t = T, …, 11+Δcost 11+ _costT do 5 Calculating zt−1z_t-1 from ztz_t is performed in the decoupled cross-attention (Eq.(1) and Eq.(9) of the mainbody) of ϵθ(Cte,Cco,Cst) _θ(C_te,C_co,C_st), the substituted frequency bands are used as the key KimcostmloweK^m^e_low_im_cost and value VimcostmloweV^m^e_low_im_cost to reconstruct the query foreground and background of the denoised personalized image; 6 end for 7 for t = 11+Δcost 11+ _costT, …, 0 do 8 Calculating zt−1z_t-1 from ztz_t is performed in the decoupled cross-attention (Eq.(1) and Eq.(9) of the mainbody) of ϵθ(Cte,Cco,Cst) _θ(C_te,C_co,C_st), the substituted frequency bands are used as the key KimcostmlowlK^m^l_low_im_cost and value VimcostmlowlV^m^l_low_im_cost to reconstruct the query foreground and background of the denoised personalized image; 9 end for 10 return Output the personalized generated image Icost=(z0)I_cost=D(z_0) Appendix D More details about the implementation details As mentioned in Sec.3.1 of the mainbody, due to page limitation, we further provide more details about the implementation details as follows: Datasets. We evaluate Dual-FDM on combined datasets covering image customization, style transfer, and text-to-image (T2I) generation: • Image Customization: 1) DreamBench Ruiz et al. (2023) + TI benchmark Gal et al. (2022): a well-known single-object personalized image generation dataset comprising 38 subjects, covering unique objects and pets such as backpacks, stuffed animals, dogs, cats, sunglasses, and cartoons, with each subject associated with 25 prompts; 2) DreamBench++ Peng et al. (2025): the latest human-aligned personalized image generation benchmark, comprising 65 object classes, 20 human classes, and 45 animal classes. For each image, 9 text prompts are generated by GPT-4o to span a range of difficulty levels. • Style Transfer: 1) IS-Plus benchmark Wang et al. (2024b): a style evaluation benchmark containing 104 images across diverse styles; 2) StyleBench Gao et al. (2024): a style evaluation benchmark covering 73 distinct styles, ranging from paintings, flat illustrations, and 3D rendering to sculptures with varying materials. For each style, 5–7 distinct images with variations are collected. In total, StyleBench contains 490 images across diverse styles. • T2I generation: T2I-CompBench++ Huang et al. (2025a): an enhanced benchmark for compositional text-to-image generation. T2I-CompBench++ comprises 8,000 compositional text prompts categorized into four primary groups: attribute binding, object relationships, generative numeracy, and complex compositions. These are further divided into eight sub-categories. Based on the above, we construct several combined datasets for different joint tasks: 1) for customization style transfer (small-scale), we concatenate DreamBench and TI benchmark as the customization references, and use IS-Plus benchmark as the style references; 2) for customization style transfer (large-scale), we take DreamBench++ as the customization references and StyleBench as the style references; 3) for color style transfer, we combine InstantStyle-Plus and StyleBench as the color and style references, while text prompts are provided by T2I-CompBench++ for stylized image generation. Metrics. We consider 9 evaluation metrics from three aspects to evaluate our Dual-FDM: • Text and Image Reference Alignment: we follow previous methods and adopt two representative models, CLIP22 2 Since CLIP is trained on real image–text (T) pairs, it is particularly sensitive to image style, which may distort text–image similarity. To alleviate this, we propose CLIP-TS by measuring similarity with style reference (S), thereby reducing the style influence . In addition, the generated background aligned with text prompts can interfere with the customized reference (C); hence we propose CLIP-TC, which incorporates text prompts to suppress such interference. We also provide CLIP-T (image–text similarity) and CLIP-CS (image–image reference similarity) for comparison. and DINO Caron et al. (2021), for evaluation. Specifically, CLIP-T measures the similarity between generated images and the given text prompts; CLIP-TC and CLIP-TS evaluate the similarity between generated images and customized (or style) reference images while jointly considering the text prompts; finally, CLIP-CS assesses the similarity between generated images and dual-reference images. • Image Generation Quality: we employ Image Reward (IR) Xu et al. (2023) to model human preferences for image quality and HPS V2 Wu et al. (2023) that combines perceptual studies with deep learning for comprehensive visual and semantic assessment. Both metrics measure the semantics consistency of the high-quality image generation aligned with human standards; • Time Complexity: We employ the inference time (Times) and Video Random Access Memory (VRAM) consumption to evaluate the computational efficiency and memory footprint of different methods during image generation. Model Configuration. Our main experiments are conducted on the pre-trained IP-Adapter-Plus Ye et al. (2023) with SDXL Podell et al. (2023) model as the text-to-image diffusion model, configured with 30 sampling steps and and a fixed random seed of 42, to ensure high-quality outputs. We use the recommended hyper-parameters for each method across all images to ensure fair comparison. All experiments are implemented in PyTorch and conducted on a NVIDIA RTX 4090 GPU. Appendix E More Analysis of Computational Efficiency in the Sec.3.2 of the Mainbody As indiacted in the mainbody, we further provide the inference time and VRAM consumption of our proposed Dual-FDMcust model, and compared it with several representative baselines, which are reported in Table. 10. The inference time of our Dual-FDMcust is 10.26s. This demonstrates superior efficiency compared with state-of-the-art methods. Our Dual-FDMcust also achieves the the lowest result in VRAM consumption. Specifically, SSR-Encoder adopts SD v1.5 and Dreamcache is designed for lightweight usage, so both consume less VRAM. Table 10: Efficiency comparisons between Dual-FDMcust and other diffusion-based personalization models. IP-Adapter Arxiv’23 SSR-Encoder CVPR’24 MIP-Adapter AAAI’25 Dreamcache CVPR’25 PatchDPO CVPR’25 MS-Diffusion ICLR’25 Dual-FDMcust Baseline + StyleID (CVPR’24) Time(s) ↓ 13.30 11.03 15.85 12.17 13.11 14.68 10.26 VRAM(GB) ↓ 16.92 13.88 16.82 13.80 16.92 16.85 12.48 Figure 10: Generation results of different T2I models. Appendix F More generation results of different T2I Models As mentioned in Sec.3.2 of the mainbody, we further report the generated results of several representative text-to-image (T2I) models in Fig. 10, including different finetuned variants of Stable Diffusion XL (SDXL), namely RealisticVisionXL and DreamshaperXL, as well as the DiT-based approach HunyuanDiT. These three models exhibit distinct characteristics due to differences in their training datasets: • RealisticVisionXL: fine-tuned on high-quality, photography-oriented datasets, emphasizing realism, lighting, and fine-grained textures, making it particularly effective for photorealistic inpainting tasks. As shown in Fig. 10(a), the customized contents and the style can be well balanced. • DreamshaperXL: incorporates a mixture of artistic and stylized datasets, leading to visually rich outputs with enhanced creativity and aesthetic variations, though sometimes at the expense of strict semantic fidelity to the input prompts. As shown in Fig. 10, the personalized images emphasize style more strongly. • HunyuanDiT: trained on large-scale, multi-domain datasets covering both natural scenes and human-centric images, which provides stronger generalization ability across diverse prompts and better semantic alignment. Benefiting from its realism-oriented training, it is more favorable to the image customization of our Dual-FDM, as illustrated in Fig. 10(a). Figure 11: Ablation studies about the influence of different denoising timesteps between the early and late stages on generating the personalized images. Appendix G Additional generation results for the ablation study about the denoising timesteps between the early and late stages As mentioned in Sec.3.3 of the mainbody, due to page limitation, we further provide more generation results to analyse the influence of the denoising timesteps between the early and late stages on generating the personalized images. Fig. 11 reports that when the proportion is set to 0.1, the generated images show the highest similarity to the style reference while being least similar to the customized reference; conversely, at 0.5, the opposite trend is observed. We can also conclude that our Adapting (Eq.(12) of the mainbody) lies between the results of the 0.4 and 0.5 proportions, demonstrating that our dynamic adjustment strategy can effectively regulate the ratio of early to late stages, leading to a well-balanced performance in terms of both text alignment and dual-image similarity, confirming the intuition of Sec.2.5 of the mainbody. Appendix H Additional generation results for the ablation study about the different masks to filter mid-and-low frequency bands As mentioned in Sec.3.3 of the mainbody, due to page limitation, we further provide more generation results to analyse the the different masks to filter mid-and-low frequency bands on generating the personalized images. For the mid-frequency band, we define different thresholds (thmp1,thmp2)(th_mp1,th_mp2), where (thmp1=0.1,thmp2=0.9)(th_mp1=0.1,th_mp2=0.9) represents the widest mid-frequency range and (thmp1=0.4,thmp2=0.6)(th_mp1=0.4,th_mp2=0.6) represents a narrower range. Fig. 12 suggests the following: when the mid-frequency range is the widest (i.e., thmp1=0.1,thmp2=0.9th_mp1=0.1,th_mp2=0.9), the generated images show the highest similarity to the customized reference and the lowest similarity to the style reference. As the mid-frequency range becomes narrower (i.e., thmp1=0.4,thmp2=0.6th_mp1=0.4,th_mp2=0.6), the opposite trend is observed, suggesting that a narrower mid-frequency band enhances the representation of style features. For the low-frequency band, we vary thlp∈0.8,0.6,0.4,0.2th_lp∈\0.8,0.6,0.4,0.2\, where thlp=0.8th_lp=0.8 indicates the largest proportion of low-frequency components. Decreasing thlpth_lp from 0.8 to 0.2 (thus reducing the proportion of low-frequency bands) leads to higher similarity between the generated images and the style reference, but worsens color consistency, as illustrated in Fig. 13, which is consistent with Sec.2.4 of the mainbody. Figure 12: Ablation studies about the influence of the different masks to filter mid-frequency band on generating the personalized images. Figure 13: Ablation studies about the influence of the different masks to filter low-frequency band on generating the personalized images. Appendix I Limitations and Broader Impacts Limitations. As illustrated in Fig. 14, for customization style transfer, when the style reference contains extremely sparse patterns (e.g., line-based structures), the mid-frequency customized cues may attenuate high-frequency style details, leading to incomplete style information. We plan to address this issue in our future work. Broader Impacts. Our Dual-FDM can generate the personalized image according to the text prompt and reference images, our method may raise certain ethical concerns. The personalized image could potentially be misused in the creation of misleading information. We strongly advocate for the establishment of clear accountability in the use of such technologies, along with enhanced legal and technical oversight, to ensure they are applied responsibly. Figure 14: Failure examples for customization style transfer Appendix J Code We implement the proposed Dual-FDM in pytorch framework under the running environment as: python3.6, pytorch1.7 and cuda10.0. The codes are available in the Supplementary Material with the folder name Dual-FDM. NeurIPS Paper Checklist 1. Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: We confirm the main claims reflect the paper’s contributions and scope. Guidelines: • The answer [N/A] means that the abstract and introduction do not include the claims made in the paper. • The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A [No] or [N/A] answer to this question will not be perceived well by the reviewers. • The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. • It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. 2. Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: The paper have conducted a discussion of limiations in the Appendix. Guidelines: • The answer [N/A] means that the paper has no limitation while the answer [No] means that the paper has limitations, but those are not discussed in the paper. • The authors are encouraged to create a separate “Limitations” section in their paper. • The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. • The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. • The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. • The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. • If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. • While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren’t acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. 3. Theory assumptions and proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [N/A] Justification: We did not include theoretical results in our paper. Guidelines: • The answer [N/A] means that the paper does not include theoretical results. • All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. • All assumptions should be clearly stated or referenced in the statement of any theorems. • The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. • Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. • Theorems and Lemmas that the proof relies upon should be properly referenced. 4. Experimental result reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? Answer: [Yes] Justification: We have described the model details in the implementation details in Sec. D of the Appendix and carefully described the experimental metrics in the experimental chapter. The information we provide is sufficient and detailed for replication purposes. Guidelines: • The answer [N/A] means that the paper does not include experiments. • If the paper includes experiments, a [No] answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. • If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. • Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general. releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. • While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example (a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. (b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. (c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). (d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. 5. Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: We make a detailed description of the implementation details to ensure that they are repeatable and use publicly available personalized image generation datasets. We provide our code in the Appendix. Guidelines: • The answer [N/A] means that paper does not include experiments requiring code. • Please see the NeurIPS code and data submission guidelines (https://neurips.c/public/guides/CodeSubmissionPolicy) for more details. • While we encourage the release of code and data, we understand that this might not be possible, so [No] is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). • The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://neurips.c/public/guides/CodeSubmissionPolicy) for more details. • The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. • The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. • At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). • Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. 6. Experimental setting/details Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer) necessary to understand the results? Answer: [Yes] Justification: We have carried out a detailed narration in the implementation details in Sec. D of the Appendix. Guidelines: • The answer [N/A] means that the paper does not include experiments. • The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. • The full details can be provided either with the code, in appendix, or as supplemental material. 7. Experiment statistical significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [N/A] . Justification: Based on our experimental experience, the reproducibility of the experiments involved in this work is high, with results that are replicable and stable, rather than simply reporting the highest outcomes. Additionally, previous related work Ye et al. (2023); Aiello et al. (2025); Huang et al. (2025b); Zhang et al. (2024); Huang et al. (2025b) has also not reported error bars. We thus do not run the statistical significance test. Guidelines: • The answer [N/A] means that the paper does not include experiments. • The authors should answer [Yes] if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. • The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). • The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) • The assumptions made should be given (e.g., Normally distributed errors). • It should be clear whether the error bar is the standard deviation or the standard error of the mean. • It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a 96% CI, if the hypothesis of Normality of errors is not verified. • For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g., negative error rates). • If error bars are reported in tables or plots, the authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. 8. Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: We state this detailed information of computer resources in Sec. 3.1 of the Mainbody . Guidelines: • The answer [N/A] means that the paper does not include experiments. • The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. • The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. • The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn’t make it into the paper). 9. Code of ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.c/public/EthicsGuidelines? Answer: [Yes] Justification: We confirm that the research involved in the article complies with the NeurIPS Code of Ethics in all respects. Guidelines: • The answer [N/A] means that the authors have not reviewed the NeurIPS Code of Ethics. • If the authors answer [No] , they should explain the special circumstances that require a deviation from the Code of Ethics. • The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). 10. Broader impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [Yes] Justification: The paper have conducted a discussion of broader impacts in the Appendix. Guidelines: • The answer [N/A] means that there is no societal impact of the work performed. • If the authors answer [N/A] or [No] , they should explain why their work has no societal impact or why the paper does not address societal impact. • Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. • The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate Deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. • The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. • If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). 11. Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pre-trained language models, image generators, or scraped datasets)? Answer: [N/A] Justification: We utilize the publicly available personalized image generation dataset Ruiz et al. (2023); Wang et al. (2024b); Gal et al. (2022); Peng et al. (2025); Gao et al. (2024); Huang et al. (2025a) and the pretrained generation models Stable Diffusion XL, there is no relevant description we will set up safeguards when we release the code of Dual-FDM. Guidelines: • The answer [N/A] means that the paper poses no such risks. • Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. • Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. • We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. 12. Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes] Justification: We utilize the publicly available personalized image generation dataset Ruiz et al. (2023); Wang et al. (2024b); Gal et al. (2022); Peng et al. (2025); Gao et al. (2024); Huang et al. (2025a) and the pretrained generation models Stable Diffusion XL, both of them state the dataset and model license. Guidelines: • The answer [N/A] means that the paper does not use existing assets. • The authors should cite the original paper that produced the code package or dataset. • The authors should state which version of the asset is used and, if possible, include a URL. • The name of the license (e.g., C-BY 4.0) should be included for each asset. • For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. • If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. • For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. • If this information is not available online, the authors are encouraged to reach out to the asset’s creators. 13. New assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [N/A] Justification: We do not introduce new assets in the paper. Guidelines: • The answer [N/A] means that the paper does not release new assets. • Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. • The paper should discuss whether and how consent was obtained from people whose asset is used. • At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. 14. Crowdsourcing and research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? Answer: [N/A] Justification: The paper does not involve crowdsourcing nor research with human subjects. Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects. • Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. • According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. 15. Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? Answer: [N/A] Justification: The paper does not involve crowdsourcing nor research with human subjects. Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects. • Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. • We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. • For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. 16. Declaration of LLM usage Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigor, or originality of the research, declaration is not required. Answer: [N/A] Justification: The paper does not involve LLM as any important, original, or non-standard components. Guidelines: • The answer [N/A] means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. • Please refer to our LLM policy in the NeurIPS handbook for what should or should not be described.