Paper deep dive
LumosX: Relate Any Identities with Their Attributes for Personalized Video Generation
Jiazheng Xing, Fei Du, Hangjie Yuan, Pengwei Liu, Hongbin Xu, Hai Ci, Ruigang Niu, Weihua Chen, Fan Wang, Yong Liu
Intelligence
Status: succeeded | Model: google/gemini-3.1-flash-lite-preview | Prompt: intel-v1 | Confidence: 95%
Last extracted: 3/23/2026, 12:13:14 PM
Summary
LumosX is a framework for personalized multi-subject video generation that addresses face-attribute misalignment and intra-group consistency. It introduces a data collection pipeline using MLLMs to extract relational priors and a model architecture featuring Relational Self-Attention and Relational Cross-Attention to explicitly bind identities with their attributes.
Entities (6)
Relation Signals (3)
LumosX ā builton ā Wan2.1
confidence 100% Ā· LumosX is built upon the Wan2.1 text-to-video backbone
LumosX ā utilizes ā Relational Self-Attention
confidence 95% Ā· LumosX explicitly encodes face-attribute bindings into coherent subject groups through two dedicated modules: Relational Self-Attention and Relational Cross-Attention.
LumosX ā utilizes ā Relational Cross-Attention
confidence 95% Ā· LumosX explicitly encodes face-attribute bindings into coherent subject groups through two dedicated modules: Relational Self-Attention and Relational Cross-Attention.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Recent advances in diffusion models have significantly improved text-to-video generation, enabling personalized content creation with fine-grained control over both foreground and background elements. However, precise face-attribute alignment across subjects remains challenging, as existing methods lack explicit mechanisms to ensure intra-group consistency. Addressing this gap requires both explicit modeling strategies and face-attribute-aware data resources. We therefore propose LumosX, a framework that advances both data and model design. On the data side, a tailored collection pipeline orchestrates captions and visual cues from independent videos, while multimodal large language models (MLLMs) infer and assign subject-specific dependencies. These extracted relational priors impose a finer-grained structure that amplifies the expressive control of personalized video generation and enables the construction of a comprehensive benchmark. On the modeling side, Relational Self-Attention and Relational Cross-Attention intertwine position-aware embeddings with refined attention dynamics to inscribe explicit subject-attribute dependencies, enforcing disciplined intra-group cohesion and amplifying the separation between distinct subject clusters. Comprehensive evaluations on our benchmark demonstrate that LumosX achieves state-of-the-art performance in fine-grained, identity-consistent, and semantically aligned personalized multi-subject video generation. Code and models are available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2603.20192v1
- Canonical: https://arxiv.org/abs/2603.20192v1
Trouble viewing inline? Open PDF directly ā
Full Text
99,690 characters extracted from source content.
Expand or collapse full text
LumosX : Relate Any Identities with Their Attributes for Personalized Video Generation Jiazheng Xing ā1,4,2 , Fei Du ā2,3 Hangjie Yuan ā2,3,1 , Pengwei Liu 1,2 , Hongbin Xu 4 , Hai Ci 4 , Ruigang Niu 2,3 , Weihua Chen ā 2,3 , Fan Wang 2 , Yong Liu ā 1 1 Zhejiang University, 2 DAMO Academy, Alibaba Group, 3 Hupan Lab, 4 National University of Singapore * Equal contribution, ā Corresponding authors. jiazhengxing@zju.edu.cn, kugang.cwh@alibaba-inc.com, yongliu@iipc.zju.edu.cn Project Page: https://jiazheng-xing.github.io/lumosx-home/ Recent advances in diffusion models have significantly improved text-to-video generation, enabling personalized content creation with fine-grained control over both foreground and background elements. However, precise faceāattribute alignment across subjects remains challenging, as existing methods lack explicit mechanisms to ensure intra-group consistency. Addressing this gap requires both explicit modeling strategies and face-attribute-aware data resources. We therefore proposeLumosX, a framework that advances both data and model design. On the data side, a tailored collection pipeline orchestrates captions and visual cues from independent videos, while multimodal large language models (MLLMs) infer and assign subject-specific dependencies. These extracted relational priors impose a finer-grained structure that amplifies the expressive control of personalized video generation and enables the construction of a comprehensive benchmark. On the modeling side, Relational Self-Attention and Relational Cross-Attention intertwine position-aware embeddings with refined attention dynamics to inscribe explicit subjectāattribute dependencies, enforcing disciplined intra-group cohesion and amplifying the separation between distinct subject clusters. Comprehensive evaluations on our benchmark demonstrate thatLumosXachieves state-of-the-art performance in fine-grained, identity-consistent, and semantically aligned personalized multi-subject video generation. Date: March 23, 2026 1 Introduction In recent years, diffusion models [11,12,33] have driven remarkable progress, establishing new performance standards in text-to-video generation [13,3,28,51], particularly through the adoption of Diffusion Transformer (DiT) architectures [31]. These advances have laid a solid groundwork for customized video generation [27,48, 52,5,15,8,24], where high-degree-of-freedom personalization unlocks transformative applications ranging from virtual theatrical production to e-commerceāenabling fine-grained control over both backgrounds and foregrounds, including multiple interacting subjects. Yet, realizing open-set personalized multi-subject video generation under such flexible and complex conditions remains profoundly challenging. The task requires not only the precise integration of diverse and interrelated conditioning signals but also the preservation of temporal coherence and identity fidelity across all subjects. In the realm of open-set personalized video generation, prior studies have pushed the field forward from distinct angles. Certain approaches [27,10,48,50,52] concentrate narrowly on foreground facial customization, preserving identity fidelity from reference images yet affording only limited flexibility in input specification. In contrast, more recent methods [5,15,8,24] enable highly versatile multi-subject video personalization with controllable foregrounds and backgrounds, but they largely neglect the intrinsic dependency structures that govern multi-subject conditions. Crucially, during fine-grained multi-condition injection, conditioning signals for each subject are typically decomposed into facial exemplars and attribute descriptors (e.g., man: blond hair, white T-shirt, sunglasses). Absent an explicit mechanism to bind identity with its associated attributes, such formulations are inherently fragile and frequently yield attribute entanglement or faceāattribute misalignment 1 arXiv:2603.20192v1 [cs.CV] 20 Mar 2026 A man in a black shirt with graying hair and a woman in a gray shirt sit close to each other under a green tent in a park, engaged in a serious conversation. man black shirt woman gray shirt green tent park LumosX Two men are seated next to each other in a casual indoor setting, captured from an eye-level camera angle. The man on the left has long, wavy hair and a beard, wearing a green button-up shirt. The man on the right is wearing a black t-shirt and a black baseball cap. Between them are two green glass bottles, placed on a table. Behind them, a framed painting with an abstract design hangs on a light-colored wall. LumosX man on the left man on the right Figure 1 LumosX supports flexible personalized multi-subject video generation. across subjects. Although implicit modeling via textual captions can capture simple multi-subject dependencies during video generation, ambiguity often arises when captions contain similar subject nouns, such as āA man on the left with ... and a man on the right with ...," leading to confusion in subjectāattribute associations. To overcome this limitation, under fine-grained multi-subject inputs, explicit constraints must be imposed at both the data and model levels. (1) Data level: When visual references are provided, the correspondence between each face and its associated attributes should be clearly specified. (2) Model level: During generation, each face-attribute pair is explicitly bound into an independent subject group, with intra-group correlation enhanced and inter-group interference suppressed. To address the challenge of modeling face-attribute dependencies in multi-subject video generation, we present LumosX, a novel framework for personalized multi-subject synthesis. On the data side, the absence of public datasets with annotated dependency structures motivates us to construct a collection pipeline that supports open-set entities. This pipeline extracts captions and foregroundābackground visual conditions from independent videos, while multimodal large language models (MLLMs) infer and assign subject-specific dependencies. In particular, it produces customized single- and multi-subject data with explicit faceāattribute correspondences, which not only enhance personalization during modeling but also enable the construction of a comprehensive benchmark. On this basis, the benchmark further defines two evaluation tasks, identity- consistent and subject-consistent generation, which allow a systematic assessment of a modelās ability to preserve identity and align multi-subject relationships. On the modeling side,LumosXexplicitly encodes face-attribute bindings into coherent subject groups through two dedicated modules: Relational Self-Attention and Relational Cross-Attention. The Relational Self-Attention module incorporates Relational Rotary Position Embedding (R2PE) and a Causal Self-Attention Mask (CSAM) to model dependencies at the positional encoding and spatio-temporal self-attention stages. In addition, the Relational Cross-Attention module introduces a Multilevel Cross-Attention Mask (MCAM), which reinforces intra-group coherence, suppresses cross-group interference, and refines the semantic representation of visual condition tokens.LumosXis built upon the Wan2.1 [40] text-to-video backbone, with our modules seamlessly integrated to support flexible and high-fidelity personalized multi-subject generation, as shown in Fig. 1. Extensive experiments demonstrate Lumos-Xās strong capability in producing fine-grained, identity-consistent, and semantically aligned personalized videos, achieving state-of-the-art results across diverse benchmarks. The contributions of our LumosX can be summarized as follows: ā¢Data Side. We build a collection pipeline for open-set multi-subject generation that extracts captions and foregroundābackground condition images with explicit faceāattribute dependencies from independent videos. This yields finer-grained relational priors that enhance personalized video customization and enable the construction of reliable benchmarks. 2 ā¢Model Side. We introduce Relational Self-Attention and Relational Cross-Attention, which integrate relational positional encodings with structured attention masks to explicitly encode faceāattribute bind- ings. This reinforces intra-group coherence, mitigates cross-group interference, and ensures semantically consistent multi-subject video generation. ā¢Overall Performance. Through extensive experiments and comparative evaluations,LumosXachieves state-of-the-art results in generating fine-grained, identity-consistent, and semantically aligned personal- ized multi-subject videos, decisively outperforming advanced open-source approaches including Phantom and SkyReels-A2. 2 Related Works Video Generation. Video generation has advanced rapidly in recent years, becoming one of the most dynamic research areas. Early works based on generative adversarial networks (GANs) [39,38] demonstrated initial video synthesis but struggled with temporal coherence and fidelity. Latent diffusion models (LDMs) [33], powered by UNet [34], marked a significant milestone by enabling high-quality video generation through denoising in compressed latent spaces. These works typically add a temporal module to an image generation model, such as Make-A-Video [36] and Animatediff [9]. However, these models often face scalability challenges when scaling to larger parameter sizes or higher resolutions. Diffusion Transformers (DiTs) [31], replacing the Unet backbone with Transformer blocks, have shown superior performance in visual generation. By incorporating spatio-temporal attention mechanisms, video DiTs achieved unprecedented performance in modeling long-range dependencies across both spatial and temporal dimensions, significantly enhancing video realism and consistency. Models like Hunyuan Video [20], Wan2.1 [40], and MAGI-1 [35] have scaled the parameters of video DiTs to more than 10 billion, achieving significant advancement. Despite these advances, controllability remains a critical bottleneck: text-driven generation often fails to precisely align with user intentions due to ambiguities in natural language descriptions. This work focuses on multi-subject video customization to address this limitation, enabling precise content-control video generation. Multi-Subject Video Customization. In recent years, subject-driven video generation has attracted growing interest. Several works focus on ID-consistent video generation, such as Magic-Me [27], ID-Animator [10], ConsisID [48], Magic Mirror [49], FantasyID [50], and Concat-ID [52]. These works generate videos that show consistent identity with the reference images, mainly focusing on facial identity. For arbitrary subject customization, VideoBooth [17] incorporates high-level and fine-level visual cues from an image prompt to the video generation model via cross attention and cross-frame attention. DreamVideo [44] customizes both subject and motion, with motion extracted from a reference video. Although they have demonstrated capabilities in generating single-subject consistent videos, neither the data processing nor the model can be easily transferred to the more challenging multi-subject customization. CustomVideo [43] generates multi- subject identity-preserving videos by composing multiple subjects in a single image and designs an attention control strategy to disentangle them. However, it requires test-time finetuning for different subjects. Recently, several works [5,15,8,24] propose to customize multiple subjects in the video DiTs. Different subjects are usually concatenated and fed into the video DiT network without distinguishing between them. This lack of differentiation can lead to semantic ambiguity, especially when there are numerous targets and hierarchical relationships among them. In this work, we design several strategies to differentiate between various subjects and their hierarchical relationships, achieving consistent customization while ensuring harmony among different objectives and the textās adherence capabilities. 3 Methods 3.1 Preliminary In this work, we bulid upon the latest text-to-video generative model, Wan2.1 [40], which comprises the 3D variational autoencoder (VAE)E, the text encoderT, and the denoising DiT [31] backboneĪµ Īø combined with Flow Matching [23]. Within the DiT architecture, full spatio-temporal Self-Attention is used to capture complex dynamics, while Cross-Attention is employed to incorporate text conditions. Specifically, given a videoX=x i N i=1 withNframes,Ecompresses it into a latent representation zā R TĆHWĆC along the 3 Caption Generation Original Video Video Frames Extraction A man and a woman are seated at a table in a lush garden, sharing a meal. The man, wearing a black shirt and a black watch on his left wrist, gestures with his hands while speaking, while the woman, in a white top, listens attentively and occasionally responds. Utensils are set neatly on the tables as they enjoy their meal together... Human Detection Entity Words Retrieval & Subject-Attributes Matching Word Tags Subjects man: black shirt, black watch woman: white top Objects utensils Background lush garden GroundingDINO + SAM Face Cropping +SAM Subject & Object Masking +Inpainting Random Sampling man woman black shirt black watach white top utensils lush garden Subjects Objects Background 0.050.50 0.95 Step 1Step 2 Step 3 0.050.50 0.95 Figure 2 Dataset construction pipeline for personalized multi-subject video generation. We build the training dataset from raw videos in three steps: (1) generate a caption and detect human subjects in extracted frames; (2) retrieve entity words from the caption and match subjects with their attributes; (3) use these entity tags to localize and segment target subjects and objects, producing a clean background image. spatiotemporal dimensions, whereT,HW, andCdenote the temporal, spatial, and channel dimensions, respectively. The text encoderTtakes the text prompt and encodes it into a textual embeddingc text . The denoising DiTĪµ Īø then processes the latent representation z and the textual representationc text to predict the distribution of video content. Each DiT Block incorporates 3D Rotary Position Embedding (3D-RoPE [37]) within the full spatio-temporal attention module to better capture both temporal and spatial dependencies. 3.2 Dataset Construction As illustrated in Fig. 2, our training dataset and inference benchmark for personalized multi-subject video generation are constructed from raw videos through the following three steps. Caption Generation and Human Detection. To obtain richer textual descriptions for downstream tasks, we replace the original video captions with captions generated by the large visionālanguage model VILA [22]. We sample three frames from the beginning, middle, and end of each video (5%, 50%, and 95% positions) and apply human detection [41] to extract human subjects for subsequent faceāattribute matching. Entity Words Retrieval and Face-Attribute Matching. In this step, our goal is to retrieve entity words from the caption, which can be classified into three categories: human subjects with attributes (e.g., man: black shirt, black watch), objects (e.g., utensils), and background (e.g., lush garden). During this process, if multiple human subjects are present, we need to assign different attributes to the corresponding subjects. In particular, when the caption contains multiple instances of the same subject noun (e.g., woman), we rely on visual information to assist in distinguishing between them. Therefore, we employ the multimodal large language model Qwen2.5-VL [1] to retrieve multiple entity words from the caption, while leveraging prior visual information from human detection results to achieve precise face-attribute matching. Obtaining Condition Images. For subjects, we apply face detection [41] within human detection boxes to extract face crops and use SAM [19] to segment attribute masks. For objects, GroundingDINO [25] combined with SAM segments each entity within the global image. For backgrounds, we remove subjects and objects using the crops and masks, then apply the diffusion inpainting model FLUX [21] to generate a clean background. Finally, from the valid results of the three key frames, we randomly select one per entity as its condition 4 DiT Block ... Relational Cross-Attention Feed Forward DiT Block ... Subject ASubject BObjects Background Video Tokens In a cozy kitchen setting, a woman in a yellow scarf is mixing ingredients in a large blue bowl, while a young woman in a black hoodie watches her attentively. A small metal measuring cup is also visible on the counter... Caption Text Encoder FrozenTrainable Relational Rotary Position Embedding (R2PE) 3D VAE Decoder 3D VAE Encoder Relational Self-Attention Video Tokens Objects Tokens Background Tokens Subject A Tokens Subject B Tokens Width Height (0,0,0)(0,0,1) (0,1,0)(0,1,1) ... (T-1, 0,0) (T-1, 0,1) (T-1, 1,0) (T-1, 1,1) (T,0,0)(T,0,1) (T,1,0)(T,1,1) (T+1, 0,0) (T+1, 0,1) (T+1, 1,0) (T+1, 1,1) (T+2, 0,0) (T+2, 0,1) (T+2, 1,0) (T+2, 1,1) (T+3, 0,0) (T+3, 0,1) (T+3, 1,0) (T+3, 1,1) (T+3, 2,2) (T+3, 2,3) (T+3, 3,2) (T+3, 3,3) (T+4, 0,0) (T+4, 0,1) (T+4, 1,0) (T+4, 1,1) (T+4, 2,2) (T+4, 2,3) (T+4, 3,2) (T+4, 3,3) Figure 3 Overview of LumosX. Built on the T2V model Wan2.1 [40], our framework encodes all condition images into image tokens via a VAE encoder, concatenates them with denoising video tokens, and feeds the result into DiT [31] blocks. Within each block, the proposed Relational Self-Attention and Relational Cross-Attention enable causal conditional modeling, enhance visual token representations, and ensure precise faceāattribute alignment. imageāmatching the inference process, where each condition uses a single reference imageāwhile ensuring data diversity by preventing all selections from a single frame. Through these three steps, we obtain the visual condition images for the subjects, objects, and background, along with their paired word tags derived from the input text caption. Note that a subject is defined as a single human face paired with its corresponding attributes. The face is expected to present clear facial features without significant occlusion, and the associated attributes can include clothing (top or bottom), accessories (e.g., glasses, earrings, or necklaces), or hairstyle. 3.3 LumosX As shown in Fig. 3, our framework builds on the T2V model Wan2.1 [40]. To enable personalized multi-subject video generation, all condition images are encoded into image tokens via a VAE encoder, concatenated with denoising video tokens, and fed into DiT [31] blocks. Within each block, we introduce Relational Self- Attention with Relational Rotary Position Embedding (R2PE) and a Causal Self-Attention Mask to support spatio-temporal and causal conditional modeling. Additionally, Relational Cross-Attention with a Multilevel Cross-Attention Mask (MCAM) incorporates textual conditions, strengthens visual token representations, and aligns faceāattribute relationships. 3.3.1 Relational Self-Attention Relational Rotary Position Embedding (R2PE). In T2V models like Wan2.1 [40], utilizing 3D Rotary Position Embedding (3D-RoPE) to assign position indices (i,j,k) to video tokens is necessary, which can affect the interaction among these tokens. In T2V tasks, the original 3D-RoPE assigns position indices (i,j,k) sequentially to the video tokens zā R TĆHWĆC , whereiā[0,T),j ā[0,W), andk ā[0,H). In personalized multi-subject video generation tasks, it is essential to not only extend 3D-RoPE to the reference condition images but also to preserve the face-attribute dependency throughout this process. 5 Video Noise Video Noise Bg Face A Attr A Attr B Face B Bg Obj A Obj B Face A Attr A Face B Attr B Obj A Obj B Bg: Background (cozy kitchen setting) Attr A: Attribute A (yellow scarf) Face A: Human Face A (woman) Subject A Subject B : True: False Key Quer y (a) Causal Self-Attention Mask (CSAM) Face B: Human Face B (young woman) Attr B: Atrribute B (black hoodie) Obj A: Object A (small metal measuring cup) Obj B: Object B (large blue bowl) Objects Visual Tokens Visual Tokens Other Text Tokens Video Noise woman yellow scarf black hoddie young woman cozy kitchen setting small metal measuring cup large blue bowl Face A Attr A Face B Attr B Obj A Obj B Textual Tokens Visual Tokens (b) Multilevel Cross-Attention Mask (MCAM) Bg Strong Correlation Weak Correlation Correlation Quer y Key Figure 4 Illustration of Attention Mask Design: Causal Self-Attention Mask (CSAM) and Multilevel Cross-Attention Mask (MCAM). We present a specific customized task case as an example. Given the concatenated VAE tokensz ā² = [z; z c ]ā R (T+N c )ĆHWĆC , wherez c ā R N c ĆHWĆC represents the condition tokens, we introduce the Relational Rotary Position Embedding (R2PE), as illustrated in Fig. 3. The condition tokens z c are composed of subject tokens z sub , object tokens z obj , and background tokens z bg , i.e.,z c = [z sub ; z obj ; z bg ]. In R2PE, for the video tokens z, we adopt the standard 3D-RoPE (i,j,k) position assignment method, while for the background z bg and object z obj tokens, we sequentially extend each entity along thei-index. For the subject tokens z sub , which are composed of human face tokens z face and human attribute tokens z attr , we strictly adhere to the face-attribute dependency when assigning position indices to the subject tokens. Therefore, for the human face tokens and their corresponding attribute tokens within the same group, they share the samei-index and are extended along thej-index andk-index. Specifically, the position index for the condition tokens z c is defined as: (i ā² ,j ā² ,k ā² ) = ( i bg/obj + T,j,k ,when z bg and z obj i sub + T + N bg/obj ,j + W ā N g i sub ,k + H ā N g i sub , when z sub (1) wherei bg/obj ā[0,N bg/obj ), withN bg/obj denoting the total number of background and object entity.i sub ā [0,N sub ) whereN sub represents the total number of face-attribute subject groups. And,N g i sub ā[0,N i sub ), whereN i sub denotes the total number of face and attribute entity within thei sub subject group. The proposed R2PE effectively inherits and extends the implicit positional correspondence of the original Wan2.1 model, while preserving the face-attribute dependency within each group of the subject condition. Causal Self-Attention Mask (CSAM). The Causal Self-Attention Mask is a boolean matrix, as illustrated in Fig. 4 (a), with the following two rules governing its mechanism: (I) Calculations are performed within each conditional branch, where the human face and its corresponding attributes are treated as a unified subject condition branch; (I) Video denoising tokens apply unidirectional attention to the condition tokens only. Given the concatenated tokens z ā² ā R (T+N c )ĆHWĆC , the mask can be formulated as: M SA q,k = ( True, if qā z or q == k or q, kā z g sub False, otherwise (2) where q and k denote the categories of the tokens corresponding to the query-key matrix in Self-Attention, both of which belong to the visual concatenated tokens z ā² . z and z g sub represent the denoising video tokens and the face/attribute tokens within the same subject group. This causal mask enforces constraints on the range of interactions during the Self-Attention process. This design efficiently prevents unidirectional attention from the conditional branch to the denoising branch, while enabling the denoising branch to independently aggregate conditional signals and efficiently bind the face-attribute dependencies within the conditional branch. To enable efficient computation, we employ the MagiAttention mechanism proposed in [35]. 6 3.3.2 Relational Cross-Attention Multilevel Cross-Attention Mask (MCAM). In the Cross-Attention process of the T2V task, all visual tokens interact with all textual tokens. However, the requirements may differ for customized video generation tasks. Intuitively, all textual tokens are of equal importance for video denoising tokens. However, for visual condition tokens in customized tasks, each has a corresponding textual token, such as: face image ā āman". Therefore, we aim to enhance the interaction of visual condition tokens with the corresponding textual tokens in the cross-attention process to improve the semantic representation of visual tokens. Furthermore, for subject condition tokens, we seek to strengthen the face-attribute dependency within the same subject group in the Cross-Attention process, while reducing the mutual influence between different subject groups. Based on the aforementioned motivation, we propose the Multilevel Cross-Attention Mask (MCAM), as shown in Fig. 4(b). MCAM is a numerical mask in which we have defined three levels of correlation: Strong Correlation (1), Correlation (0), and Weak Correlation (-1). Specifically, Strong Correlation applies to the interaction between the visual condition tokens and their corresponding textual tokens, as well as between visual subject (face & attribute) condition tokens and all textual tokens within the same subject group. Weak Correlation applies to the interaction between visual subject tokens and the textual tokens from different subject groups. And all other cases remain as Correlation. Therefore, this mask M CA q,k can be formulated as: M CA q,k = ļ£± ļ£“ ļ£² ļ£“ ļ£³ 1 (Strong Correlation), if q, k belong to the same semantic entity or subject group ā1 (Weak Correlation), if q, k belong to the different subject group 0 (Correlation),otherwise (3) where q and k denote the categories of the tokens corresponding to the query-key matrix in Cross-Attention, with the query and key representing the visual and textual tokens, respectively, in this context. Subsequently, we inject this constraint mask M CA q,k into the Cross-Attention as follows: Cross-Attention(Q, K, V) = Softmax QK ā¤ + M CA q,k Ā· sĀ· r ā d K ! V(4) where Q denotes concatenated visual features, and K and V are textual features. The hyperparameterr controls the strength of the M CA q,k constraint. Because similarity scores between query and key tokens vary across positions, a uniform mask template cannot be applied directly. To address this, we introduce a dynamic scaling factor s to adjust M CA q,k at each position. The most straightforward strategy is to use the absolute value of the similarity matrix itself as s. However, existing accelerated Attention computation modules based on Pytorch do not support customized numerical masks like this, and recomputing the similarity scores between Q and K outside the Attention module would incur significant computational overhead. To balance accuracy and efficiency, we propose an approximate method to compute the similarity matrix and derive s outside the Attention module as: s = Repeat Q ds K ā¤ , shape QK ā¤ (5) where Q ds denotes Q downsampled by a factor ofdĆ dvia local average pooling on its spatial dimensions. TheRepeat(Ā·, shape(Ā·)) operation restores the downsampled similarity matrix to its original size. Overall, the proposed MACM effectively strengthens relational dependency consistency and enhances the semantic-level representation of visual condition tokens. 4 Experiments 4.1 Experimental settings Datasets. Our personalized multi-subject video generation training dataset is built on Panda70M [4]. After the cleaning and processing steps described in Sec. 3.2, we obtain 1.57M samples: 1.31M single-subject, 0.23M two-subject, and 0.03M three-subject videos. Benchmark. For the testing benchmark, 500 videos crawled from YouTube are processed using the method described in Sec. 3.2, including 220 single-subject, 230 two-subject, and 50 three-subject videos. To rigorously 7 assess personalized multi-subject video generation, we establish two tasks in this benchmark: identity-consistent and subject-consistent video generation. For identity-consistent video generation, 1 to 3 facial reference images are provided, with the facial similarity between the reference images and generated videos assessed using FaceSim-Arc (ArcSim), based on ArcFace [6], and FaceSim-Cur (CurSim), based on CurricularFace [14], following [52] and [48]. We also employ VideoCLIPXL (ViCLIP-T) [42] to measure semantic similarity between generated videos and text prompts. For subject-consistent video generation, the model takes multiple reference images, including faces, attributes, objects, and the background, as input. To comprehensively evaluate this task, we consider two aspects: 1) Evaluation of the entire video, where we use ViCLIP-T [42] and ViCLIP-V [42] to assess semantic similarity between the generated videos and text prompts, as well as between the generated videos and the ground-truth videos, respectively. Additionally, to prevent the generated videos from exhibiting copy-paste artifacts, we assess their dynamic degree, following [16]. 2) Evaluation on subjects, where we first use Florence-2 [47] to detect the person subjects based on the text prompts, and then apply OWLv2 [29] to detect the bounding boxes of the personās attributes and other objects. Once all subjects are located, we compute CLIP-T [32] between the cropped image regions and the corresponding text prompts, as well as DINO-I [30] and CLIP-I [32] between these regions and the corresponding reference images. ArcSim [6] is also used here to assess the identity similarities. If a subject is not detected in the generated video, the score is set to zero. Implementation Details. LumosXis fine-tuned from Wan2.1ās T2V(1.3B) [40] model, built on the DiT [31] architecture. Video generation is performed at 480p resolution, with each training clip containing 81 frames (5 seconds at 16 FPS). In MACM, the downsampling factordand hyperparameterrare set to 8 and 0.5, respectively. During training, each subjectās face is associated with up to three attributes. Therefore, we recommend providing no more than three attributes per subject group during inference to maintain consistency with the training setup. LumosXis trained in two phases: 15k iterations on single-subject data, followed by 16k iterations on mixed multi-subject data. Training uses the Adam [18] optimizer with a learning rate of 1e-5, EMA decay of 0.99, weight decay of 1e-4, gradient clipping at 1.0, a batch size of 64, and random text-conditioning dropout at 10%. Our full training process required approximately 883 GPU-days on H20 GPUs. During inference, we use 50 steps and set the CFG scale to 6. 4.2 Main Results In the following experiments, all baseline configurations are fixed when compared against our LumosX (Wan2.1-1.3B), including ConsisID (CogVideoX-5B) [48], Concat-ID (Wan2.1-1.3B) [52], SkyReels-A2 (Wan2.1-14B) [8], and Phantom (Wan2.1-1.3B) [24]. For the Identity-Consistent Video Generation setting, all methods use the same inputs: each subjectās face image and a shared global text prompt. For the Subject-Consistent Video Generation setting, we also enforce strict input parity. All methods receive exactly the same inputs as LumosX, including each subjectās face image, all associated attribute images, object reference images, the background image, and the shared global prompt. Identity-consistent video generation. In this experiment, we use only face reference images as input to evaluate identity preservation with ArcSim and CurSim. We first compare our LumosXwith face-specific customization methods, ConsisID [48] and Concat-ID [52], on a single-face test set of 220 videos, as ConsisID supports only single-face customization and Concat-ID has only released weights for the single-face setting, with results presented in Tab. 1. Additionally, we compare LumosXwith SkyReels-A2 [8] and Phantom [24], two general multi-subject video customization methods, on the full test set of 500 videos, as shown in Tab. 2. Across both settings, LumosXachieves SOTA identity similarity scores, demonstrating its strong ability to preserve identity consistency in video generation. The qualitative comparison shown in Fig. 5 (a) further highlights the advanced performance of our LumosX. Subject-consistent video generation. This experiment focuses on comprehensive multi-subject video customiza- tion, with reference inputs including faces, attributes, objects, and background images. For the evaluation of the entire video, we adopt ViCLIP-T and ViCLIP-V to assess the semantic similarity, while Dynamics detects potential copy-paste artifacts in the motion consistency. For evaluation on subjects, CLIP-T, CLIP-I, and DINO-I measure the semantic similarity, while ArcSim and CurSim evaluate facial similarity within the extracted subject regions, collectively capturing the accuracy of face-attribute associations in multi-subject 8 Table 1 Comparison of different methods for single-face identity-consistent video generation. Methods Identity Consistency Prompt Following ArcSim ā CurSim āViCLIP-T ā ConsisID [48]0.4580.4740.263 Concat-ID [52]0.4670.4850.261 LumosX0.5420.5750.262 Table 2 Comparison of different methods for identity- consistent video generation. Methods Identity Consistency Prompt Following ArcSim ā CurSim āViCLIP-T ā SkyReels-A2 [8]0.3820.4010.261 Phantom [24]0.5080.5360.264 LumosX0.5100.5400.262 Table 3 Comparison of different methods for subject-consistent video generation, including evaluation of the entire video and evaluation on the extracted subjects. Methods Entire VideoExtracted Subjects Dynamic ā ViCLIP-T ā ViCLIP-Vā CLIP-Tā CLIP-I ā DINO-Iā ArcSimā CurSimā SkyReels-A2 [8]0.6710.2510.8390.1780.6060.1920.2710.290 Phantom [24]0.6610.2540.8650.1850.6470.2160.4440.477 LumosX0.7230.2600.9320.2010.6920.2610.4540.483 video generation. As shown in Tab. 3, our method achieves SOTA performance in both the entire video generation quality and the accuracy of face-attribute associations within the subject regions, outperforming the advanced models SkyReels-A2 [8] and Phantom [24], thus demonstrating the robustness of our LumosX. For qualitative comparison, Fig. 5(b) shows that our method supports flexible multi-subject foreground and background customization while maintaining accurate faceāattribute matching across multiple subjects. In contrast, competing methods exhibit incorrect faceāattribute pairings at a relatively high frequency, such as the mismatches visible in the lower-left SkyReels-A2 and Phantom results and the lower-right SkyReels-A2 result in Fig. 5(b), which further underscores the robustness of our approach. 4.3 Ablation Study In this paper, we focus on enabling flexible foreground-background customization while addressing the depen- dency of face-attribute across multiple subjects. Therefore, we conduct the ablation study of our LumosXon subject-consistent video generation. We conduct experiments under a relatively lightweight setting, where the training dataset consists of 300K video samples while maintaining the same subject distribution as the full dataset,andthevideogenerationresolutionissetto240p. Table 4 Ablation study for the subject-consistent video generation. The value in the MACM parentheses isr. MethodsCLIP-Tā ArcSim ā None0.1840.316 +R2PE0.1780.363 +R2PE + CSAM0.1820.363 +R2PE + CSAM + MCAM (0.1)0.1820.364 +R2PE + CSAM + MCAM (0.5)0.1860.429 +R2PE + CSAM + MCAM (1.0)0.1870.384 We ablate each component in LumosX, with the results presented in Tab. 4. The results show that R2PE (Row 2) significantly im- proves ArcSim, thanks to the binding of faces and attributes during the positional encoding, which helps the model avoid face confusion in generation. But CLIP-T shows a slight decrease, which we believe may be due to the shared T-idx within the subject group, affect- ing the semantic representation of individual entities. Furthermore, our CSAM (Row 3) shows an improvement over CLIP-T, as it effectively blocks the interaction between different conditional signals and allows the denoising branch to independently aggregate these signals. For MCAM (Rows 4ā6), we evaluated performance under different values of the hyperparameterr. The results indicate that MCAM yields substantial improvements in both CLIP-T and ArcSim, as it not only enhances the semantic representation of each visual condition but also effectively optimizes intra-group and inter-group correlations within subject groups. ArcSim achieves its best performance atr= 0.5 (Row 5), while CLIP-T peaks atr= 1.0 (Row 6). Since the improvement in ArcSim is more significant atr= 0.5, and considering that ArcSim better reflects the accuracy of face-attribute affiliation matching, we ultimately choose r = 0.5 in LumosX. 9 SkyReels-A2 Phantom LumosX Reference Images A man in a dark shirt stands in front of a brightly lit "XMAS" sign with speakers and a pink curtain with a decorative pattern in the background, pointing and speaking to the camera. Caption man dark shirt speakers pink curtain with a decorative pattern Two men are seated in a cozy room. The man on the right wears a gray collared shirt, and is speaking. The man on the left wears a blue polo shirt, listening attentively. In the background, there is a bookshelf with books and a small bottle. SkyReels-A2 Phantom LumosX man on the right gray collared shirt man on the left blue polo shirt small bottle cozy room Caption Reference Images ConsisID Concat-ID LumosX Reference Image Caption SkyReels-A2 Phantom LumosX A man with dark hair and a beard, wearing a light brown zip-up sweater over a black shirt, stands in front of a dark, solid-colored wall. To his left, there is a large, industrial-looking spotlight mounted on a black tripod stand. To his right, there is a wooden ukulele with a glossy finish, featuring a sound hole, a bridge, and four tuning pegs. Reference Images A man with a beard, wearing a white t-shirt, and a woman with her hair tied back, wearing a black top, stand together in a modern, minimalistic interior. The man gestures with his hands while speaking, and the woman listens attentively. In the background, there is a white wall, a round mirror with a thin, dark frame, a kitchen counter with cabinets, and a potted plant with broad, green leaves. Caption (a) Identity-consistent video generation (b) Subject-consistent video generation manmanwoman person on the left Three individuals are in a modern kitchen, engaged in a cooking demonstration. The person on the left, wearing a headband, is smiling at the middle person in a black top holding a utensil. The person on the right in a striped shirt is holding food. In front of them are a wooden cutting board, a metal mixing bowl, and a white baking dish. head -band middle person black top metal mixing bowl person on the right striped shirt wooden cutting board Caption SkyReels-A2 Phantom LumosX white baking dish modern kitchen SkyReels-A2 Phantom LumosX Reference Images In a warmly lit room, a woman in a white shirt discusses a book with another woman wearing a striking red long skirt, both seated comfortably across from each other. Caption woman white shirt woman Reference Images red long skirt Figure 5 Qualitative comparison for identity-consistent and subject-consistent video generation. 5 Conclusion This paper proposes LumosX, a novel framework designed for personalized multi-subject video generation that explicitly models face-attribute dependencies. To solve the lack of annotated data tailored for multi- subject generation, we develop a data collection pipeline that supports open-set entities with subject-specific dependencies. Built upon Wan2.1ās T2V model, LumosXintroduces Relational Self-Attention and Relational Cross-Attention, which incorporate position embedding and attention mechanisms to explicitly bind face- attribute pairs into coherent subject groups and optimize both intra-group and inter-group correlations. Extensive experiments validate the effectiveness of LumosXin generating fine-grained and personalized multi-subject videos, achieving state-of-the-art performance across diverse benchmarks. 10 Acknowledgments This work was supported by the State Key Laboratory of Industrial Control Technology, China (Grant No. ICT2024A09). References [1]Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025. [2] Baidu. Paddleocr. https://github.com/PaddlePaddle/PaddleOCR, 2025. [3] Andreas Blattmann, Robin Rombach, Huan Ling, Tim Dockhorn, Seung Wook Kim, Sanja Fidler, and Karsten Kreis. Align your latents: High-resolution video synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 22563ā22575, 2023. [4] Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Ekaterina Deyneka, Hsiang-wei Chao, Byung Eun Jeon, Yuwei Fang, Hsin-Ying Lee, Jian Ren, Ming-Hsuan Yang, et al. Panda-70m: Captioning 70m videos with multiple cross-modality teachers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 13320ā13331, 2024. [5]Tsai-Shien Chen, Aliaksandr Siarohin, Willi Menapace, Yuwei Fang, Kwot Sin Lee, Ivan Skorokhodov, Kfir Aberman, Jun-Yan Zhu, Ming-Hsuan Yang, and Sergey Tulyakov. Multi-subject open-set personalization in video generation. arXiv preprint arXiv:2501.06187, 2025. [6]Jiankang Deng, Jia Guo, Niannan Xue, and Stefanos Zafeiriou. Arcface: Additive angular margin loss for deep face recognition. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 4690ā4699, 2019. [7] Patrick Esser, Sumith Kulal, Andreas Blattmann, Rahim Entezari, Jonas MĆ¼ller, Harry Saini, Yam Levi, Dominik Lorenz, Axel Sauer, Frederic Boesel, et al. Scaling rectified flow transformers for high-resolution image synthesis. In Forty-first international conference on machine learning, 2024. [8]Zhengcong Fei, Debang Li, Di Qiu, Jiahua Wang, Yikun Dou, Rui Wang, Jingtao Xu, Mingyuan Fan, Guibin Chen, Yang Li, et al. Skyreels-a2: Compose anything in video diffusion transformers. arXiv preprint arXiv:2504.02436, 2025. [9]Yuwei Guo, Ceyuan Yang, Anyi Rao, Zhengyang Liang, Yaohui Wang, Yu Qiao, Maneesh Agrawala, Dahua Lin, and Bo Dai. Animatediff: Animate your personalized text-to-image diffusion models without specific tuning. arXiv preprint arXiv:2307.04725, 2023. [10]Xuanhua He, Quande Liu, Shengju Qian, Xin Wang, Tao Hu, Ke Cao, Keyu Yan, and Jie Zhang. Id-animator: Zero-shot identity-preserving human video generation. arXiv preprint arXiv:2404.15275, 2024. [11] Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. Advances in neural information processing systems, 33:6840ā6851, 2020. [12]Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022. [13]Wenyi Hong, Ming Ding, Wendi Zheng, Xinghan Liu, and Jie Tang. Cogvideo: Large-scale pretraining for text-to-video generation via transformers. arXiv preprint arXiv:2205.15868, 2022. [14]Yuge Huang, Yuhan Wang, Ying Tai, Xiaoming Liu, Pengcheng Shen, Shaoxin Li, Jilin Li, and Feiyue Huang. Curricularface: adaptive curriculum learning loss for deep face recognition. In proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 5901ā5910, 2020. 11 [15]Yuzhou Huang, Ziyang Yuan, Quande Liu, Qiulin Wang, Xintao Wang, Ruimao Zhang, Pengfei Wan, Di Zhang, and Kun Gai. Conceptmaster: Multi-concept video customization on diffusion transformer models without test-time tuning. arXiv preprint arXiv:2501.04698, 2025. [16]Ziqi Huang, Yinan He, Jiashuo Yu, Fan Zhang, Chenyang Si, Yuming Jiang, Yuanhan Zhang, Tianxing Wu, Qingyang Jin, Nattapol Chanpaisit, et al. Vbench: Comprehensive benchmark suite for video generative models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21807ā21818, 2024. [17]Yuming Jiang, Tianxing Wu, Shuai Yang, Chenyang Si, Dahua Lin, Yu Qiao, Chen Change Loy, and Ziwei Liu. Videobooth: Diffusion-based video generation with image prompts. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6689ā6700, 2024. [18]Diederik P Kingma. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014. [19]Alexander Kirillov, Eric Mintun, Nikhila Ravi, Hanzi Mao, Chloe Rolland, Laura Gustafson, Tete Xiao, Spencer Whitehead, Alexander C Berg, Wan-Yen Lo, et al. Segment anything. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4015ā4026, 2023. [20] Weijie Kong, Qi Tian, Zijian Zhang, Rox Min, Zuozhuo Dai, Jin Zhou, Jiangfeng Xiong, Xin Li, Bo Wu, Jianwei Zhang, et al. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603, 2024. [21] Black Forest Labs. Flux. https://github.com/black-forest-labs/flux, 2023. [22]Ji Lin, Hongxu Yin, Wei Ping, Pavlo Molchanov, Mohammad Shoeybi, and Song Han. Vila: On pre-training for visual language models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 26689ā26699, 2024. [23] Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022. [24]Lijie Liu, Tianxiang Ma, Bingchuan Li, Zhuowei Chen, Jiawei Liu, Qian He, and Xinglong Wu. Phantom: Subject-consistent video generation via cross-modal alignment. arXiv preprint arXiv:2502.11079, 2025. [25]Shilong Liu, Zhaoyang Zeng, Tianhe Ren, Feng Li, Hao Zhang, Jie Yang, Qing Jiang, Chunyuan Li, Jianwei Yang, Hang Su, et al. Grounding dino: Marrying dino with grounded pre-training for open-set object detection. In European Conference on Computer Vision, pages 38ā55. Springer, 2024. [26]Xiaoran Liu, Hang Yan, Shuo Zhang, Chenxin An, Xipeng Qiu, and Dahua Lin. Scaling laws of rope-based extrapolation. arXiv preprint arXiv:2310.05209, 2023. [27]Ze Ma, Daquan Zhou, Chun-Hsiao Yeh, Xue-She Wang, Xiuyu Li, Huanrui Yang, Zhen Dong, Kurt Keutzer, and Jiashi Feng. Magic-me: Identity-specific video customized diffusion. arXiv preprint arXiv:2402.09368, 2024. [28] Willi Menapace, Aliaksandr Siarohin, Ivan Skorokhodov, Ekaterina Deyneka, Tsai-Shien Chen, Anil Kag, Yuwei Fang, Aleksei Stoliar, Elisa Ricci, Jian Ren, et al. Snap video: Scaled spatiotemporal transformers for text-to-video synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7038ā7048, 2024. [29] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby. Scaling open-vocabulary object detection. Advances in Neural Information Processing Systems, 36:72983ā73007, 2023. [30] Maxime Oquab, TimothĆ©e Darcet, ThĆ©o Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023. [31]William Peebles and Saining Xie. Scalable diffusion models with transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 4195ā4205, 2023. [32] Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from 12 natural language supervision. In International conference on machine learning, pages 8748ā8763. PmLR, 2021. [33]Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bjƶrn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 10684ā10695, 2022. [34]Olaf Ronneberger, Philipp Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In Medical image computing and computer-assisted interventionāMICCAI 2015: 18th international conference, Munich, Germany, October 5-9, 2015, proceedings, part I 18, pages 234ā241. Springer, 2015. [35] Sand-AI. Magi-1: Autoregressive video generation at scale, 2025. [36] Uriel Singer, Adam Polyak, Thomas Hayes, Xi Yin, Jie An, Songyang Zhang, Qiyuan Hu, Harry Yang, Oron Ashual, Oran Gafni, et al. Make-a-video: Text-to-video generation without text-video data. arXiv preprint arXiv:2209.14792, 2022. [37]Jianlin Su, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, and Yunfeng Liu. Roformer: Enhanced transformer with rotary position embedding. Neurocomputing, 568:127063, 2024. [38] Sergey Tulyakov, Ming-Yu Liu, Xiaodong Yang, and Jan Kautz. Mocogan: Decomposing motion and content for video generation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pages 1526ā1535, 2018. [39]Carl Vondrick, Hamed Pirsiavash, and Antonio Torralba. Generating videos with scene dynamics. Advances in neural information processing systems, 29, 2016. [40]Ang Wang, Baole Ai, Bin Wen, Chaojie Mao, Chen-Wei Xie, Di Chen, Feiwu Yu, Haiming Zhao, Jianxiao Yang, Jianyuan Zeng, et al. Wan: Open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314, 2025. [41]Chien-Yao Wang, I-Hau Yeh, and Hong-Yuan Mark Liao. Yolov9: Learning what you want to learn using programmable gradient information. In European conference on computer vision, pages 1ā21. Springer, 2024. [42] Jiapeng Wang, Chengyu Wang, Kunzhe Huang, Jun Huang, and Lianwen Jin. Videoclip-xl: Advancing long description understanding for video clip models. arXiv preprint arXiv:2410.00741, 2024. [43]Zhao Wang, Aoxue Li, Lingting Zhu, Yong Guo, Qi Dou, and Zhenguo Li. Customvideo: Customizing text-to-video generation with multiple subjects. arXiv preprint arXiv:2401.09962, 2024. [44]Yujie Wei, Shiwei Zhang, Zhiwu Qing, Hangjie Yuan, Zhiheng Liu, Yu Liu, Yingya Zhang, Jingren Zhou, and Hongming Shan. Dreamvideo: Composing your dream videos with customized subject and motion. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6537ā6549, 2024. [45]Haoning Wu, Zicheng Zhang, Weixia Zhang, Chaofeng Chen, Liang Liao, Chunyi Li, Yixuan Gao, Annan Wang, Erli Zhang, Wenxiu Sun, et al. Q-align: Teaching lmms for visual scoring via discrete text-defined levels. arXiv preprint arXiv:2312.17090, 2023. [46] Shaojin Wu, Mengqi Huang, Wenxu Wu, Yufeng Cheng, Fei Ding, and Qian He. Less-to-more gener- alization: Unlocking more controllability by in-context generation. arXiv preprint arXiv:2504.02160, 2025. [47]Bin Xiao, Haiping Wu, Weijian Xu, Xiyang Dai, Houdong Hu, Yumao Lu, Michael Zeng, Ce Liu, and Lu Yuan. Florence-2: Advancing a unified representation for a variety of vision tasks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 4818ā4829, 2024. [48] Shenghai Yuan, Jinfa Huang, Xianyi He, Yunyuan Ge, Yujun Shi, Liuhan Chen, Jiebo Luo, and Li Yuan. Identity-preserving text-to-video generation by frequency decomposition. arXiv preprint arXiv:2411.17440, 2024. 13 [49]Yuechen Zhang, Yaoyang Liu, Bin Xia, Bohao Peng, Zexin Yan, Eric Lo, and Jiaya Jia. Magic mirror: Id-preserved video generation in video diffusion transformers. arXiv preprint arXiv:2501.03931, 2025. [50]Yunpeng Zhang, Qiang Wang, Fan Jiang, Yaqi Fan, Mu Xu, and Yonggang Qi. Fantasyid: Face knowledge enhanced id-preserving video generation. arXiv preprint arXiv:2502.13995, 2025. [51]Zangwei Zheng, Xiangyu Peng, Tianji Yang, Chenhui Shen, Shenggui Li, Hongxin Liu, Yukun Zhou, Tianyi Li, and Yang You. Open-sora: Democratizing efficient video production for all. arXiv preprint arXiv:2412.20404, 2024. [52]Yong Zhong, Zhuoyi Yang, Jiayan Teng, Xiaotao Gu, and Chongxuan Li. Concat-id: Towards universal identity-preserving video synthesis. arXiv preprint arXiv:2503.14151, 2025. 14 Appendix of LumosX In this Appendix, we provide additional content organized as follows: ā¢ Sec. A details the data collection pipeline and training dataset, including: ā Sec. A.1 Entity words retrieval and subjectāattributes matching. ā Sec. A.2 Obtaining condition images. ā Sec. A.3 Data cleaning and filtering for training dataset. ā Sec. A.4 Robustness analysis of external modules in the data collection pipeline. ā¢ Sec. B provides additional details on LumosX, including: ā Sec. B.1 Training objective. ā Sec. B.2 Data augmentation details during training. ā Sec. B.3 Generalization of LumosX to diverse T2V architectures. ā¢ Sec. C presents extended ablation studies and evaluations, including: ā Sec. C.1 Fine-grained quantitative comparisons of identity-consistent video generation. ā Sec. C.2 Fine-grained quantitative comparison of subject-consistent video generation. ā Sec. C.3 Additional quantitative results from various online sources. ā Sec. C.4 Discussion on LumosXās capability for customized control for 4+ subjects. ā Sec. C.5 Quantitative assessment of temporal coherence among different methods. ā Sec. C.6 Evaluation on public personalization benchmark. ā Sec. C.7 Visual effectiveness of individual components of LumosX. ā Sec. C.8 Qualitative comparison under varying hyperparameters r in MCAM. ā Sec. C.9 Quantitative analysis of text-based attribute control. ā Sec. C.10 Quantitative comparison with image-personalizationābased initialization. ā Sec. C.11 Discussion of the importance of the inpainting model in data collection pipeline. ā Sec. C.12 Human study for multi-subject video customization. ā Sec. C.13 Analysis of computational overhead and latency in LumosX. ā¢ Sec. D provides more visualization results, including: ā Sec. D.1 Additional results of identity-consistent video generation. ā Sec. D.2 Additional results of subject-consistent video generation. ā¢ Sec. E discusses the limitations and future work. A Details of Data Collection Pipeline and Training Dataset A.1 Entity Words Retrieval and Subject-Attributes Matching In Sec. 3.2 of the manuscript, we propose utilizing Qwen2.5-VL-32B [1] to retrieve multiple entity words from the caption, while leveraging prior visual information from human detection results to achieve precise face-attribute matching. The process of processing a human-detecting frame with the corresponding caption can be divided into four steps, as shown in Fig. 6. The result of Data Organization (step 2) serves as the input to Qwen2.5-VL and includes: captions (text), bounding boxes (text), and frames with colored bounding boxes (visual). During the Data Processing (third step) phase, we design a tailored prompt for Qwen2.5-VL to 15 better address the task requirements, as shown below. The results of Data Processing, as shown in the fourth column (āData Annotationā) of Fig. 6, involve extracting key entity words from the captions and categorizing them into Subjects, Objects, and Background. For Subjects, we obtain human-attribute dependency relations, which strictly follow the matching structure presented in the human-detecting frame. A man and a woman are seated at a table in a lush garden, sharing a meal. The man, wearing a black shirt and a black watch on his left wrist, gestures with his hands while speaking, while the woman, in a white top, listens attentively and occasionally responds. Utensils are set neatly on the tables as they enjoy their meal together... You are an information extraction assistant. Analyze the given video caption, frame (with colored bounding boxes), and bboxes data. Data Collection Caption Data Organization Caption: "A man and a woman are seated at a table in a lush garden, sharing a meal. The man, wearing a black shirt and a black watch on his left wrist..." Bboxes: "bbox1": "coordinate": [0, 0, 365, 480], "color": "periwinkle", "bbox2": "coordinate": [375, 64, 694, 480], "color": "orange" Frame: <image with man in blue bbox1, woman in orange bbox2> Data ProcessingData Annotation Subjects man: black shirt, black watch woman: white top Objects utensils Background lush garden Figure 6 Process of entity words retrieval and subject-attributes matching with four steps: Data Collection, Data Organization, Data Processing, and Data Annotation. Prompt design for entity words retrieval and subject-attribute matching in Qwen2.5-VL PROMPT = """ You are an information extraction assistant. Analyze the given video caption, frame (with colored bounding boxes), and bounding box data. Rules: 1. For PERSONS: - Identify how many persons by the caption and bboxes data, list ALL human-related terms (human/person/man/woman/girl/ boy etc.). Merge the same person to avoid redundancy. - Extract ONLY concrete attributes for each person: [OK] Pick maximum THREE most frequent concrete attributes. [X] The attibutes can be clothing descriptions (e.g., "black suit", "red dress", "glasses", "hat") or physical features (e.g., "blonde hair", "beard") [X] REJECT abstract attributes (e.g., "skin tone", emotions/actions like "smiling" or "running") [X] REJECT additional descriptions of the attributes (e.g., red hoodie with a black design on the front -> red hoodie) [X] REJECT long phrases: keep under 5 words per attribute - Crucially, use the Frame and BBoxes input: Visually match each person described in the Caption to the person inside the corresponding colored bbox (red, green, blue) depicted in the Frame. The BBoxes variable maps color names (e. g., "red") to bbox identifiers (e.g., "bbox1"). Assign the correct bbox identifier to the matched person to prevent mismatches. 2. For OTHER SUBJECTS: [OK] Pick maximum THREE most frequent concrete objects [OK] Use exact object names (e.g., "dog", "oak tree", "wooden chair") [X] REJECT objects that cannot be directly described with specific nouns (e.g., small rectangular object) [X] REJECT words that may refer to multiple objects (e.g., "leaves") [X] REJECT long phrases: keep under 5 words per attribute 3. For BACKGROUND: Nouns indicating the overall environment or backdrop of the video [OK] Extract the exact environmental phrase (e.g., "park", "coffee shop") [X] REJECT inferred locations 4. NOTES: - The extracted phrases must strictly come from the caption, without adding, removing, or rewriting words and punctuation marks. - Ensure bbox references correspond to the correct colored bbox from the input. - Bbox coordinate format: [xmin, ymin, xmax, ymax]. """ 16 GroundingDINO + SAM Face Cropping Subject & Object Masking +Inpainting Random Sampling SAM Subjects man: black shirt, black watch woman: white top Objects utensils Background lush garden man woman black shirt white top Subjects utensils Objects lush garden Background Figure 7 Process of acquiring condition images of subjects, objects, and the background. A.2 Obtaining condition images The detailed process for obtaining condition images, described in Sec. 3.2, is illustrated in Fig. 7. For subjects, we first apply face detection [41] within the bounding boxes obtained from human detection to extract facial crops. Then, guided by the extracted word tags from Qwen2.5-VL [1], we use SAM [19] to segment the corresponding attribute regions within these human bounding boxes. For objects, we first employ GroundingDINO [25], guided by object tags, to detect bounding boxes from the global image. We then apply SAM within these bounding boxes to segment the mask regions corresponding to each object entity. As for the background, we first remove the subjects and objects based on the previously obtained crops and masks in face cropping and SAM. Since SAM occasionally produces imprecise boundaries, we dilate the foreground masks before inpainting. We then utilize the diffusion-based inpainting model FLUX [21], guided by the prompt: ābackground, empty, nothing, there is nothing,ā where background corresponds to the extracted background tag. Finally, based on the valid outputs from the three key frames, we randomly select one result for each entity to serve as its corresponding condition image. This design aligns with the inference process, where each condition typically corresponds to a single reference image. Additionally, the random selection prevents all condition images from originating from the same frame, thereby enhancing data diversity. A.3 Data Cleaning and Filtering for Training Dataset Our personalized multi-subject video generation training dataset is constructed based on Panda70M [4]. However, not all data in the Panda70M dataset meets the quality standards required for video generation training. To ensure the suitability of the dataset, we perform a filtering and cleaning process based on the following criteria and operations: ā¢ Removing subtitles using a text detection pipeline based on PaddleOCR [2]. ā¢ Cropping black-and-white borders using a simple Hough transform-based approach. ā¢ Excluding grayscale or black-and-white videos to ensure sufficient visual quality and aesthetic appeal. ā¢ Retaining only samples with QAlign quality>3.5 [45] and aesthetic score>2.0 [45], enforcing constraints on semantic alignment and visual aesthetics. ā¢ Retaining only samples with flow strength within the range of 0.05 to 2.0, constraining motion intensity to an appropriate level for video generation. 17 ā¢Retaining only videos in which 1 to 3 people are detected by Yolov9 [41], as videos with more than 3 individuals typically exhibit lower visual and semantic quality. ā¢ Removing duplicate videos by clustering based on VideoCLIP [42] embeddings. After data cleaning and filtering, we collect a total of 1.57M video samples, comprising 1.31M single-subject, 0.23M two-subject, and 0.03M three-subject videos. A.4 Robustness Analysis of External Modules in the Data Collection Pipeline To ensure the quality and robustness of the data collection pipeline, we adopt deliberate component choices and strict filtering strategies to reduce upstream prediction errors: ā¢ Careful Selection of High-Reliability Components: 1) For faceāattribute tag extraction, we used Qwen- VL-32B, which incorporates visual priors, instead of the pure language model Qwen-2.5-32B, improving accuracy (see Sec. A); 2) For background inpainting, we selected FLUX over Stable Diffusion 2.0 due to its superior realism. ā¢ Strict Data Filtering Policies: 1) Qwen-VL predictions were accepted only if the predicted entities ap- peared in the captions; 2) Conservative thresholds were used in detection modules (box_threshold=0.5, text_threshold=0.45in Grounding DINO and SAM) to eliminate unreliable region-text matches; 3) For background inpainting, we retained only samples where the foreground occupied less than 50% of the image to maintain consistency and plausibility. Additional filters are detailed in Sec. A.3. While isolating the robustness of each external module is computationally expensive, we conduct a simple test of Qwenās robustness in word tag extraction. A case was marked as incorrect if either the face or attribute word was missing from the caption, or if the pairing was semantically invalid. With visual priors, Qwen-VL-32B achieved 95.2% accuracy, compared to 78.4% for Qwen-2.5-32B without visual input. Overall, our stringent quality-driven filtering is evident in the data selection ratio: from 70 million Panda70M samples, only 1.57 million videos were retained (2.2%). This highlights our strong emphasis on quality control and minimizing error propagation. B Details and Discussions of LumosX B.1 Training Objective Due to our LumosXbuilt upon Wan2.1ās [40] T2V model, we adopt Flow Matching[23] to formulate the training objective. Flow Matching is a generative paradigm that learns continuous-time dynamics using ordinary differential equations (ODEs). It directly estimates optimal transport paths between distributions, enabling stable training without iterative denoising steps. Given a video latent representation z, a random noise z 0 , and a timesteptā[0,1] drawn from a logit-normal distribution, an intermediate latent z t is obtained as the training input. Following Rectified Flows (RF) [7], the intermediate state z t is defined via linear interpolation between z and z 0 , formulated as: z t = (1ā t)Ā· z 0 + tĀ· z,(6) The corresponding ground-truth velocity is given by the time derivative of x t : v t = dz t dt = zā z 0 .(7) The model is trained to estimate this velocity field. The training objective is defined as the Mean Squared Error (MSE) between the predicted velocity and the ground-truth velocity: L = E z 0 ,z,c text ,z c ,t ā„u(z t , c text , z c ,t;Īø)ā v t ā„ 2 ,(8) where c text and z c denote the umT5 text embedding and reference visual embedding, respectively.Īøis the model weights, and u(z t , c text , z c ,t;Īø) denotes the output velocity predicted by the generation model. 18 B.2 Data Augmentation Details During Training During training, all conditional reference images are resized to match the video aspect ratio. Subject and object reference images are augmented with both numerical transformations (e.g., brightness, blur) and geometric transformations (e.g., rotation, horizontal flip), while background images are augmented only with numerical augmentations. B.3 Generalization of LumosX to Diverse T2V Architectures While our current implementation is based on Wan2.1-T2V with a single-tower DiT architecture, the proposed modules in LumosXare architecture-agnostic and compatible with other DiT-style backbones, such as the dual-tower M-DiT in HunyuanVideo [20] and the Parallel Attention in MAGI-1 [35]. Specifically, R2PE simply reorders relative position indices and can be seamlessly integrated into any DiT-based model, as it is independent of tower design. CSAM and MCAM function as attention masks for Self-Attention (among visual tokens) and Cross-Attention (between visual and textual tokens), respectively. In M-DiT, these operations occur after dual-stream fusion and can adopt the same masking logic. Similarly, Parallel Attention includes equivalent attention interfaces that support our mask-based modules directly. In essence, regardless of whether the base model is Wan2.1, HunyuanVideo, or MAGI-1, all DiT-based T2V architectures share two key components: intra-modal interactions among visual tokens and cross-modal attention between visual and textual tokens. Our modules are explicitly designed to operate at these points, enabling straightforward extension to diverse architectures. C More Ablation Discussion C.1 Fine-Grained Quantitative Comparisons of Identity-Consistent Video Generation We have added comparisons with Phantom and SkyReels-A2 in the single-face and multi-faces identity- consistent video generation setting, as shown in Tab 5. In the single-face setting (rows 1-3), our method underperforms Phantom in terms of identity preservation metrics (ArcSim and CurSim). However, our LumosX demonstrates a clear performance advantage in multi-face scenarios for identity preservation (rows 4ā6). Table 5 Fine-grained quantitative comparison of identity-consistent video generation across different numbers of faces. MethodsNumbers Identity Consistency Prompt Following ArcSim ā CurSim āViCLIP-T ā SkyReels-A2 [8] 1 face0.5090.5400.258 Phantom [24]1 face0.6020.6370.261 LumosX1 face0.5420.5750.262 SkyReels-A2 [8] ā„ 2 faces0.2820.2920.263 Phantom [24] ā„ 2 faces0.4340.4570.266 LumosXā„ 2 faces0.4850.5130.263 C.2Fine-Grained Quantita- tive Comparison of Subject- Consistent Video Generation To further demonstrate the supe- riority of our model in handling multi-subject settings, we conduct a more fine-grained quantitative evaluation of subject-consistent video generation across varying numbers of subjects (1-3 subjects), as shown in Tab. 6. It can be ob- served that as the number of sub- jects increases, LumosXdemon- strates increasingly superior performance across the majority of metrics, with the performance gap over other methods becoming more pronounced. These experimental results further validate the effectiveness of the module we designed in LumosX for modeling multi-subject face-attribute dependency relationships. C.3 Additional Quantitative Results from Various Online Sources We also provide additional qualitative results using condition images collected from various online sources. The test includes 50 sample cases (20 single-subject, 20 two-subject, and 10 three-subject cases). Since ground truth videos are not available for these samples, ViCLIP-V cannot be measured. As shown in Tab. 7, our 19 Table 6 Fine-grained quantitative comparison of subject-consistent video generation across different numbers of subjects. (Ā·) denotes that the respective method yields a lower value than our LumosXon this metric, whereas (Ā·) indicates better performance. Values in parentheses are recalibrated differences computed as (Methodā LumosX ) for the same setting and metric. MethodsNumbers Entire VideoExtracted Subjects Dynamic ā ViCLIP-T ā ViCLIP-VāCLIP-TāCLIP-I āDINO-IāArcSimāCurSimā SkyReels-A2 1 subject 0.772 (0.080) 0.252 (0.007) 0.870 (0.064) 0.177 (0.022) 0.604 (0.082) 0.199 (0.069) 0.356 (0.133) 0.380 (0.136) Phantom1 subject 0.848 (0.004) 0.254 (0.005) 0.880 (0.054) 0.186 (0.013) 0.646 (0.040) 0.215 (0.053) 0.539 (0.050) 0.581 (0.065) LumosX1 subject0.8520.2590.9340.1990.6860.2680.4890.516 SkyReels-A2 2 subjects 0.655 (0.009) 0.251 (0.011) 0.820 (0.110) 0.182 (0.021) 0.620 (0.079) 0.185 (0.064) 0.216 (0.227) 0.233 (0.241) Phantom2 subjects 0.526 (0.138) 0.255 (0.007) 0.856 (0.074) 0.188 (0.015) 0.658 (0.041) 0.215 (0.034) 0.396 (0.047) 0.424 (0.050) LumosX2 subjects0.6640.2620.9300.2030.6990.2490.4430.474 SkyReels-A2 3 subjects 0.268 (0.156) 0.243 (0.017) 0.795 (0.141) 0.161 (0.039) 0.546 (0.144) 0.189 (0.097) 0.150 (0.202) 0.161 (0.219) Phantom3 subjects 0.411 (0.013) 0.250 (0.010) 0.840 (0.096) 0.169 (0.031) 0.601 (0.089) 0.220 (0.066) 0.247 (0.105) 0.269 (0.111) LumosX3 subjects0.4240.2600.9360.2000.6900.2860.3520.380 SkyReels-A2 4 subjects 0.257 (0.047) 0.241 (0.022) 0.761 (0.165) 0.111 (0.095) 0.375 (0.329) 0.116 (0.149) 0.122 (0.167) 0.129 (0.190) Phantom4 subjects 0.286 (0.018) 0.240 (0.023) 0.790 (0.136) 0.165 (0.041) 0.615 (0.089) 0.200 (0.065) 0.191 (0.098) 0.200 (0.119) LumosX4 subjects0.3040.2630.9260.2060.7040.2650.2890.319 Table 7 Comparison of different methods for subject-consistent video generation, including evaluation of the entire video and evaluation on subjects. Methods Entire VideoExtracted Subjects Dynamic ā ViCLIP-T ā CLIP-Tā CLIP-I ā DINO-Iā ArcSimā CurSimā SkyReels-A2 [8]0.8390.2350.2070.6140.1730.1600.166 Phantom [24]0.9290.2550.2260.6230.1950.2710.305 LumosX0.8040.2470.2320.6680.3360.3170.346 Caption "10 people looking directly at the camera with fully visible frontal faces. They toss and catch glowing objects between each other in a high-tech arena. Fast coordinated motion, but faces always unobstructed and well-lit." Wan2.1-T2v-1.3B Figure 8 Visualization of 10-person T2V results of our baseline model Wan2.1-T2V-1.3B. method continues to outperform other approaches on most metrics, particularly those related to subject-specific customization (CLIP-T, CLIP-I, DINO-I, ArcSim, CurSim). C.4 Discussion on LumosX ās Capability for Customized Control for 4+ Subjects LumosXis designed with scalability in mind. Although the training dataset contains videos with up to three subjects, the model architecture, including R2PE, CSAM, and MCAM, is inherently scalable and capable of handling more subjects during inference without retraining. To validate this capability, we conduct quantitative comparisons under a four-subject setting using 50 videos, without any retraining. The results in Tab. 6 show that even in the four-subject setting, our method consistently outperforms the baselines. Moreover, performance remains stable compared to the three-subject setting, showing no significant drop. It is worth noting that although R2PE in LumosXinherits the extrapolation capability of RoPE, increasing the number of subjects slightly beyond what was seen during training (e.g., from 3 to 4) does not lead to a noticeable drop in performance during inference without re-training. However, when the subject count increases substantially (e.g. from 3 to 10+), the risk of extrapolation instability rises significantly: higher RoPE dimensions may not have seen a full rotation period during training [26], causing positional encodings to become Out-of-Distribution (O.O.D), which may degrade attention alignment. To support long-range inference without retraining, we plan to try to integrate NTK-Aware Scaled RoPE (NTK-RoPE) [26], a training-free length extrapolation method that extends RoPEās context range by adjusting the base of its sinusoidal encoding rather than learning new positional parameters. This change enables the model to handle positions beyond the training window with 20 Table 8 Quantitative comparison of temporal coherence metrics on subject-consistent video generation. MethodsSubject Consistency ā Background Consistency ā Motion Smoothness ā Face Consistency ā SkyReels-A2 [8]0.6540.7980.9790.739 Phantom [24]0.7680.8530.9860.883 LumosX0.9620.9460.9880.895 minimal perplexity degradation and no fine-tuning required. In addition to RoPE extrapolation, we also find that the generation capability of the underlying base model itself plays a crucial role in determining the upper limit of how many subjects the customization model can realistically handle. For example, Wan2.1-1.3B-T2V, which we rely on, also struggles significantly when generating videos containing more than ten distinct human subjects. Fig. 8 presents visualizations of Wan2.1-1.3B-T2V under a 10-subject setting, where the facial quality is severely degraded and the model fails to strictly follow the prompt, generating only nine people instead of ten. This reinforces our point that the base modelās capability is the primary bottleneck when extending LumosXto scenarios involving substantially more subjects. These results indicate that current video-generation models like Wan-2.1 lack the capacity to learn stable representations for such large numbers of subjects, largely because high-quality 10+ subject videos are extremely rare. C.5 Quantitative Assessment of Temporal Coherence Among Different Methods To evaluate temporal coherence, we additionally report four metrics: Subject Consistency, Background Con- sistency, Motion Smoothness, and Face Consistency. The first three are adopted from VBench [16]. Subject Consistency measures how consistently the subject (e.g., person or object) appears across consecutive frames. Background Consistency evaluates the temporal stability of the background, while Motion Smoothness quanti- fies the fluidity and natural continuity of motion. In contrast, Face Consistency is a metric we define to capture frame-to-frame facial similarity. Specifically, we use ArcFace embeddings to compute the similarity of the same face across consecutive frames. The results are shown in Tab. 8, evaluated on the test set for subject-consistent video generation, which contains 500 videos. Our LumosXoutperforms other methods across all metrics. Table 9 Quantitative comparison on MSRVTT-personalization [5] benchmark. boldfacen and underlinefont indicate the highest and the second highest results. Method#Base ModelText-S ā Vid-Sā Subj-Sā Dync-Dā SkyReels-A2 [8] Wan2.1-14B-T2V0.2530.7810.5540.783 Phantom [24]Wan2.1-1.3B-T2V0.2700.6960.5340.461 LumosXWan2.1-1.3B-T2V 0.2580.7070.5490.786 C.6Evaluation on Pub- licPersonalizationBench- mark To further evaluate our method, we conducted a quantitative comparison be- tween our LumosXand other approaches, including SkyReels-A2 and Phantom, on the MSRVTT-personalization [5] benchmark. The MSRVTT-personalization benchmark provides two evaluation modes: subject-mode and face-mode. Face-mode focuses solely on face- similarity metrics, whereas we are interested in assessing the overall subject performance (face + attributes). Therefore, we conduct our evaluation under the subject-mode setting. Our evaluation is conducted using one reference image for the subject and one for the background, and, the results are reported in Tab. 9. It is worth noting that MSRVTT-personalization is a single-subject benchmark, and its definition of āsubjectā differs from ours: a subject may refer to either a human category (e.g., man, woman) or an object category (e.g., car, horse, clothes, hat). Moreover, for human subjects under the benchmarkās subject-mode setting, the subject is provided as a holistic entity without decoupled face and attribute references, and therefore does not involve the faceāattribute matching problem that our method is designed to address. As a result, our method does not benefit from its explicit faceāattribute relational modeling in this setting. In addition, both LumosX and Phantom are built on the Wan2.1-1.3B-T2V base model, whereas SkyReels-A2 adopts the much stronger Wan2.1-14B-T2V base model and therefore further benefits from a significantly more capable backbone. Despite these disadvantages, our LumosXstill achieves the second-best overall performance, demonstrating the robustness and generalization ability of our approach even outside its primary multi-subject setting. 21 old man Reference Images Caption In a sunny park, an old man in a white shirt enjoys a leisurely walk beside a woman dressed in a vibrant yellow long skirt, both sharing light-hearted conversations. None R2PE R2PE+CSAM white shirt woman yellow long skirt R2PE+CSAM +MCAM(0.5) Figure 9 Visual effectiveness of individual components proposed in our LumosX. C.7 Visual Effectiveness of Individual Components of LumosX To assess the impact of each proposed component in LumosX, we conduct a qualitative evaluation, as illustrated in Fig. 9. Without incorporating our proposed modules (first row), the generated video exhibits noticeable identity confusion. Specifically, an old white man is misrepresented as black, and a young black girl is erroneously depicted as old. With the introduction of our designed R2PE (second row), coarse-grained identity attributes, including age and skin tone, are effectively rectified. While the introduction of the CSAM (third row) further improves fine-grained facial identity preservation, certain artifacts still persist in the generated videos. Ultimately, the incorporation of the MCAM (fourth row) leads to improved identity consistency in the generated videos, while also enhancing video quality to a certain degree. To further validate the effectiveness of our CSAM and MCAM, we also visualize the Attention Similarity Scores averaged over all 12 heads in the last DiT layer for the case shown in Fig. 9. The results are presented in Fig. 10. It is worth noting that our generated videos contain 81 frames, and after VAE encoding, the temporal dimension is downsampled by a factor of four, resulting in a video noise token length of 21. The results in this figure can be compared with those in Fig. 4 in the main paper. From this comparison, we observe that CSAM (Fig. 10(a) vs. Fig. 10(b)) enables the denoising branch to independently aggregate conditional signals and to effectively bind faceāattribute dependencies within the conditional branch. Moreover, MCAM (Fig. 10(c) vs. Fig. 10(d)) explicitly enhances intra-group and inter-group correlations among subject groups and strengthens the semantic-level representation of the visual condition tokens. C.8 Qualitative Comparison under Varying Hyperparametersr in MCAM When MCAM withr= 0.1 is introduced (third row), the artifacts in the generated videos are significantly alleviated. Nevertheless, identity consistency is still unsatisfactory; for instance, a short-sleeved T-shirt may appear as long-sleeved in the video, and facial similarity is still insufficient. Atr= 0.5 (fourth row), we observe a notable improvement in identity consistency, especially in the resemblance between facial appearances in the generated videos and the reference image, along with a further enhancement in video quality. Increasingr to 1.0 (fifth row) results in continued improvements in video quality, though it introduces a slight decline in identity similarity. The above findings align with the quantitative results reported in Sec. 4.3 of the manuscript. Accordingly, we chooser= 0.5 as the final setting based on a balanced trade-off between identity consistency and video quality. 22 (i) Self-Attention Similarity Score(i) Cross-Attention Similarity Score Visual Visual Texual Visual (a) w/o CSAM (b) w/ CSAM (c) w/o MCAM (d) w/ MCAM Visual Visual Texual Visual Visual 0-20: Video Tokens 2122 23 24 Textual 7-8: old man 11-12: white shirt 22: woman 28-30: yellow long skirt Figure 10 Visualization of attention maps. (i): Self-Attention similarity score. (i): Cross-Attention similarity score. man Reference Images Caption A man in a yellow shirt and a woman in a black shirt sit across from each other at a coffee shop, engaged in a lively conversation. yellow shirt woman black shirt None R2PE+CSAM +MCAM(0.1) R2PE+CSAM +MCAM(1.0) R2PE+CSAM +MCAM(0.5) R2PE+CSAM Figure 11 Qualitative comparison under varying control hyperparametersr in MCAM. C.9Quantitative Analysis of Text-Based Attribute Control Without Images in Multi-Subject Video Customization Our method supports attribute control using text alone, and this setting is already included in both the main paper and the appendix under Identity-Consistent Video Generation (Sec. 4.1, main paper). To more comprehensively evaluate this setting in multi-subject video customization scenarios, we perform subject-region 23 quantitative evaluations under this setting, using the same metrics as in Subject-Consistent Video Generation, to compare text-based attribute control across different methods. The results are presented in the Tab. 11. Overall, the CLIP-T scores are higher under text-only control, which is expected since the attributes are directly specified via text prompts, naturally yielding higher textāvideo similarity. CLIP-I trends largely mirror those of CLIP-T because both measure semantic similarity. Notably, for LumosX(rows 6-7, col. 4), the performance gap between text-only control and text&visual control is very small, whereas for the other methods (rows 2-5, col. 4), performance under text&visual control drops noticeably compared to text-only control. This suggests that those methods become more prone to attribute confusion once visual conditions are introduced, likely due to the absence of explicit faceāattribute relational modeling, whereas LumosX avoids this issue. For DINO-I, text-only control is clearly weaker than text &visual control in LumosX. At the same time, the substantial improvement under text&visual control indicates that our faceāattribute relational constraints are accurately matched and effectively utilized. In comparison, Phantom and SkyReels-A2 do not show notable DINO-I improvements when visual attribute conditions are added, suggesting that their faceāattribute alignment is less reliable. Meanwhile, under the text-based attribute control setting, LumosX performs on par with Phantom and clearly outperforms SkyReels-A2, demonstrating its strong generalization ability even without attribute images. C.10 Quantitative Comparison with Image-PersonalizationāBased Initialization To evaluate whether multi-subject image personalization models can serve as an initial image generator prior to video synthesis, we conducted an additional experiment using UNO [46] to produce a multi-subject customized image, which was then fed into Wan2.1-14B-I2V [40] for video generation. We quantitatively compared this pipeline (UNO + Wan-I2V) against our method on our benchmark. The results are shown in Tab. 12. Overall, our method outperforms UNO + Wan2.1-I2V-14B, particularly on ViCLIP-V, DINO-I, FaceSim and CurSim, which reflects facial identity consistency. This improvement is expected, as UNO is not specifically designed for face-aware attribute binding, whereas our approach explicitly models faceāattribute relational constraints, leading to significantly better identity preservation in video personalization. Our ViCLIP-T and CLIP-T scores are slightly weaker than those of UNO + Wan2.1-I2V-14B, which may be attributed to the latterās use of a substantially larger base model (Wan-14B). C.11 Discussion of the Importance of the Inpainting Model in Data Collection Pipeline To quantitatively assess the realism of FLUX compared with other inpainting models, we conducted a large-scale evaluation on 2,130 test cases, applying both FLUX [21] and Stable-Diffusion-2 [33] for inpainting. We then computed FID scores against the COCO 2017 Val set to measure distributional realism. In addition, we randomly sampled 100 cases and asked GPT-4o to independently judge which inpainting result appeared more realistic. The instruction provided to GPT-4o is as follows: Instruction design for inpainting model assessment in GPT-4o PROMPT = """ - I will give you three images. The first two images are background inpainting results obtained by masking out the foreground in the original image and then applying two different inpainting models. The third image is the corresponding foreground mask. Please evaluate which of the first two images provides a better inpainting result based on pixel-level consistency, scene continuity, and the overall plausibility of the inpainted background. You should output only "1" or "2", indicating whether the first or the second image is better. Do not output anything else. - Note that this is background completion, so the masked region should be inpainted with appropriate background content . Pay particular attention to the continuity and coherence of the background in the inpainted region. """ The results of both evaluations are summarized in the Tab. 10. The results show that FLUX achieves better performance under both evaluation metrics. To further understand how background quality affects downstream video generation, we performed qualitative comparison using the two inpainting outputs as inputs to video generation. As shown in Fig. 12, artifacts introduced during background inpainting are clearly propagated into the generated video, degrading overall video quality. This confirms that inpainting realism plays a 24 Reference Images Caption A woman with red lipstick and a leopard-print top, seated in a modern office setting, makes a peace sign, and holds up a white box. woman leopard print top modern office setting Stable-Diffusion-2-InpaintingFLUX-InpaintingBackground Image with Mask Stable-Diffusion-2 -Inpainting FLUX-Inpainting Figure 12 Qualitative comparison between Stable-Diffusion-2 and FLUX for background inpainting. crucial role in high-quality video generation and motivates our choice of FLUX as the inpainting module. MethodFID ā AI Judgement ā Stable-Diffusion-2-Inpainting 96.3236% FLUX-Inpainting92.8364% Table 10 Comparison between Stable-Diffusion-2 [33] and FLUX [21] for background inpainting. C.12Human Study for Multi-subject Video Customization To complement automated metrics, we con- ducted a user study on a randomly sampled set of 24 video cases, including 6 single- subject, 12 two-subject, and 6 three-subject customization scenarios. The comparison includes LumosX, SkyReels-A2, and Phan- tom, with a total of 30 participants. Each participant evaluated videos along four dimensions: ā¢ FaceāAttribute Alignment: whether each face is correctly matched with its corresponding attributes (e.g., clothing, accessories, or hairstyle). ā¢ Face Similarity: how closely each generated face resembles the provided visual reference. ā¢ Video Naturalness: the overall visual quality and coherence of the generated video. ā¢ Prompt Adherence: whether the generated video follows the instructions specified in the text prompt. For each case, participants ranked the three methods, and scores were assigned as follows: 1 point for first place, 0.5 points for second place, and 0 points for third place. Final scores for each dimension were computed as the weighted average across all participants. The results are shown below Fig 13. Overall, our method achieves superior performance across all four evaluation dimensions. C.13 Analysis of Computational Overhead and Latency in LumosX To evaluate the computational impact of our relational attention mechanisms, we report detailed inference- stage statistics on computational overhead and latency. Specifically, we present the average latency for each Self-Attention and Cross-Attention operation, the average per-step latency, per-step FLOPs, and GPU 25 (a)(b) Figure 13 Human study results for multi-subject video customization. (a) Results across the four evaluation dimensions. (b) Average scores computed over all four dimensions. Table 11 Quantitative comparison of text-only vs. text&visual attribute control across methods. MethodsAttribute Control CLIP-T CLIP-I DINO-I SkyReels-A2 [8] text-only0.1920.6430.192 text&visual0.1780.6060.192 Phantom [24] text-only0.2070.6870.209 text&visual0.1850.6470.216 LumosX text-only0.2050.6840.210 text&visual0.1930.6810.265 Table 12 Comparison of image-personalizationābased method and LumosXfor subject-consistent video generation, including evaluation of the entire video and evaluation on subjects. Methods Entire VideoExtracted Subjects Dynamic ā ViCLIP-T ā ViCLIP-Vā CLIP-Tā CLIP-I ā DINO-Iā ArcSimā CurSimā UNO + Wan2.1-I2V-14B0.5470.2610.8800.2010.6500.1970.2370.244 LumosX0.7230.2600.9320.2010.6920.2610.4540.483 memory consumption. All measurements are obtained on an H20 GPU under the same video customization scenario. The results are shown in Tab. 13 and the key observations are summarized below: ā¢ R2PE incurs no extra compute or memory overhead (row 1 vs. row 2), since it reorders relative position indices rather than introducing new parameters or operations. ā¢The original implementation of the Wan2.1-T2V model utilizes FlashAttention 2.0 for acceleration, which does not support custom masks. In our CSAM module, we replace it with MagiAttention [35] (see Sec. 3.3.1 in the main paper), which supports custom masking and achieves faster inference (row 1 vs. row 3). The integration of CSAM results in only a modest overhead (row 4 vs. row 5). ā¢For MCAM (Cross-Attention), MagiAttention cannot handle numeric masks, so we use PyTorchās native implementation. Since the key comes from relatively short T5 token sequences (512 tokens), the computation remains lightweight. As a result, MCAM does not noticeably affect Cross-Attention latency (row 5 vs. row 6, col 2), and the additional cost of computing the dynamic scaling matrix s is relatively small and well within acceptable bounds (row 5 vs. row 6, cols 3ā5). In summary, through the use of efficient modules (MagiAttention) and optimized strategies (e.g., lightweight scaling matrix design), LumosXachieves high computational efficiency with minimal additional inference overhead (row 6 vs. rows 1 and 4). 26 Table 13 Inference-stage computational overhead and latency statistics for each module under the same video customization case on an H20 GPU. MethodsSelf-Attn Latency Cross-Attn Latency Latency/step FLOPs/step GPU Memory Usage None0.1440s0.0026s8.66s195.44 T21.5G +R2PE0.1441s0.0025s8.66s195.44 T21.5G None (MagiAttention in Self-Attn)0.0935s0.0025s5.79s195.44 T21.5G +R2PE (MagiAttention in Self-Attn)0.0936s0.0026s5.79s195.44 T21.5G +R2PE+CSAM0.0966s0.0026s5.81s195.44 T21.5G +R2PE+CSAM+MCAM0.0965s0.0045s6.11s195.46 T22.7G D More Visualization Results D.1 Additional Results of Identity-Consistent Video Generation The qualitative identity-consistent video generation results are shown in Fig. 14. Under the single-subject setting (Case 1 and Case 2), our method demonstrates significantly superior identity preservation performance compared to ConsisID [48] and Concat-ID [52]. Under the multi-subject setting (Case 3 and Case 4), our approach demonstrates superior performance over SkyReels-A2 [8] and Phantom [24] in both identity preservation and the alignment between multiple subjects and captions, effectively preventing character confusion (Case 3 SkyReels-A2) and positional confusion (Case 3 Phantom). D.2 Additional Results of Subject-Consistent Video Generation The qualitative comparison of subject-consistent video generation is shown in Fig. 15 and Fig. 16. The experimental results demonstrate that our model supports flexible multi-subject foreground-background video customization. Compared to SkyReels-A2 [8] and Phantom [24], our approach achieves superior subject consistency, accurately matching the human faces and their corresponding attributes, and maintaining the reference identity throughout the generation process. In contrast, SkyReels-A2 struggles when handling multiple customized subjects, often exhibiting character confusion (Cases 2, 4 in Fig. 15 and Case 1 in Fig. 16) and subject disappearance (Case 3 in Fig. 15). Similarly, Phantom also suffers from character confusion (Cases 1,2 in Fig. 16) in comparable scenarios. Furthermore, the videos generated by our LumosXare natural and realistic. Unlike Phantom, which suffers from noticeable quality degradation as the number of reference condition images increases, resulting in visual artifacts (Case 1 in Fig. 15 and Cases 2, 3, 4 in Fig. 16) or an unintended cartoon-like style (Cases 2, 3, 4 in Fig. 15). E Limitations and Future Work Although our LumosXeffectively addresses the key challenge of multi-subject video personalization by explicitly modeling the dependency of faceāattribute within the subject group, it remains constrained by limitations in model size as well as the diversity and scale of training data. Consequently, the performance of LumosXhas not yet reached its full potential. Looking ahead, we plan to deploy LumosXon Wan2.1-14B-T2V model [40] and train it on a larger-scale, higher-quality, and more diverse dataset to further enhance its performance and generalization capabilities. Moreover, to push the boundaries of dynamic behavior understanding and better capture complex motion patterns, we also recognize the value of incorporating motion-aware constraintsāfor instance, augmenting data collection with motion descriptions (e.g., walking, running) and integrating motion cues within the MCAM module to strengthen correlations between visual tokens and motion-aware textual tokens, which would improve alignment for dynamic behaviors and multi-subject interactions (e.g., hugging, handshaking, passing objects). 27 ConsisID Concat-ID LumosX Reference Image Caption A man with a beard, wearing a brown cap and a light brown shirt, stands in front of a white shed with a red roof, speaking and occasionally pointing to his right. The shed is adorned with a yellow wheelbarrow, a red pot, and a green hose coiled on the ground. To the right of the shed, a wooden fence with a red ball attached to it is visible, along with a tree featuring a thick trunk and green leaves. The ground is covered with brown mulch, and the scene is set in a sunny outdoor environment. man man on the left Reference Images Caption Two men sit side by side on a light gray sofa, engaged in a conversation. The man on the left, wearing a dark blue shirt, smiles and looks directly at the camera, while the man on the right, in a light blue shirt, occasionally looks at his companion with a subtle smile. Behind them, a brick wall serves as the backdrop, with a dark brown wooden shelf mounted on it. On the shelf, two white candles are placed side by side. man on the right SkyReels-A2 Phantom LumosX ConsisID Concat-ID LumosX Reference Image Caption A woman with shoulder-length wavy hair, wearing a white off-the-shoulder top and a black choker necklace with small, round, brown beads, is seen speaking and gesturing with her hands. She has a neutral expression and is wearing makeup, including eyeshadow and lipstick. Behind her, there is a black metal shelf with three tiers. On the top shelf, there is a small wooden box with a heart-shaped cutout on the lid. woman Reference Images Caption Three people are standing in a kitchen in front of a stainless steel refrigerator. On the left, a blonde woman in a light blue top looks slightly to the side. In the center, a young woman with dark hair tied in a top knot, wearing a light gray sweatshirt, holds a piece of bread. On the right, a man in a dark long-sleeve shirt smiles while looking at her. The refrigerator behind them has magnets and papers, with white cabinets above it. SkyReels-A2 Phantom LumosX woman young woman man Figure 14 Qualitative comparison for identity-consistent video generation. 28 young man Reference Images Caption A young man with short brown hair, wearing a light gray hoodie and a black watch, drives a sleek black convertible through a scenic area with trees and greenery. The side mirror is rectangular with a curved edge. SkyReels-A2 Phantom LumosX light gray hoodie black watch scenic area with trees and greenery man on the left Reference Images Caption Two men are seated at a round wooden table, engaged in a casual conversation. The man on the left, wearing a green plaid shirt, is laughing and clapping his hands, while the man on the right, dressed in a dark green shirt, holds a white mug. A black smartphone is near the man on the left. In the background, a film reel is placed on a shelf. The room is warmly lit, creating a relaxed and friendly atmosphere. SkyReels-A2 Phantom LumosX green plaid shirt room man Reference Images Caption A man in a white shirt and denim jacket converses with a woman in a light brown sweater on a bed in a cozy, warmly lit room. A wooden bedside table with a white lamp is in the background. SkyReels-A2 Phantom LumosX denim jacket cozy, warmly lit room man on the right dark green shirt film reel white mug woman light brown sweater blad man Reference Images Caption A bald man in a black jacket and gray shirt speaks to a woman with short blonde hair in a dark dress and a man in a dark suit in a television studio. The small gray table next to them holds a black cup and two red dice. SkyReels-A2 Phantom LumosX black jacket television studio woman dark dress gray shirt man dark suit black cup red dice Figure 15 Qualitative comparison for subject-consistent video generation. 29 3 woman Reference Images Caption A man wearing a vibrant red hat and a woman adorned in a flowing yellow skirt are leisurely strolling under the trees by seaside. The gentle sea breeze tugs playfully at their attire, adding a dynamic flutter to the woman's skirt and causing the man's hat to tilt charmingly. Their serene silhouettes against the backdrop of the tranquil blue ocean paint a picturesque scene of peaceful leisure. SkyReels-A2 Phantom LumosX yellow skirt under the trees by seaside old man Reference Images Caption At the entrance of the zoo, a young girl in a bright pink tulle dress stands beside an old man dressed in a denim shirt, khaki belt, and a worn cowboy hat. They both face a calm capybara nearby, creating a warm, tranquil moment. The lush greenery and rustic structures in the background suggest a peaceful, nature-filled setting, emphasizing the gentle connection between the three figures. SkyReels-A2 Phantom LumosX cowboy hat man red hat man Reference Images Caption A man in a white shirt and a woman wearing a red long skirt walk side by side through a bustling outdoor market, discussing the items on sale. SkyReels-A2 Phantom LumosX white shirt woman red long skirt man Reference Images Caption Two men stand at a bustling city crosswalk. The first man wears a crisp white jacket, while the second man contrasts sharply in a sleek black jacket, both deep in conversation. SkyReels-A2 Phantom LumosX white jacket man black jacket girl capbara Figure 16 Qualitative comparison for subject-consistent video generation. 30