Low-Rank Adaptation (LoRA) fusion enables the composition of subject and style representations for controllable generation without retraining. However, existing approaches primarily operate through weight-level merging, without explicitly modeling how independently trained LoRAs interact in the shared parameter space. We adopt a geometric perspective on LoRA fusion, interpreting content and style LoRAs as occupying overlapping, non-orthogonal low-rank subspaces, where such overlap can lead to conflicting parameter updates that affect generation quality. This observation motivates us to reformulate LoRA fusion not merely as parameter combination, but as a problem of controlling how updates from overlapping subspaces are combined. Based on this insight, we propose Null Space Projection LoRA (NP-LoRA), a training-free framework that employs projection as a fusion operator to explicitly modulate cross-LoRA interactions. Specifically, NP-LoRA uses principal directions of the style LoRA to define a projection subspace and projects the content LoRA onto the complementary subspace (i.e., the null space of the style LoRA), suppressing interference along dominant style directions while preserving complementary information. To avoid the overly aggressive suppression of hard projection, we further formulate soft projection as a regularized optimization problem that balances content preservation against style-subspace suppression. This objective admits a closed-form solution, yielding a projection operator controlled by a single parameter that continuously interpolates between linear merging and hard projection. Extensive experiments across multiple pretrained LoRA pairs show that NP-LoRA achieves more balanced content-style composition compared to strong baselines, without requiring retraining. Code is available at https://github.com/blackzebra7/NP-LoRA.
Figures & tables
Figure 1: Illustration of our motivation. (a) The task is to compose a content LoRA (subject identity) with a style LoRA (artistic appearance) for controllable generation. (b) Weighted merging can degrade stylistic fidelity, potentially due to interference, as independently trained LoRAs may occupy correlated, non-orthogonal low-rank subspaces of the diffusion feature space. (c) NP-LoRA projects the content LoRA into the null space of the style LoRA, reducing overlap and better preserving stylistic characteristics.
Figure 2: Overview of the proposed method. NP-LoRA takes pretrained content and style LoRAs as inputs. The style LoRA is decomposed via singular value decomposition (SVD) to construct a null space, onto which the content LoRA is projected. This design enables training-free fusion through a closed-form projection operator.
Figure 3: Singular value spectrum of a LoRA and perturbation effects. Perturbing dominant directions produces more noticeable stylistic changes than perturbing less principal directions.
Figure 4: Comparison of LoRA projection. Projecting style into the content null space or direct merging degrades style, while projecting content into the style null space better balances content and style.
Figure 5: Effect of the projection strength μ in NP-LoRA. (a) and (e) are the content and style references, respectively. (c) corresponds to direct merging ( μ=0 ), while (d)–(g) show increasing projection strengths from soft to hard projection. Smaller μ values tend to exhibit stronger content–style interference, whereas larger μ values better preserve style but may weaken content fidelity.
Method
ScontentCLIP
SstyleCLIP
ScontentDINO
SstyleDINO
Sarith↑
Sharm↑
μ=0
0.75
0.55
0.54
0.23
0.51
0.42
μ=0.1
0.74
0.57
0.53
0.25
0.53
0.45
μ=0.5
0.73
0.59
0.52
0.33
0.55
0.50
μ=1
0.71
0.60
0.48
0.35
0.53
0.50
μ=10
0.69
0.60
0.40
0.37
0.52
0.48
μ→∞
0.68
0.60
0.39
0.40
0.52
0.48
Table 1: Projection-strength sensitivity on the same 32 LoRA pairs and evaluation protocol as Tab. 2 . Increasing μ attenuates the content component aligned with the style subspace.
Method
ScontentCLIP
SstyleCLIP
ScontentDINO
SstyleDINO
Sarith↑
Sharm↑
Direct
0.75
0.55
0.54
0.23
0.51
0.42
ZipLoRA
0.75
0.56
0.55
0.24
0.52
0.44
DuoLoRA
0.75
0.55
0.56
0.26
0.53
0.46
K-LoRA
0.76
0.55
0.57
0.22
0.52
0.42
LoRA.rar
0.70
0.53
0.50
0.30
0.51
0.46
NP-LoRA
0.73
0.59
0.52
0.33
0.55
0.50
Table 2: Quantitative and preference-based evaluation of content–style LoRA fusion. (a) Automatic evaluation using CLIP/DINO similarities, with aggregated scores ( Sarith , Sharm ) summarizing the overall trade-off. (b) Preference evaluation by human users and GPT-5, reporting the proportion of cases where each method is selected as the best.
Figure 6: Qualitative comparison with Direct Weight Merge, ZipLoRA, DuoLoRA, K-LoRA, LoRA.rar, and our proposed NP-LoRA, illustrating the trade-off between subject fidelity and style preservation.
Method
Merge (s † ) ↓
Gen./Img. (s) ↓
Direct
13.47
22.90
ZipLoRA
>10 min
29.93
DuoLoRA
>5 min
23.93
K-LoRA
12.25 ∗
60.39 ∗
NP-LoRA
13.44
23.21
Table 3: Efficiency comparison of LoRA merging strategies. Merge includes loading time; Gen./Img. is per-image generation time. LoRA.rar Shenaj et al. (2025) is excluded for its needs pretraining.
Figure 7: Qualitative comparison on FLUX. Each row shows a subject–style pair and results from Direct Merge, K-LoRA, and NP-LoRA.
Method
ScontentCLIP
SstyleCLIP
ScontentDINO
SstyleDINO
Sarith↑
Sharm↑
Direct
0.71
0.55
0.59
0.36
0.55
0.52
K-LoRA
0.73
0.54
0.62
0.32
0.55
0.50
NP-LoRA
0.70
0.57
0.58
0.40
0.56
0.54
Table 4: Matched quantitative comparison on FLUX using the same LoRA pairs, prompts, and random seeds.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 8: Comparison of output-Space (U) and parameter-space (V) projections for null-space construction. The U-space projection fails to remove style interference, leading to distorted stylistic features, whereas the V-space projection effectively eliminates such interference and preserves the intended style appearance.
Figure 9: Qualitative comparison with joint training. Joint training exhibits unstable performance and often fails to merge content and style effectively, while N-LoRA achieves consistent and well-balanced fusion without retraining.
Figure 10: Our method effectively modifies the object’s actions and environment while maintaining the original style.
Figure 11: Results obtained with randomly selected seeds demonstrate the stability and robustness of our NP-LoRA.
Method
ScontentCLIP
SstyleCLIP
ScontentDINO
SstyleDINO
Sarith↑
Sharm↑
Direct
0.77
0.56
0.58
0.22
0.53
0.42
ZipLoRA
0.79
0.57
0.59
0.23
0.54
0.44
DuoLoRA
0.78
0.56
0.60
0.26
0.55
0.46
K-LoRA
0.80
0.55
0.60
0.22
0.54
0.43
LoRA.rar
0.73
0.53
0.53
0.31
0.52
0.47
NP-LoRA
0.78
0.60
0.58
0.30
0.56
0.50
Appendix
Table 5: Comparison under different LoRA ranks.
Figure 12: Qualitative results of NP-LoRA on the Flux backbone using diverse publicly available LoRAs. Each image corresponds to the combination of the content LoRA shown above and the style LoRA shown on the left, illustrating the results produced by our method.
Figure 13: Qualitative results of NP-LoRA on the Flux backbone using diverse publicly available LoRAs. Each image corresponds to the combination of the conten t LoRA shown above and the style LoRA shown on the left, illustrating the results produced by our method.
Method
ScontentCLIP
SstyleCLIP
ScontentDINO
SstyleDINO
Sarith↑
Sharm↑
Joint Training
0.67
0.61
0.49
0.25
0.50
0.43
B-LoRA
0.72
0.55
0.51
0.26
0.51
0.44
NP-LoRA-H
0.68
0.60
0.39
0.40
0.52
0.48
NP-LoRA
0.73
0.59
0.52
0.33
0.55
0.50
Appendix
Table 6: Comparison with joint training methods under the same experimental setting as Tab. 2 . Evaluation uses CLIP and DINO similarity scores for content preservation and style alignment. We emphasize the aggregated metrics ( Sarith and Sharm ), as they best reflect the overall trade-off between content and style.
Figure 14: Questionnaire screenshot used in the human preference study. Participants selected the result that best balanced subject fidelity and style consistency based on the generated outputs and the reference subject/style images.
Method
ScontentCLIP
SstyleCLIP
ScontentDINO
SstyleDINO
Sarith
Sharm↑
K-LoRA
0.73
0.54
0.62
0.32
0.55
0.50
NP-LoRA
0.70
0.57
0.58
0.40
0.56
0.54
Appendix
Table 7: Quantitative comparison on community-trained FLUX LoRAs. NP-LoRA achieves a better overall content–style trade-off than K-LoRA, especially in terms of style preservation and harmonic aggregation.