RecycleLoRA: Rank-Revealing QR-Based Dual-LoRA Subspace Adaptation for Domain Generalized Semantic Segmentation

RecycleLoRA: Rank-Rev ealing QR-Based Dual-LoRA Subspace Adaptation f or Domain Generalized Semantic Segmentation Chanseul Cho Seokju Y un Jeaseong Jeon Seungjae Moon Y oungmin Ro * Machine Intelligence Laboratory , Uni versity of Seoul, K orea { chanseul2001, wsz871, jasonjun1121, msj0243, youngmin.ro } @uos.ac.kr https://github.com/chanseul01/RecycleLoRA.git Abstract Domain Generalized Semantic Se gmentation (DGSS) aims to maintain r ob ust performance acr oss unseen tar get do- mains. V ision F oundation Models (VFMs) of fer rich multi-domain knowledge that can enhance generalization. However , strate gies for actively exploiting the rich sub- space structur es within VFMs r emain under-e xplor ed, with many existing methods focusing primarily on preserving pr e-trained knowledge. Furthermor e, their LoRA com- ponents often suffer from limited r epr esentational diver- sity and inefficient parameter utilization. W e pr opose Re- cycleLoRA, which addresses both challenges by employ- ing Rank-Revealing QR Decomposition (RRQR) to system- atically exploit VFM’ s subspace structur es and enhance LoRA ’ s r epr esentational richness. Our main adapter lever- ages minor subspace dir ections identified by RRQR to learn diverse and independent features, achie ving competitive performance even when used alone. W e further intr oduce a sub adapter that car efully r efines major dir ections with minimal adjustments, pr oviding complementary impr ove- ments to the main adapter’ s str ong baseline performance. This design enables the dual adapters to learn distinct r epr esentations without r equiring additional r e gularization losses. Our systematic exploitation of pr e-trained subspace structur es thr ough RRQR-based initialization leads to su- perior domain g eneralization performance. RecycleLoRA achie ves state-of-the-art performance on both synthetic- to-r eal generalization and r eal-to-r eal generalization tasks without complex ar chitectur es or additional inference la- tency . 1. Introduction Semantic segmentation assigns semantic labels to ev ery pixel in an image and plays a crucial role in autonomous driving, medical imaging, and robotics. Howe v er , domain shift, the phenomenon where models trained on one do- main experience performance degradation when applied to another , limits real-world deployment. T o address this GT RecycleLoRA (Ours) RecycleLoRA (Main adapter Only) SoMA 71.82 72.92 73.01 61.31 61.22 61.77 71.67 71.75 72.07 68.27 68.63 68.95 GT A V à Cityscapes GT A V à BDD100K GT A V à Mapillary Average SoMA (Previous SOT A) RecycleLoRA (Main adapter only) RecycleLoRA (Ours) Figure 1. Comparison of synthetic-to-real generalization perfor- mance (mIoU,%) between our proposed Rec ycleLoRA and the previous SO T A method, SoMA. challenge, Domain Generalized Semantic Segmentation (DGSS) has been introduced, aiming to de velop models that maintain robust performance across di verse domains with- out target domain data. This capability is particularly es- sential in safety-critical applications where models must re- liably handle varying conditions. T raditional DGSS approaches focused on data augmen- tation and domain-inv ariant feature learning but used back- bones trained on limited datasets [ 2 , 8 , 14 , 58 , 59 , 69 ], whose knowledge was confined to specific domains and thus had limited generalization capability [ 16 , 27 , 64 ]. With the emergence of V ision Foundation Models (VFMs) such as DINOv2 [ 55 ] and CLIP [ 60 ]—trained on large-scale, div erse datasets that already capture rich and transferable knowledge across domains—the emphasis in DGSS is shift- 1 ing from div ersifying inputs to preserving and efficiently adapting VFM’ s world knowledge. T o that end, recent stud- ies hav e explored Parameter-Ef ficient Fine-Tuning (PEFT) methods for adapting VFMs, achieving competitiv e perfor- mance with minimal computational overhead. In DGSS, Rein [ 71 ] introduced learnable tokens, while SoMA [ 76 ] lev eraged Low-Rank Adaptation (LoRA) [ 29 ] to selectiv ely adjust minor components through Singular V alue Decom- position (SVD). Meanwhile, Tqdm [ 56 ] and MFuser [ 77 ] lev eraged V ision-Language Models (VLMs) to enhance cross-domain generalization. Howe v er , existing methods face limitations in fully ex- ploiting the potential of VFMs. First, while SVD-based approaches such as SoMA [ 76 ] have sho wn promising re- sults by focusing on minor singular components for pre- serving pre-trained knowledge, it remains underexplored whether SVD is the most effecti ve decomposition method for adapting V ision Foundation Models. In particular, SVD prioritizes v ariance preservation, which, while mathemat- ically optimal for data reconstruction, does not necessar- ily guarantee the most rele v ant directions for downstream adaptation [ 1 , 41 ]. Moreov er , SoMA adjusts only the mi- nor directions, leaving potentially useful major components untouched during adaptation. This restricted view may limit the model’ s capacity to handle complex new tasks or fully exploit the rich representations in VFMs. Moreov er , many LoRA-based methods suffer from limited represen- tational div ersity due to learning redundant representations among their basis vectors, which leads to inef ficient param- eter utilization(e.g., T ab. 2 , Fig. 3 ) In domain generaliza- tion research, enhancing representational di versity has been shown to improve generalization performance by enabling models to capture a wider range of features [ 30 , 52 , 70 , 73 ]. This issue of representational collapse or rank deficiency in LoRA has been noted and explored in several recent stud- ies [ 25 , 31 , 38 , 40 , 46 ]. T o address these problems, we introduce an initializa- tion strategy based on Rank-Revealing QR Decomposition (RRQR). In contrast to SVD, which finds new orthogonal bases that preserv e global v ariance, RRQR selects informa- tiv e columns directly from the original weight matrix us- ing greedy column pi voting [ 7 ]. At each step, it identifies the column with the largest orthogonal component relative to the pre viously selected subspace, thereby minimizing re- dundancy and preserving directional independence. Since this approach constructs the basis vectors based on columns selected directly from the original weight matrix, the unique structural information held by those columns is well reflected in the ne w basis. By selecting basis vectors from the actual columns of the weight matrix, RRQR re- tains localized structural information and preserves the cor - respondence between weight dimensions and learned rep- resentations. This leads to LoRA adapters that are both in- terpretable and div erse in representation. Importantly , our RRQR-based initialization helps mitigate the representa- tional redundanc y often observ ed among LoRA ’ s basis vec- tors. RRQR’ s greedy selection process promotes directional independence, naturally constructing a LoRA adapter with enhanced representational capacity . By recycling struc- turally informative directions, our method enhances both parameter ef ficiency and adaptation capacity while preserv- ing the core knowledge embedded in VFMs. Furthermore, while methods that focus only on minor directions are ef- fectiv e for preserving the VFM’ s pre-trained kno wledge, they can struggle to adapt to new , comple x tasks. Recent work such as PiSSA [ 51 ] has demonstrated that effecti v e task adaptation can be achie ved by tuning only the major directions of the pretrained weights. Motivated by this, we extend our design with a complementary sub-adapter that carefully refines major directions, further improving gener- alization performance. Building upon these insights, we propose RecycleLoRA , a nov el approach to utilizing pre-trained weights through RRQR decomposition. As demonstrated in Fig. 1 , our main adapter , initialized with minor directions identified by RRQR, achieves state-of-the-art performance by learn- ing div erse and independent features. This demonstrates that strategically rec ycling these minor directions alone sur- passes existing methods. T o further enhance performance, we introduce a sub adapter that carefully refines major di- rections with minimal adjustments, providing complemen- tary improvements. This strate gic design enables the two adapters to naturally learn complementary features with- out additional regularization losses or complex training regimes. As shown in our analysis, the two adapters learn to operate in distinct subspaces and induce dif ferent types of modifications in the feature space. Experimental results demonstrate that RecycleLoRA achiev es top performance in both synthetic-to-real and real- to-real generalization tasks without VLMs or complex ar- chitectures. This sho ws that superior performance in do- main generalization can be achiev ed by effecti vely exploit- ing pre-trained subspace structures. Our main contributions are as follo ws. • W e propose a novel initialization strategy for LoRA based on Rank-Re vealing QR Decomposition, which mitigates representational redundancy by selecting structurally di- verse directions from the original weight matrix. This im- prov es both parameter utilization and task-specific adapt- ability in VFM fine-tuning. • W e design a dual-adapter structure that combines a main adapter leveraging minor directions with a sub adapter re- fining major directions, enabling the model to naturally learn complementary feature representations without ex- plicit regularization. • Our method achieves state-of-the-art performance, with 2 68.95 mIoU in synthetic-to-real generalization and 72.10 mIoU in real-to-real generalization. 2. Related W ork 2.1. Domain Generalized Semantic Segmentation Domain Generalized Semantic Segmentation (DGSS) aims to train models that can generalize to unseen target do- mains without access to target domain data during train- ing. Early approaches primarily focused on alleviating do- main shift through data augmentation and adv ersarial train- ing techniques, b ut their performance w as constrained by the limited representational power of the con ventional back- bones they relied on, which were often trained on limited datasets [ 8 , 20 , 36 , 37 , 39 , 43 , 54 , 57 – 59 , 69 , 72 , 78 ]. The emergence of V ision Foundation Models (VFMs) has introduced ne w paradigms for DGSS. SoMA [ 76 ] in- troduces a method that lev erages the subspace structure of pre-trained weights by selectively tuning minor singu- lar components through singular value decomposition, ef- fectiv ely preserving the generalization capacity of VFMs while acquiring task-specific kno wledge. Rein [ 71 ] pro- poses a parameter-ef ficient approach utilizing learnable tokens that refine feature maps layer-by-layer , enabling instance-lev el refinement within the backbone architecture. Meanwhile, methods such as MFuser [ 77 ] and tqdm [ 56 ] hav e improv ed generalization performance by leveraging the domain-in v ariant properties of text information based on VLMs. While recent VFM-based DGSS methods hav e primarily focused on preserving pre-trained kno wledge, ap- proaches that systematically recycle and exploit their rich internal subspace structures remain underexplored. 2.2. V ision F oundation Models V ision Foundation Models have emerged as powerful tools for various computer vision tasks, offering strong general- ization capabilities across div erse domains. Among promi- nent VFMs, DINOv2 [ 55 ] utilizes self-supervised learning techniques to learn robust visual representations from di- verse visual data, enabling broad applicability across var - ious downstream tasks. EV A02-CLIP [ 21 ] is a V ision- Language Model that provides robust, domain-in variant representations by aligning visual features with textual se- mantics. CLIP [ 60 ] has established itself as a foundational V ision-Language Model through joint training on image- text pairs, enabling zero-shot classification and cross-modal understanding. 2.3. Parameter -Efficient Fine-T uning Parameter -Efficient Fine-T uning (PEFT) has become a stan- dard for adapting large models, as it enables fine-tuning with only a small fraction of the total parameters. Among these techniques, Lo w-Rank Adaptation (LoRA) [ 29 ] is a prominent method that freezes the original weights and in- jects trainable, low-rank matrices to model weight updates, achieving comparable performance to full fine-tuning with high parameter efficienc y . Recent dev elopments in PEFT ha ve explored more so- phisticated initialization strategies that le verage subspace structures. SoMA [ 76 ] utilizes singular value decomposi- tion to identify and tune minor singular components, specif- ically tar geting the less dominant singular values while pre- serving the major ones to maintain the pre-trained kno wl- edge. This approach focuses on knowledge preserv ation by selectiv ely modifying the subspace components that con- tribute less to the original representation, primarily concen- trating on the minor components. Methods like PiSSA [ 51 ] also use singular v alue decomposition for initialization to lev erage the principal directions of weight matrices. These approaches have shown that consideration of the underlying structure in pre-trained weights can significantly impact the effecti v eness of low-rank adaptation. Existing PEFT methods for domain generalization ex- hibit two main limitations. First, they tend to focus on pre- serving pre-trained kno wledge rather than activ ely exploit- ing it. Second, many LoRA-based approaches suffer from inefficient parameter utilization and limited representational capacity , which leads to under-utilized subspace informa- tion and parameter redundancy [ 18 , 48 ]. Our approach ad- dresses both of these challenges by systematically recycling subspace components to simultaneously improve LoRA ’ s parameter efficienc y and representational capabilities. 3. Proposed Methods In this section, we introduce our proposed method, Re- cycleLoRA. Section 3.1 provides the necessary technical background on Low-Rank Adaptation (LoRA) and Rank- Rev ealing QR Decomposition (RRQR), which are founda- tional to our method. Subsequently , in Section 3.2 , we present the detailed design of RecycleLoRA, validating its effecti v eness through an in-depth in vestigation. 3.1. Preliminaries Low-Rank Adaptation (LoRA). LoRA is a parameter- efficient fine-tuning technique that models weight updates through trainable lo w-rank decomposition while freezing pre-trained weights W 0 ∈ R d × k . The weight update ∆ W is represented as: W = W 0 + ∆ W = W 0 + BA (1) where B ∈ R d × r , A ∈ R r × k , and the rank r is much smaller than the original dimensions ( r ≪ min( d, k ) ), con- straining the adaptation to a low-dimensional subspace. The learned low-rank matrices can be merged into the original weights during inference, introducing no additional infer - ence latency . 3 A (c) Main Adapter c (a) (b) Pretrained Weight A (d) Sub Adapter RRQR based Subspace Selection Frozen Parameter Tra i na b le wi t h S ta nd a rd LR Tra i na b le wi t h Lo w er LR Initialization with “minor” directions Initialization with “major” directions Dual - Adapter Residual Weight RecycleLoRA B B Figure 2. RecycleLoRA Framework Overview . This figure illustrates the overall workflo w of RecycleLoRA. (a) Rank-Revealing QR Decomposition (RRQR) is applied to the pre-trained weight matrix to identify subspace directions ranked by importance. (b) Among the recyclable subspaces selected through RRQR, the minor directions are assigned as initialization v alues for the main adapter , while the major directions are assigned to the sub adapter . (c) The main adapter’ s B matrix is initialized with the minor directions, and its A matrix is sparsely initialized by mapping these directions to their corresponding column indices. (d) The sub adapter’ s B matrix is initialized with the major directions, and its A matrix is sparsely initialized by mapping these directions to their corresponding column indices. Rank-Revealing QR Decomposition (RRQR). The as- sessment of parameter importance has been a significant research topic across various areas of deep learning. A substantial body of work has established that weight mag- nitudes serve as effecti ve indicators of parameter impor- tance. Classical pruning methods such as magnitude-based pruning demonstrate that parameters with larger magni- tudes typically contribute more significantly to model per- formance [ 19 , 26 , 42 , 44 ]. This principle has been ex- tended to various contexts, including structured pruning, where entire channels or layers are ranked by their norm- based importance scores [ 28 , 32 , 45 , 67 ]. Recent advances in parameter-ef ficient fine-tuning ha ve similarly le veraged magnitude-based importance measures [ 12 , 25 , 48 , 50 ]. These findings establish weight magnitude as an effecti ve indicator of parameter importance, moti v ating our adoption of RRQR decomposition, which systematically ranks ma- trix columns by their norm-based importance. F or a matrix W ∈ R m × n , the Rank-Revealing QR (RRQR) decomposi- tion is expressed as: WP = QR (2) where P ∈ R n × n is a permutation matrix, Q ∈ R m × n is an orthogonal matrix ( Q T Q = I ), and R ∈ R n × n is an upper triangular matrix whose diagonal elements capture the magnitude of each column’ s orthogonal component, and whose off-diagonal elements encode the dependencies be- tween columns. At each step k , the algorithm selects the next column from the set of remaining columns. The cho- sen column is the one that has the largest norm after being projected onto the orthogonal complement of the subspace spanned by the previously selected columns. Specifically , giv en the already selected columns W P 1 , . . . , W P k − 1 , the algorithm chooses the next column W P k that maximizes:    WP k − proj span ( WP 1 ,..., WP k − 1 ) ( WP k )    2 (3) where proj span ( · ) denotes orthogonal projection onto the subspace spanned by the argument. This greedy selection process typically produces a strong tendency for the diago- nal elements of R to satisfy: | r 11 | ≥ | r 22 | ≥ · · · ≥ | r nn | (4) indicating that most of the matrix energy is concentrated in the leading components. The permutation matrix P records this importance ordering, where P [ i ] indicates the original column index of the i -th most important direction. The orthogonal matrix Q provides the corresponding orthonor- mal basis, where each column q i represents the normalized direction of the orthogonal component of the P [ i ] -th col- umn. Thus, RRQR provides two ke y insights: the permuta- tion matrix P identifies the importance ordering of original columns, while the orthogonal matrix Q provides the cor- responding geometric directions. This structural characteri- zation enables effecti v e LoRA initialization through sparse mapping of specific input dimensions, enhancing represen- tation div ersity and parameter utilization. 3.2. RecycleLoRA RecycleLoRA is a dual-adapter methodology that uses RRQR decomposition to separate pre-trained weights into minor and major directions, which initialize a main and sub adapter , respectiv ely (Figure 2 ). T o preserve the initial out- put of the pre-trained weights, we construct a residual ma- trix by subtracting the initial adapter values from the orig- inal weights before training. This matrix is then frozen, so that only the two adapters are trained. T o demonstrate 4 T able 1. Comparison of ℓ 2 -norm statistics between selected and non-selected columns in LoRA matrix A after training. LoRA Column-wise Norm Statistics Mean ℓ 2 norm Mean ℓ 2 norm A verage Maximum (selected) (non-selected) ratio ratio All layers 0.138245 0.113566 1.22 × 1.63 × T able 2. Effecti ve rank and rank ef ficiency comparison. Rank efficienc y is calculated as the ratio of effec tiv e rank to target rank, measuring parameter utilization. RecycleLoRA consistently achiev es higher efficiency than SoMA across dif ferent rank set- tings. Effective Rank & Rank-Efficiency Statistics T arget rank RecycleLoRA (Ours) SoMA Effecti ve rank Rank efficiency Effective rank Rank efficiency 16 13.60 0.850 9.78 0.611 32 24.65 0.770 20.80 0.650 the effecti veness of our approach, we compare our method against SoMA [ 76 ], which is the previous state-of-the-art and a LoRA-based method. Main Adapter . RecycleLoRA lev erages RRQR decompo- sition to design a novel initialization strategy that systemat- ically recycles pre-trained kno wledge from VFMs while si- multaneously enhancing LoRA ’ s representational efficiency for improved domain generalization. Specifically , we per- form RRQR decomposition on each linear layer’ s weight matrix W 0 ∈ R d × k to obtain: W 0 P = QR (5) where Q ∈ R d × k is an orthogonal matrix and R ∈ R k × k is upper triangular . In practice, the permutation matrix P is returned as an index array P ∈ N k , where P [ i ] indicates the original column index of the i -th most important direction. The main adapter is initialized as: B main = Q [: , − r main :] (6) A main [ i, j ] = ( 1 , if j = P [ k − r main + i ] 0 , otherwise (7) where i ∈ { 0 , 1 , ..., r main − 1 } and j ∈ { 0 , 1 , ..., k − 1 } . This sparse initialization is designed to focus initial changes on specific input dimensions identified by RRQR. T o v alidate the effecti veness of this initialization strat- egy , we analyzed the ev olution of LoRA parameters throughout training. T able 1 presents the column-wise norms of the A matrix after training, re vealing that columns containing sparsely initialized positions with value 1 main- tain norms that are on av erage 1 . 22 × larger , and up to 1 . 63 × larger , than columns initialized entirely to 0. This Figure 3. Cosine similarity heatmaps of LoRA components for (a) RecycleLoRA and (b) SoMA at different ranks (r=16, 32). Left: pairwise similarity among rows of A . Right: pairwise similarity among columns of B . Darker blue colors represent lower similar- ity . suggests that parameter updates during training were rel- ativ ely concentrated on the initially selected directions, demonstrating that our sparse initialization maintains struc- tural bias to some extent throughout the training process. Furthermore, we ev aluated the div ersity of learned rep- resentations and the ef ficiency of parameter utilization. Fig- ure 3 visualizes the cosine similarity between LoRA com- ponents for both SoMA and Rec ycleLoRA, sho wing that RecycleLoRA consistently exhibits lower similarity across various rank settings. Specifically , we focused our analy- sis on the dimensions that actually determine the rank in LoRA ’ s low-rank structure: the similarity between rows of A ∈ R r × k and the similarity between columns of B ∈ R d × r . This is because in LoRA ’ s weight update ∆ W = BA , the i -th row of A and the i -th column of B together form a single low-rank component. Therefore, low similarity among rows of A and low similarity among columns of B indicate that each low-rank component cap- tures distinct, independent features. This reduced similarity encourages each low-rank component to learn more inde- pendent and distinctiv e features, thereby improving the uti- lization efficienc y of limited parameters. This di verse representation learning directly translates to enhanced model expressi veness. As shown in T a- ble 2 , RecycleLoRA consistently achiev es higher Effecti ve Rank [ 63 ] compared to SoMA. The Effecti ve Rank quanti- fies the dimensional richness of learned representations, and recent studies ha ve shown that higher Ef fectiv e Rank corre- lates with improv ed representational capacity and general- ization performance [ 22 , 31 , 38 , 46 ]. The consistent im- prov ements observed across various settings demonstrate that our method utilizes LoRA ’ s limited parameters more efficiently to enhance representational capacity . Sub Adapter . A key insight in the design of RecycleLoRA is that different subspace components within VFM weights can contribute complementarily to domain generalization performance. Building on this observation, we introduce 5 Subspace Similarity Subspace Similarity between Main & Sub Adapters Figure 4. Block-wise subspace similarity ϕ between main and sub adapters measured using Grassmann Distance on the lo w-rank ma- trices. Lower v alues indicate more orthogonal subspaces. a sub adapter that complements the main adapter . By inv estigating recent LoRA initialization strategies, we uncovered a critical yet underexplored relationship be- tween the choice of initialization subspace and the optimal learning rate. Specifically , we observed that methods that initialize LoRA with major directions, such as PiSSA [ 51 ], tend to adopt lower learning rates. W e can infer that this is because major directions encode the V ision Foundation Model’ s core, generalizable knowledge, making them sen- sitiv e to large updates that could risk catastrophic for get- ting. A lower learning rate thus enables a careful refine- ment of these critical components, preserving foundational knowledge while adapting to the new task. Con versely , we observ ed that approaches using minor directions, like SoMA [ 76 ], tend to employ relati vely higher learning rates. Since these minor directions contrib ute less to the model’ s pretrained capabilities, they can pro vide a safer subspace for learning new representations. A higher learning rate al- lows for more aggressive and efficient adaptation within this subspace without jeopardizing the model’ s core representa- tions. This suggests an inherent relationship between the initialization strategy and the optimal learning rate, rooted in the trade-of f between knowledge preservation and task adaptation. Motiv ated by this observation, we in vestigated the in- terplay between initialization methods and learning rates. Our experiments, detailed in Section 4.3 (T ab 6 ), sug- gest a tendency where the optimal learning rate is contin- gent upon the nature of the initialized directions. Specif- ically , the sub adapter , initialized with RRQR’ s top direc- tions, tended to e xhibit impro ved performance at a lower learning rate (5e-5), which implies that the major directions encoding the VFM’ s core knowledge require more careful optimization. In contrast, both the main adapter initial- ized with minor directions and standard LoRA with Kaim- ing initialization achiev ed peak performance at the standard learning rate (1e-4), with their performance declining when the learning rate was reduced. Notably , this performance drop w as more pronounced for the main adapter than for (a) (b) (c ) Figure 5. V isualization of adapter-induced feature modifications via PCA projection. (a) Input. (b) main adapter PCA V isualiza- tion. (c) sub adapter PCA V isualization. The div ergent activ ation patterns rev eal complementary feature learning between the dual adapters. the Kaiming-initialized LoRA. This result suggests that the minor directions provide a safer subspace for learning new , task-specific features, thereby benefiting from more aggres- siv e updates. These contrasting findings validate our design choice of employing a dif ferentiated learning rate scheme in our dual-adapter architecture, where each adapter is opti- mized according to the sensitivity of its assigned subspace. Based on these findings, we incorporate a sub adapter that leverages the top directions from RRQR. While our main adapter alone already surpasses existing state-of-the- art methods (as will be demonstrated in T able 5 , Section 4.3), the sub adapter provides complementary performance improv ements. The sub adapter is initialized as: B sub = Q [: , : r sub ] (8) A sub [ i, j ] = ( 1 , if j = P [ i ] 0 , otherwise (9) The sub adapter is trained with a much smaller rank (32 → 4) and lo wer learning rate (1e-4 → 5e-5) than the main adapter , allowing careful adjustment of the major directions in the pre-trained weights. T o verify that the two adapters learn distinct representa- tions, we conducted several analyses. Figure 4 presents the subspace similarity measured using Grassmann Distance. Specifically , we compute the similarity between the sub- spaces spanned by the left singular vectors ( U main , U sub ) 6 from the SVD of each adapter’ s matrix A : ϕ = ∥ U T main U sub ∥ 2 F min( r main , r sub ) ∈ [0 , 1] (10) This metric, introduced in the LoRA paper [ 29 ], ranges from 0 to 1, where 1 indicates identical subspaces and 0 signifies complete orthogonality . As sho wn in Figure 4 , our RecycleLoRA framework yields a consistently low simi- larity between the main and sub adapters. As a baseline, we trained a dual-adapter model that, while initialized with the standard Kaiming method, otherwise shared the iden- tical configuration of RecycleLoRA, including the dif fer - entiated rank and learning rate settings for the main and sub adapters. This baseline exhibited significantly higher subspace similarity than RecycleLoRA. This comparative analysis demonstrates that our RRQR-based initialization is instrumental in guiding the adapters to operate in distinct, nearly orthogonal subspaces. Furthermore, Figure 5 presents PCA projections of the feature differences produced by the main and sub adapters in the final block of DINOv2. The main adapter induces localized, salient modifications around foreground objects, whereas the sub adapter yields broader shifts that spread across the background. These complementary patterns sup- port the claim that the RRQR-based initialization and the differentiated rank and learning-rate design steer the two adapters tow ard learning complementary representations. This integration of RRQR-based initialization, differen- tiated rank allocation, and carefully tuned learning rates en- ables Rec ycleLoRA to systematically exploit VFM’ s multi- domain knowledge, achieving superior domain generaliza- tion performance without additional regularization. 4. Experiments 4.1. Experimental Settings Datasets. W e ev aluate the effecti veness of RecycleLoRA using widely adopted benchmark datasets for Domain Gen- eralized Semantic Segmentation. For synthetic data, we use GT A V [ 61 ], which consists of 12,403 training images, 6,382 validation images, and 6,181 test images. For real- world data, we employ Cityscapes [ 15 ] with 2,975 training images and 500 v alidation images, Berkele y Deep Driv- ing dataset [ 75 ] with 1,000 validation images, and Mapil- lary [ 53 ] with 2,000 validation images. Implementation Details. W e use DINOv2-Large as the backbone and Mask2Former as the segmentation head. Re- cycleLoRA is applied to all linear layers within the self- attention modules and MLP layers of the transformer . The main adapter uses a rank of 32, while the sub adapter’ s rank is set to 4 for the synthetic-to-real setting and 2 for the real- to-real setting. The learning rate multipliers for the main and sub adapters are 1.0 and 0.5, respecti vely . All experi- ments use 512×512 cropped images for training. T able 3. Domain generalization results (mIoU %) under the synthetic-to-real setting. Our method is highlighted in gray . Bold and underlined indicate best and second-best results. Synthetic-to-Real Generalization Method V enue Backbone T rained on GT A V → Citys. → BDD → Map. A vg. CLOUDS [ 4 ] CVPR2024 CLIP-CN-L 60.20 57.40 67.00 61.50 VL TSeg [ 33 ] ACCV2024 EV A02-L 65.30 58.30 66.00 63.20 DoRA [ 49 ] ICML2024 DINOv2-L 66.12 59.31 67.07 64.17 VPT [ 35 ] ECCV2022 DINOv2-L 68.75 58.64 68.32 65.24 SET [ 74 ] TIP2021 DINOv2-L 68.06 61.64 67.68 65.79 tqdm [ 56 ] ECCV2024 EV A02-L 68.88 59.18 70.10 66.05 Rein † [ 71 ] CVPR2024 DINOv2-L 69.19 60.01 69.06 66.09 F AD A [ 5 ] NeurIPS2024 DINOv2-L 68.23 61.94 68.09 66.09 AdaptFormer [ 10 ] NeurIPS2022 DINOv2-L 70.10 59.81 68.77 66.23 PEGO [ 30 ] ECCV2024 DINOv2-L 68.86 61.44 68.61 66.30 SSF [ 47 ] NeurIPS2022 DINOv2-L 68.97 61.30 68.77 66.35 LoRA [ 29 ] ICLR2022 DINOv2-L 70.13 60.13 70.42 66.89 DepthForge [ 11 ] ICCV2025 DINOv2-L 69.04 62.82 69.22 67.03 DPMFormer [ 34 ] ICCV2025 EV A02-L 70.08 60.48 70.66 67.07 Mfuser [ 77 ] CVPR2025 EV A02-L 70.19 63.13 71.28 68.20 SoMA [ 76 ] CVPR2025 DINOv2-L 71.82 61.31 71.67 68.27 RecycleLoRA - DINOv2-L 73.01 61.77 72.07 68.95 Method V enue Backbone T rained on GT A V + Synthia → Citys. → BDD → Map. A vg. Rein † [ 71 ] CVPR2024 DINOv2-L 72.17 61.53 70.69 68.13 SoMA [ 76 ] CVPR2025 DINOv2-L 73.16 61.90 72.73 69.26 RecycleLoRA - DINOv2-L 73.71 61.87 72.68 69.42 Method V enue Backbone T rained on GT A V + Synthia + UrbanSyn → Citys. → BDD → Map. A vg. Full Fine-T uning - DINOv2-L 75.90 60.93 72.80 69.88 SoMA [ 76 ] CVPR2025 DINOv2-L 77.33 62.78 74.93 71.68 RecycleLoRA - DINOv2-L 78.66 63.46 74.83 72.32 4.2. Comparison with State-of-the-Art Methods T o demonstrate the effecti veness of our method, we com- pare RecycleLoRA against e xisting state-of-the-art DGSS methods. F or a fair comparison, some methods are reim- plemented using publicly av ailable official checkpoints, and these reimplemented results are denoted with † . Synthetic-to-Real Generalization. W e compare Recy- cleLoRA against a wide range of recent state-of-the-art methods. As shown in T able 3 , in the setting where models are trained on synthetic GT A V data and ev alu- ated on real-world Cityscapes, BDD, and Mapillary data, RecycleLoRA achieves state-of-the-art performance, out- performing all existing methods. Notably , our method achiev es 73.01 mIoU on GT A V → Cityscapes, representing a substantial improv ement of 1.19 mIoU over the previ- ous best-performing method. It also achie ves a 0.4 mIoU improv ement on GT A V → Mapillary , establishing state-of- the-art performance with an av erage improv ement of 0.68 mIoU. T o further assess the robustness and scalability of our method, we also ev aluate RecycleLoRA in a multi- source generalization setting where the model is trained on combined synthetic datasets. As detailed in T able 3 , when trained on GT A V and Synthia, RecycleLoRA achie ves an av erage mIoU of 69.42, surpassing the pre vious state-of- the-art method, SoMA [ 76 ]. W e further extend the exper- iment by adding UrbanSyn to the training sources. In this three-source setting, RecycleLoRA again demonstrates su- perior performance, achieving an average mIoU of 72.32 7 T able 4. Domain generalization results (mIoU %) under the real- to-real setting. Our method is highlighted in gray . Bold and underlined indicate best and second-best results. Real-to-Real Generalization Method V enue Backbone T rained on Cityscapes → BDD → Map. A vg. HGFormer [ 17 ] CVPR2023 Swin-L 61.50 72.10 66.80 CMFormer [ 6 ] AAAI2024 Swin-L 62.60 73.60 68.10 PD AF [ 9 ] ICCV2025 Swin-L 63.00 74.10 68.55 SET [ 74 ] TIP2021 DINOv2-L 65.07 75.67 70.37 VL TSeg [ 33 ] A CCV2024 EV A02-L 64.40 76.40 70.40 tqdm [ 56 ] ECCV2024 EV A02-L 64.72 76.15 70.44 DPMFormer [ 34 ] ICCV2025 EV A02-L 64.20 76.67 70.44 F AD A [ 5 ] NeurIPS2024 DINOv2-L 65.12 75.86 70.49 Rein † [ 71 ] CVPR2024 DINOv2-L 66.53 75.18 70.86 DepthForge [ 11 ] ICCV2025 DINOv2-L 66.19 75.93 71.06 SoMA [ 76 ] CVPR2025 DINOv2-L 67.02 76.45 71.74 MFuser [ 77 ] CVPR2025 EV A02-L 65.81 77.93 71.87 RecycleLoRA - DINOv2-L 66.65 77.54 72.10 T able 5. Ablation study on the Main and Sub Adapter components of RecycleLoRA. Best results in bold . Main Sub Params. → Citys. → BDD → Map. A vg. " 1.6M 70.64 60.56 71.11 67.44 " 12.6M 72.92 61.22 71.75 68.63 " " 14.2M 73.01 ( ↑ 0.09) 61.77 ( ↑ 0.55) 72.07 ( ↑ 0.32) 68.95 ( ↑ 0.32) and widening its performance gap over SoMA. These re- sults confirm that RecycleLoRA effecti vely lev erages mul- tiple source domains and maintains its strong generalization capabilities as the div ersity of training data increases. Real-to-Real Generalization. In this setting, our method is compared with various competitiv e approaches. T able 4 shows that in the real-to-real scenario, where models are trained on Cityscapes and ev aluated on BDD and Mapil- lary , RecycleLoRA consistently demonstrates superior per- formance, achie ving state-of-the-art performance with an av erage improvement of 0.23 mIoU. These results confirm that our method ef fectiv ely handles not only synthetic-to- real domain gaps but also v ariations between different real- world domains. 4.3. Ablation Studies T o validate our design choices, we conduct a series of abla- tion studies in the synthetic-to-real generalization setting. Components Analysis. T able 5 presents an analysis of each component’ s contribution. Remarkably , using only the main adapter achiev es an average mIoU of 68.63, which already surpasses all previous state-of-the-art methods, in- cluding SoMA. This result highlights that our RRQR-based initialization strategy for the main adapter is highly effec- tiv e on its o wn. The addition of the sub adapter further boosts performance across all target domains, reaching an av erage of 68.95 mIoU. This confirms that the two adapters work in a complementary manner to maximize domain gen- eralization performance. Learning Rate Analysis. T o further in vestigate our hy- pothesis on the relationship between the initialization strat- T able 6. Comparativ e analysis of learning rate sensitivity across initialization strategies. Standard LoRA emplo ys Kaiming ini- tialization while RecycleLoRA ’ s Sub Adapter utilizes RRQR top- ranked directions. The best result for each method is shown in bold . Method lr . → Citys. → BDD → Map. A vg. Main Adapter 1e-4 72.92 61.22 71.75 68.63 Main Adapter 5e-5 69.46 ( ↓ 3.46) 62.22 ( ↑ 1.00) 68.23 ( ↓ 3.52) 66.64 ( ↓ 1.99) LoRA 1e-4 70.36 60.21 69.95 66.84 LoRA 5e-5 69.40 ( ↓ 0.96) 59.62 ( ↓ 0.59) 69.20 ( ↓ 0.75) 66.07 ( ↓ 0.77) Sub Adapter 1e-4 68.60 60.74 67.54 65.63 Sub Adapter 5e-5 70.64 ( ↑ 2.04) 60.56 ( ↓ 0.18) 71.11 ( ↑ 3.57) 67.44 ( ↑ 1.81) T able 7. Domain generalization performance (mIoU, %) on the EV A02-L backbone under the synthetic-to-real setting. Bold and underlined indicate the best and second-best results, respectiv ely . Synthetic-to-Real Generalization Method V enue Backbone T rained on GT A V → Citys. → BDD → Map. A vg. Rein [ 71 ] CVPR2024 EV A02-L 65.30 60.50 64.90 63.60 F AD A [ 5 ] NeurIPS2024 EV A02-L 66.70 61.90 66.10 64.90 DepthForge [ 11 ] ICCV2025 EV A02-L 68.00 61.70 67.50 65.73 SoMA [ 76 ] CVPR2025 EV A02-L 68.05 60.81 68.33 65.73 RecycleLoRA - EV A02-L 68.95 61.37 68.73 66.35 egy and learning rates, we present the analysis in T able 6 . The results indicate that both the main adapter, initialized with RRQR’ s minor directions, and the standard LoRA with Kaiming initialization tend to show performance degrada- tion as the learning rate is reduced, with this trend being more pronounced for the main adapter . Interestingly , the sub adapter , which is initialized with RRQR’ s top direc- tions, exhibits a contrasting and notable pattern of impro ved performance at a lo wer learning rate. These observations suggest a potential link between the directions used for ini- tialization and the optimal learning rate, a finding that sup- ports the rationale for applying dif ferent learning rates to our main and sub adapters. Perf ormance on Different VFM backbones. T o demon- strate its generality , we also ev aluated RecycleLoRA on the EV A02-L backbone. As shown in T able 7 , Recy- cleLoRA achiev es state-of-the-art performance ag ainst pure V ision Foundation Model adaptation methods, with VLM- based approaches that lev erage textual information being excluded from the comparison [ 34 , 56 , 77 ]. This result con- firms that our RRQR-based strategy is robust and ef fectiv e across different VFM architectures. 5. Conclusion In this paper, we introduce RecycleLoRA, an approach de- signed to acti vely recycle the internal subspace structures of V ision Foundation Models through Rank-Revealing QR decomposition. This strategy enables complementary fea- ture learning by lev eraging both minor and major directions, achieving state-of-the-art performance on both synthetic-to- real and real-to-real generalization tasks. 8 6. Acknowledgments This work w as supported by the National Research F ounda- tion (NRF) grant funded by the Korea government (MSIT) [RS-2025-00562400] and [RS-2022-NR068754]. References [1] Mohamed Abdelnaby and Marmar R Moussa. A benchmark- ing study of random projections and principal components for dimensionality reduction strategies in single cell analy- sis. bioRxiv , 2025. 2 [2] W oo-Jin Ahn, Geun-Y eong Y ang, Hyun-Duck Choi, and Myo-T aeg Lim. Style blind domain generalized semantic segmentation via co variance alignment and semantic consis- tence contrastive learning. In Proceedings of the ieee/cvf confer ence on computer vision and pattern reco gnition , pages 3616–3626, 2024. 1 [3] Muhammad A wais, Muzammal Naseer , Salman Khan, Rao Muhammad Anwer, Hisham Cholakkal, Mubarak Shah, Ming-Hsuan Y ang, and Fahad Shahbaz Khan. Foundation models defining a new era in vision: a survey and outlook. IEEE T ransactions on P attern Analysis and Machine Intelli- gence , 2025. [4] Y asser Benigmim, Subhankar Roy , Slim Essid, V icky Kalo- geiton, and St ´ ephane Lathuili ` ere. Collaborating foundation models for domain generalized semantic segmentation. In Pr oceedings of the IEEE/CVF Confer ence on Computer V i- sion and P attern Recognition , pages 3108–3119, 2024. 7 [5] Qi Bi, Jingjun Y i, Hao Zheng, Haolan Zhan, Y awen Huang, W ei Ji, Y uexiang Li, and Y efeng Zheng. Learning frequency- adapted vision foundation model for domain generalized se- mantic segmentation. Advances in Neural Information Pr o- cessing Systems , 37:94047–94072, 2024. 7 , 8 [6] Qi Bi, Shaodi Y ou, and Theo Gev ers. Learning content- enhanced mask transformer for domain generalized urban- scene segmentation. In Pr oceedings of the AAAI Confer ence on Artificial Intelligence , pages 819–827, 2024. 8 [7] T ony F Chan. Rank rev ealing qr factorizations. Linear alge- bra and its applications , 88:67–82, 1987. 2 [8] Prithvijit Chattopadhyay , Kartik Sarangmath, V ivek V i- jaykumar , and Judy Hoffman. Pasta: Proportional amplitude spectrum training augmentation for syn-to-real domain gen- eralization. In Proceedings of the IEEE/CVF international confer ence on computer vision , pages 19288–19300, 2023. 1 , 3 [9] I Chen, Hua-En Chang, W ei-T ing Chen, Jenq-Neng Hwang, Sy-Y en Kuo, et al. Exploring probabilistic modeling be- yond domain generalization for semantic segmentation. In Pr oceedings of the IEEE/CVF International Confer ence on Computer V ision , pages 21755–21765, 2025. 8 [10] Shoufa Chen, Chongjian Ge, Zhan T ong, Jiangliu W ang, Y ibing Song, Jue W ang, and Ping Luo. Adaptformer: Adapting vision transformers for scalable visual recogni- tion. Advances in Neural Information Processing Systems , 35:16664–16678, 2022. 7 [11] Siyu Chen, T ing Han, Changshe Zhang, Xin Luo, Meiliu W u, Guorong Cai, and Jinhe Su. Stronger , steadier & superior: Geometric consistenc y in depth vfm for ges do- main generalized semantic segmentation. arXiv pr eprint arXiv:2504.12753 , 2025. 7 , 8 [12] Tian yi Chen, Tian yu Ding, Badal Y adav , Ilya Zharkov , and Luming Liang. Lorashear: Efficient large language model structured pruning and knowledge recovery . arXiv preprint arXiv:2310.18356 , 2023. 4 [13] Bowen Cheng, Ishan Misra, Alexander G Schwing, Alexan- der Kirillov , and Rohit Girdhar. Masked-attention mask transformer for uni versal image segmentation. In Pr oceed- ings of the IEEE/CVF conference on computer vision and pattern r ecognition , pages 1290–1299, 2022. 13 [14] Sungha Choi, Sanghun Jung, Huiwon Y un, Joanne T Kim, Seungryong Kim, and Jaegul Choo. Robustnet: Improving domain generalization in urban-scene segmentation via in- stance selecti ve whitening. In Pr oceedings of the IEEE/CVF confer ence on computer vision and pattern reco gnition , pages 11580–11590, 2021. 1 [15] Marius Cordts, Mohamed Omran, Sebastian Ramos, Timo Rehfeld, Markus Enzweiler, Rodrigo Benenson, Uwe Franke, Stefan Roth, and Bernt Schiele. The cityscapes dataset for semantic urban scene understanding. In Pr oceed- ings of the IEEE confer ence on computer vision and pattern r ecognition , pages 3213–3223, 2016. 7 [16] Jia Deng, W ei Dong, Richard Socher , Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern r ecognition , pages 248–255. Ieee, 2009. 1 [17] Jian Ding, Nan Xue, Gui-Song Xia, Bernt Schiele, and Dengxin Dai. Hgformer: Hierarchical grouping transformer for domain generalized semantic segmentation. In Pr oceed- ings of the IEEE/CVF conference on computer vision and pattern r ecognition , pages 15413–15423, 2023. 8 [18] Ning Ding, Xingtai Lv , Qiaosen W ang, Y ulin Chen, Bowen Zhou, Zhiyuan Liu, and Maosong Sun. Sparse low-rank adaptation of pre-trained language models. arXiv pr eprint arXiv:2311.11696 , 2023. 3 [19] Bryn Elesedy , V arun Kanade, and Y ee Whye T eh. Lottery tickets in linear models: An analysis of iterative magnitude pruning. arXiv preprint , 2020. 4 [20] Mohammad Fahes, T uan-Hung V u, Andrei Bursuc, Patrick P ´ erez, and Raoul De Charette. A simple recipe for language- guided domain generalized segmentation. In Pr oceedings of the IEEE/CVF Conference on Computer V ision and P attern Recognition , pages 23428–23437, 2024. 3 [21] Y uxin F ang, Quan Sun, Xinggang W ang, T iejun Huang, Xin- long W ang, and Y ue Cao. Eva-02: A visual representation for neon genesis. Image and V ision Computing , 149:105171, 2024. 3 [22] Ruili Feng, Kecheng Zheng, Y ukun Huang, Deli Zhao, Michael Jordan, and Zheng-Jun Zha. Rank diminishing in deep neural networks. Advances in Neural Information Pr o- cessing Systems , 35:33054–33065, 2022. 5 [23] Jonathan Frankle and Michael Carbin. The lottery ticket hy- pothesis: Finding sparse, trainable neural networks. arXiv pr eprint arXiv:1803.03635 , 2018. 9 [24] Jose L G ´ omez, Manuel Silva, Antonio Seoane, Agn ` es Borr ´ as, Mario Noriega, Germ ´ an Ros, Jose A Iglesias- Guitian, and Antonio M L ´ opez. All for one, and one for all: Urbansyn dataset, the third musketeer of synthetic driv- ing scenes. Neur ocomputing , 637:130038, 2025. [25] Naibin Gu, Peng Fu, Xiyu Liu, Bo wen Shen, Zheng Lin, and W eiping W ang. Light-peft: Lightening parameter- efficient fine-tuning via early pruning. arXiv pr eprint arXiv:2406.03792 , 2024. 2 , 4 [26] Song Han, Jeff Pool, John Tran, and W illiam Dally . Learn- ing both weights and connections for ef ficient neural net- work. Advances in neural information pr ocessing systems , 28, 2015. 4 [27] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceed- ings of the IEEE confer ence on computer vision and pattern r ecognition , pages 770–778, 2016. 1 [28] Y ang He, Y uhang Ding, Ping Liu, Linchao Zhu, Hanwang Zhang, and Y i Y ang. Learning filter pruning criteria for deep con volutional neural networks acceleration. In Pr oceedings of the IEEE/CVF conference on computer vision and pattern r ecognition , pages 2009–2018, 2020. 4 [29] Edward J Hu, Y elong Shen, Phillip W allis, Ze yuan Allen- Zhu, Y uanzhi Li, Shean W ang, Lu W ang, W eizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR , 1(2):3, 2022. 2 , 3 , 7 , 14 [30] Jiajun Hu, Jian Zhang, Lei Qi, Y inghuan Shi, and Y ang Gao. Learn to preserve and div ersify: Parameter-ef ficient group with orthogonal re gularization for domain generalization. In Eur opean Confer ence on Computer V ision , pages 198–216. Springer , 2024. 2 , 7 [31] Qiushi Huang, T om Ko, Zhan Zhuang, Lilian T ang, and Y u Zhang. Hira: Parameter-ef ficient hadamard high-rank adap- tation for large language models. In The Thirteenth Inter- national Confer ence on Learning Representations , 2025. 2 , 5 [32] Zhongzhan Huang, W enqi Shao, Xinjiang W ang, Liang Lin, and Ping Luo. Rethinking the pruning criteria for conv olu- tional neural network. Advances in Neural Information Pr o- cessing Systems , 34:16305–16318, 2021. 4 [33] Christoph H ¨ ummer , Manuel Schwonberg, Liangwei Zhou, Hu Cao, Alois Knoll, and Hanno Gottschalk. Strong but simple: A baseline for domain generalized dense perception by clip-based transfer learning. In Proceedings of the Asian Confer ence on Computer V ision , pages 4223–4244, 2024. 7 , 8 [34] Seogkyu Jeon, Kibeom Hong, and Hyeran Byun. Exploiting domain properties in language-dri ven domain generalization for semantic segmentation. In Pr oceedings of the IEEE/CVF International Conference on Computer V ision , pages 20791– 20801, 2025. 7 , 8 [35] Menglin Jia, Luming T ang, Bor -Chun Chen, Claire Cardie, Serge Belongie, Bharath Hariharan, and Ser-Nam Lim. V i- sual prompt tuning. In European confer ence on computer vision , pages 709–727. Springer , 2022. 7 [36] Y uru Jia, Lukas Hoyer, Shengyu Huang, T ianfu W ang, Luc V an Gool, K onrad Schindler , and Anton Obukhov . Dgin- style: Domain-generalizable semantic segmentation with image diffusion models and stylized semantic control. In Eur opean Confer ence on Computer V ision , pages 91–109. Springer , 2024. 3 [37] Christoph Kamann and Carsten Rother . Increasing the ro- bustness of semantic segmentation models with painting- by-numbers. In European Confer ence on Computer V ision , pages 369–387. Springer , 2020. 3 [38] Jaeill Kim, W onseok Lee, Moonjung Eo, and W onjong Rhee. Improving forward compatibility in class incremental learn- ing by increasing representation rank and feature richness. Neural Networks , 183:106969, 2025. 2 , 5 [39] Sunghwan Kim, Dae-hwan Kim, and Hoseong Kim. T exture learning domain randomization for domain generalized seg- mentation. In Proceedings of the IEEE/CVF International Confer ence on Computer V ision , pages 677–687, 2023. 3 [40] Y oav Kurtz, Noga Bar, and Raja Giryes. Group orthogonal- ization regularization for vision models adaptation and ro- bustness. arXiv pr eprint arXiv:2306.10001 , 2023. 2 [41] Daniel Lee and H Sebastian Seung. Algorithms for non- negati ve matrix factorization. Advances in neural informa- tion pr ocessing systems , 13, 2000. 2 [42] Jaeho Lee, Sejun Park, Sangwoo Mo, Sungsoo Ahn, and Jin- woo Shin. Layer-adaptiv e sparsity for the magnitude-based pruning. arXiv preprint , 2020. 4 [43] Suhyeon Lee, Hongje Seong, Seongwon Lee, and Euntai Kim. W ildnet: Learning domain generalized semantic seg- mentation from the wild. In Proceedings of the IEEE/CVF confer ence on computer vision and pattern reco gnition , pages 9936–9946, 2022. 3 [44] Guiying Li, Chao Qian, Chunhui Jiang, Xiaofen Lu, and Ke T ang. Optimization based layer-wise magnitude-based prun- ing for dnn compression. In IJCAI , pages 2383–2389, 2018. 4 [45] Hao Li, Asim Kadav , Igor Durdanovic, Hanan Samet, and Hans Peter Graf. Pruning filters for ef ficient con vnets. arXiv pr eprint arXiv:1608.08710 , 2016. 4 [46] Vladislav Lialin, Sherin Muckatira, Namrata Shi vagunde, and Anna Rumshisky . ReloRA: High-rank training through low-rank updates. In The T welfth International Confer ence on Learning Repr esentations , 2024. 2 , 5 [47] Dongze Lian, Daquan Zhou, Jiashi Feng, and Xinchao W ang. Scaling & shifting your features: A new baseline for efficient model tuning. Advances in Neural Information Pr ocessing Systems , 35:109–123, 2022. 7 [48] Jian Liang, W enke Huang, Guancheng W an, Qu Y ang, and Mang Y e. Lorasculpt: Sculpting lora for harmonizing gen- eral and specialized knowledge in multimodal large language models. In Pr oceedings of the Computer V ision and P attern Recognition Confer ence , pages 26170–26180, 2025. 3 , 4 [49] Shih-Y ang Liu, Chien-Yi W ang, Hongxu Y in, Pavlo Molchanov , Y u-Chiang Frank W ang, Kwang-T ing Cheng, and Min-Hung Chen. Dora: W eight-decomposed low-rank adaptation. In F orty-first International Conference on Ma- chine Learning , 2024. 7 [50] Zihang Liu, T ianyu Pang, Oleg Balabanov , Chaoqun Y ang, T ianjin Huang, Lu Y in, Y aoqing Y ang, and Shiwei Liu. Lift 10 the veil for the truth: Principal weights emerge after rank re- duction for reasoning-focused supervised fine-tuning. arXiv pr eprint arXiv:2506.00772 , 2025. 4 [51] Fanxu Meng, Zhaohui W ang, and Muhan Zhang. Pissa: Prin- cipal singular v alues and singular vectors adaptation of large language models. Advances in Neural Information Pr ocess- ing Systems , 37:121038–121072, 2024. 2 , 3 , 6 , 14 [52] Rang Meng, Xianfeng Li, W eijie Chen, Shicai Y ang, Jie Song, Xinchao W ang, Lei Zhang, Mingli Song, Di Xie, and Shiliang Pu. Attention diversification for domain general- ization. In European confer ence on computer vision , pages 322–340. Springer , 2022. 2 [53] Gerhard Neuhold, T obias Ollmann, Samuel Rota Bulo, and Peter K ontschieder . The mapillary vistas dataset for semantic understanding of street scenes. In Proceedings of the IEEE international confer ence on computer vision , pages 4990– 4999, 2017. 7 [54] Joshua Niemeijer , Manuel Schwonberg, Jan-Aike T erm ¨ ohlen, Nico M Schmidt, and T im Fingscheidt. Generalization by adaptation: Dif fusion-based domain extension for domain-generalized semantic segmentation. In Pr oceedings of the IEEE/CVF winter conference on applications of computer vision , pages 2830–2840, 2024. 3 [55] Maxime Oquab, Timoth ´ ee Darcet, Th ´ eo Moutakanni, Huy V o, Marc Szafraniec, V asil Khalidov , Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby , et al. Dinov2: Learning rob ust visual features without supervision. arXiv pr eprint arXiv:2304.07193 , 2023. 1 , 3 [56] Byeonghyun Pak, Byeongju W oo, Sunghwan Kim, Dae- hwan Kim, and Hoseong Kim. T extual query-dri ven mask transformer for domain generalized segmentation. In Eur opean Confer ence on Computer V ision , pages 37–54. Springer , 2024. 2 , 3 , 7 , 8 [57] Xingang Pan, Ping Luo, Jianping Shi, and Xiaoou T ang. T wo at once: Enhancing learning and generalization capacities via ibn-net. In Pr oceedings of the eur opean conference on computer vision (ECCV) , pages 464–479, 2018. 3 [58] Duo Peng, Y injie Lei, Lingqiao Liu, Pingping Zhang, and Jun Liu. Global and local texture randomization for synthetic-to-real semantic segmentation. IEEE T ransactions on Image Pr ocessing , 30:6594–6608, 2021. 1 [59] Duo Peng, Y injie Lei, Munaw ar Hayat, Y ulan Guo, and W en Li. Semantic-aware domain generalized segmentation. In Pr oceedings of the IEEE/CVF conference on computer vi- sion and pattern r ecognition , pages 2594–2605, 2022. 1 , 3 [60] Alec Radford, Jong W ook Kim, Chris Hallacy , Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry , Amanda Askell, P amela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. In International conference on mac hine learning , pages 8748–8763. PmLR, 2021. 1 , 3 [61] Stephan R Richter, V ibhav V ineet, Stefan Roth, and Vladlen K oltun. Playing for data: Ground truth from computer games. In Eur opean confer ence on computer vision , pages 102–118. Springer , 2016. 7 [62] German Ros, Laura Sellart, Joanna Materzynska, David V azquez, and Antonio M Lopez. The synthia dataset: A large collection of synthetic images for semantic segmentation of urban scenes. In Proceedings of the IEEE conference on computer vision and pattern r ecognition , pages 3234–3243, 2016. [63] Olivier Roy and Martin V etterli. The ef fectiv e rank: A mea- sure of ef fectiv e dimensionality . In 2007 15th Eur opean sig- nal pr ocessing confer ence , pages 606–610. IEEE, 2007. 5 [64] Mark Sandler , Andre w Ho ward, Menglong Zhu, Andrey Zh- moginov , and Liang-Chieh Chen. Mobilenetv2: In verted residuals and linear bottlenecks. In Pr oceedings of the IEEE confer ence on computer vision and pattern recogni- tion , pages 4510–4520, 2018. 1 [65] Namrata Shiv agunde, Mayank Kulkarni, Giannis Kara- manolakis, Jack G. M. FitzGerald, Y annick V ersley , Saleh Soltan, V olkan Cevher , Jianhua Lu, and Anna Rumshisky . Approximations may be all you need: T o wards pre-training llms with low-rank decomposition and optimizers. 2024. [66] Mingjie Sun, Zhuang Liu, Anna Bair, and J Zico K olter . A simple and effecti ve pruning approach for large language models. arXiv preprint , 2023. [67] Xinglong Sun and Humphrey Shi. T owards better structured pruning saliency by reorg anizing con volution. In Pr oceed- ings of the IEEE/CVF W inter Conference on Applications of Computer V ision , pages 2204–2214, 2024. 4 [68] PeiY uan T ang, Xiaodong Zhang, Chunze Y ang, Haoran Y uan, Jun Sun, Danfeng Shan, and Zijiang James Y ang. Un- leashing the power of visual foundation models for general- izable semantic segmentation. In Pr oceedings of the AAAI Confer ence on Artificial Intelligence , pages 20823–20831, 2025. [69] Sumanth Udupa, Prajwal Gurunath, Aniruddh Sikdar , and Suresh Sundaram. Mrfp: Learning generalizable semantic segmentation from sim-2-real with multi-resolution feature perturbation. In Pr oceedings of the IEEE/CVF Confer ence on Computer V ision and P attern Recognition , pages 5904– 5914, 2024. 1 , 3 [70] Zijian W ang, Y adan Luo, Ruihong Qiu, Zi Huang, and Mahsa Baktashmotlagh. Learning to diversify for single do- main generalization. In Pr oceedings of the IEEE/CVF in- ternational confer ence on computer vision , pages 834–843, 2021. 2 [71] Zhixiang W ei, Lin Chen, Y i Jin, Xiaoxiao Ma, T ianle Liu, Pengyang Ling, Ben W ang, Huaian Chen, and Jinjin Zheng. Stronger fe wer & superior: Harnessing vision foundation models for domain generalized semantic segmentation. In Pr oceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition , pages 28619–28630, 2024. 2 , 3 , 7 , 8 , 13 [72] Zhenyao Wu, Xinyi W u, Xiaoping Zhang, Lili Ju, and Song W ang. Siamdoge: Domain generalizable semantic segmen- tation using siamese network. In European Conference on Computer V ision , pages 603–620. Springer, 2022. 3 [73] Seunghan Y ang, Seokeon Choi, Hyunsin Park, Sungha Choi, Simyung Chang, and Sungrack Y un. Feature di versifica- tion and adaptation for federated domain generalization. In Eur opean Confer ence on Computer V ision , pages 52–70. Springer , 2024. 2 11 [74] Jingjun Y i, Qi Bi, Hao Zheng, Haolan Zhan, W ei Ji, Y awen Huang, Y uexiang Li, and Y efeng Zheng. Learning spectral- decomposited tokens for domain generalized semantic seg- mentation. In Proceedings of the 32nd ACM International Confer ence on Multimedia , pages 8159–8168, 2024. 7 , 8 [75] Fisher Y u, Haofeng Chen, Xin W ang, W enqi Xian, Y ingying Chen, Fangchen Liu, V ashisht Madhav an, and T re vor Dar- rell. Bdd100k: A diverse driving dataset for heterogeneous multitask learning. In Pr oceedings of the IEEE/CVF con- fer ence on computer vision and pattern r ecognition , pages 2636–2645, 2020. 7 [76] Seokju Y un, Seunghye Chae, Dongheon Lee, and Y oungmin Ro. Soma: Singular v alue decomposed minor components adaptation for domain generalizable representation learning. In Pr oceedings of the Computer V ision and P attern Recogni- tion Conference , pages 25602–25612, 2025. 2 , 3 , 5 , 6 , 7 , 8 , 13 , 14 [77] Xin Zhang and Robby T T an. Mamba as a bridge: Where vision foundation models meet vision language models for domain-generalized semantic segmentation. In Pr oceedings of the Computer V ision and P attern Recognition Confer ence , pages 14527–14537, 2025. 2 , 3 , 7 , 8 [78] Zhun Zhong, Y uyang Zhao, Gim Hee Lee, and Nicu Sebe. Adversarial style augmentation for domain general- ized urban-scene segmentation. Advances in neural infor- mation pr ocessing systems , 35:338–350, 2022. 3 12 RecycleLoRA: Rank-Rev ealing QR-Based Dual-LoRA Subspace Adaptation f or Domain Generalized Semantic Segmentation Supplementary Material This supplement provides additional materials omitted from the main te xt to facilitate a deeper understanding of our pro- posed RecycleLoRA. Contents A . Implementation Details 13 B . Additional Experiments and Analysis 13 B.1 . Hyperparameter Analysis . . . . . . . . . . . 13 B.2 . Ablation on Dual-Adapter Initialization . . . 14 C . Limitations and Future W orks 14 D . Qualitative Results 14 A. Implementation Details Our method is implemented based on the MMSegmen- tation codebase. W e use DINOv2-Large as the back- bone and Mask2Former as the decode head. Following the experimental setups of Rein [ 71 ] and SoMA [ 76 ], we only utilize the default data augmentation provided in Mask2Former [ 13 ] to ensure a fair comparison. All mod- els are trained on NVIDIA A6000 GPUs. Further details on hyperparameters are provided in T able 8 . Unless stated oth- erwise, all experiments in the main paper were conducted using these settings. Hyperparameter Synthetic-to-Real Real-to-Real backbone DINOv2-L DINOv2-L main rank 32 32 sub rank 4 2 main lr mult. 1.0 1.0 sub lr mult. 0.5 0.5 learning rate 1e-4 1e-4 backbone lr mult. 0.5 0.5 lr scheduler PolyLR PolyLR A WD scheduler Cosine Cosine weight decay 0.05 0.05 optimizer AdamW AdamW batch size 4 4 iterations 40,000 40,000 T able 8. Hyperparameter settings for experiments. B. Additional Experiments and Analysis B.1. Hyperparameter Analysis Rank Analysis. T o determine the optimal configuration for our dual-adapter structure, we conducted an analysis to in- vestigate the impact of the rank settings for both the Main and Sub Adapters on domain generalization performance. First, for the synthetic-to-real scenario (T able 9 ), we found that a Main Adapter rank of 32 consistently outper- formed a rank of 16. With the Main Adapter’ s rank fixed at 32, we observed that performance peaked at an aver - age mIoU of 68.95 when the Sub Adapter’ s rank was 4. Howe v er , increasing the rank further to 8 or higher led to a noticeable degradation in performance. This finding is consistent with our hypothesis that the Sub Adapter , which modifies the VFM’ s major directions, requires minimal and careful adjustments. Therefore, we adopted the (32, 4) rank configuration for the synthetic-to-real experiments (e.g., T a- ble 3 , 12 ). W e extended this analysis to the real-to-real generaliza- tion scenario (T able 10 ) as well. In this setting, the best per - formance (72.10 mIoU) was achiev ed with a Main Adapter rank of 32 and a Sub Adapter rank of 2. This result suggests that for the real-to-real setting, which has a smaller domain gap, an ev en more conserv ati ve adjustment of the VFM’ s major directions is beneficial. Consequently , we used the (32, 2) rank configuration for all real-to-real experiments (e.g., T able 4 ). Synthetic-to-Real Generalization rank lr Params. T rained on GT A V main sub main sub → Citys. → BDD → Map. A vg. 32 2 1e-4 5e-5 13.4M 72.03 60.75 72.06 68.28 32 4 1e-4 5e-5 14.2M 73.01 61.77 72.07 68.95 32 8 1e-4 5e-5 15.7M 72.83 61.13 70.81 68.26 32 16 1e-4 5e-5 18.9M 71.20 61.16 70.87 67.74 32 32 1e-4 5e-5 25.2M 72.13 60.83 69.87 67.61 16 2 1e-4 5e-5 7.1M 71.67 61.27 70.84 67.93 16 4 1e-4 5e-5 7.9M 71.74 61.54 71.25 68.18 16 8 1e-4 5e-5 9.4M 71.40 60.56 70.41 67.46 16 16 1e-4 5e-5 12.6M 71.51 60.48 70.37 67.45 T able 9. Domain generalization results (mIoU %) for Recy- cleLoRA with varying rank configurations for its Main and Sub Adapters, under the synthetic-to-real setting (G →{ C, B, M } ). Bold and underlined indicate best and second-best results. Real-to-Real Generalization rank lr Params. T rained on Cityscapes main sub main sub → BDD → Map. A vg. 32 2 1e-4 5e-5 13.4M 66.65 78.14 72.10 32 4 1e-4 5e-5 14.2M 66.76 76.64 71.70 T able 10. Domain generalization results (mIoU %) for Recy- cleLoRA with varying rank configurations for its Main and Sub Adapters, under the real-to-real setting (C →{ B, M } ). Bold and underlined indicate best and second-best results. 13 Learning rate Analysis. W e conduct an analysis to deter- mine the optimal learning rate for the Sub Adapter and v ali- date our design choice of using a different learning rate from the Main Adapter . As shown in T able 11 , we fixed the Main Adapter’ s learning rate to 1e-4 and varied the Sub Adapter’ s learning rate. The results demonstrate that the best perfor- mance is achiev ed when the Sub Adapter’ s learning rate is set to 5e-5, half that of the Main Adapter, achieving an av- erage mIoU of 68.95. Setting the learning rate for the Sub Adapter to be either the same as the Main Adapter (1e-4) or excessiv ely low (1e-5) resulted in a performance drop. This empirical evidence supports our strategy of applying a carefully tuned, lower learning rate to the Sub Adapter . This differentiation is crucial for enabling the complemen- tary learning process between the two adapters, v alidating our ov erall design. Synthetic-to-Real Generalization rank lr Params. T rained on GT A V main sub main sub → Citys. → BDD → Map. A vg. 32 4 1e-4 1e-4 14.2M 72.23 60.50 70.60 67.78 32 4 1e-4 5e-5 14.2M 73.01 61.77 72.07 68.95 32 4 1e-4 1e-5 14.2M 72.55 61.10 70.55 68.07 T able 11. Domain generalization results (mIoU %) for Recy- cleLoRA with varying learning rate configurations for its Main and Sub Adapters, under the synthetic-to-real setting (G →{ C, B, M } ). Bold and underlined indicate best and second-best results. B.2. Ablation on Dual-Adapter Initialization T o assess the impact of the initialization strategy on our dual-adapter framework, we conducted an ablation study comparing our RRQR-based approach with other represen- tativ e initialization methods. W e compare against Kaim- ing uniform initialization, a standard method that does not le verage the pre-trained weight structure, and SVD- based initialization, which utilizes subspace decomposition as seen in prior work such as SoMA [ 76 ] or PiSSA [ 51 ]. Interestingly , as presented in T able 12 , the other initial- ization methods did not synergize with the dual-adapter structure and instead exhibited performance degradation. Specifically , the dual-adapter with Kaiming initialization scored 66.31 mIoU, which is lower than the 66.89 mIoU of standard LoRA [ 29 ] (single adapter) presented in T a- ble 3 . Similarly , the SVD-based initialization achieved only 67.23 mIoU, underperforming SoMA (single adapter), which scored 68.27 mIoU as shown in T able 3 . In con- trast, our proposed RRQR-based initialization achieves an av erage mIoU of 68.95, significantly outperforming both al- ternativ es. These results underscore the importance of the initialization method and suggest that our proposed RRQR- based strategy is a more effecti ve choice for the proposed dual-adapter structure. Synthetic-to-Real Generalization Initialization Backbone T rained on GT A V → Citys. → BDD → Map. A vg. Kaiming unif. DINOv2-L 68.96 60.57 69.41 66.31 SVD DINOv2-L 70.40 60.35 70.93 67.23 RRQR DINOv2-L 73.01 61.77 72.07 68.95 T able 12. Domain generalization results (mIoU %) for the dual- adapter frame work with dif ferent initialization strategies, under the synthetic-to-real setting (G →{ C, B, M } ). Bold and underlined indicate best and second-best results. C. Limitations and Future W orks While RecycleLoRA demonstrates robust state-of-the-art performance, we identify several avenues for future re- search that could further advance its capabilities and address its current limitations. Further optimization of RecycleLoRA could be achieved by dev eloping systematic methods for tuning hyperparam- eters, such as the ranks and learning rates of the dual adapters. The current configuration was determined empir- ically , and creating automated search strategies would en- hance the practicality and replicability of our approach. The scope of our work, currently focused on Domain Generalized Semantic Segmentation, could also be broad- ened. The core principle of rec ycling pre-trained kno wl- edge through subspace analysis is likely applicable to other downstream tasks that require efficient foundation model fine-tuning, such as object detection, video analysis, and medical image segmentation. Inv estigating the effecti ve- ness of our approach in these di verse contexts presents a logical direction for future work. Another promising research direction in volves re visiting our methodology’ s binary partitioning of the VFM’ s sub- space into “major” and “minor” directions, which currently omits the intermediate directions. This simplification is based on the hypothesis that the most and least dominant directions are most critical for balancing knowledge preser - vation and new feature acquisition. Howe ver , the potential contribution of these intermediate directions remains unex- plored. Future research could in vestigate a more nuanced allocation of the entire subspace spectrum, perhaps through a third adapter or a soft-weighting scheme that utilizes all ranked directions, which may unlock further performance gains. D. Qualitative Results Figures 6 and 7 present qualitativ e comparisons against other state-of-the-art methods. These visualizations high- light that RecycleLoRA generates more accurate and de- tailed segmentation maps, which are more closely aligned with the ground truth. 14 Input MFuser SoMA RecycleLoRA G T road sidew. build. wall fence pole light sign veget. terrain sky per son rider car truc k bus train motor . bicycle n/a. Figure 6. Qualitativ e comparison of semantic segmentation on the Cityscapes. All models were trained on the GT A V . Input MFuser SoMA RecycleLoRA G T road sidew . build. wall fence pole light sign veget. terrain sky person rider car truck bus train motor. bicycle n/a. Figure 7. Qualitativ e comparison of semantic segmentation on the Mapillary . All models were trained on the GT A V . 15