MirrorWorld: Taming Video Diffusion Models for Mirror Reflection Generation

MirrorWorld: Taming Video Diffusion Models for Mirror Reflection Generation Youjun Zhao1 , Alex Warren2 , Gary K.L. Tam2 , Rynson W.H. Lau1 1

City University of Hong Kong, 2 Swansea University

https://youjunzhao.github.io/MirrorWorld/

arXiv:2608.07463v1 [cs.CV] 7 Aug 2026

Abstract Recent advances in video diffusion models (VDMs) have enabled high-fidelity video synthesis. However, generating mirror reflections remains challenging because the content within a mirror must remain consistent with the surrounding scene. Existing VDMs are not specifically designed to model sceneto-mirror relationships, which can lead to reflections with incorrect content or inconsistent spatial arrangements. We observe that mirror reflection generation involves two complementary challenges: determining what scene content should be reflected and how the reflected content should be spatially arranged within the mirror region. Motivated by this observation, we propose MirrorWorld, a reflection-aware video inpainting framework that models scene-to-mirror relationships during generation. Specifically, we introduce Semantic Relation Distillation (SRD), which transfers relational information from a frozen visual foundation model to encourage semantic associations between visible scene content and mirror regions. We further propose Geometric Transformation Alignment (GTA), which learns a transformation that guides the spatial arrangement of reflected content. The two components play complementary roles, with SRD modeling what should be reflected and GTA modeling how it should be arranged. To facilitate research on this problem, we construct a benchmark for video mirror reflection generation by repurposing four existing video mirror datasets into a unified reflection reconstruction task. Experimental results show that MirrorWorld achieves improved reflection reconstruction quality over representative image-based reflection generation methods and strong video inpainting baselines.

Introduction Recent advances in video diffusion models (VDMs) (Yang et al. 2025b; Wan et al. 2025; Kong et al. 2024; HaCohen et al. 2024) have enabled high-fidelity video synthesis with realistic objects, complex motion, and smooth camera movement. These capabilities support applications such as movie creation (Xiao et al. 2026), synthetic data generation (Wen et al. 2024), and interactive environment simulation (Bruce et al. 2024). At the same time, generating content that remains consistent across different regions of a scene is still non-trivial. Mirror reflections provide a representative case, since the content within a mirror is constrained by its relationship with the surrounding scene rather than being determined solely from local visual context. This raises an important

A wooden chair placed in front of a plane mirror, with a realistic reflection.

A white suitcase placed in front of a plane mirror, with a realistic reflection.

A kettle placed in front of a plane mirror, with a realistic reflection.

Figure 1: Video mirror reflection generation remains challenging even for advanced VDMs. Veo 3.1 generates visually realistic videos but often produces mirror content that is inconsistent with the visible scene, including incorrect reflected objects and implausible spatial arrangements. question: can current VDMs generate reflections that remain consistent with the visible scene? To investigate this question, we prompt a representative VDM to generate videos containing mirrors. As illustrated in Figure 1, Veo 3.1 (Veo Team 2025) produces visually realistic scenes, but the reflected content can be inconsistent with the visible environment. As highlighted by the red boxes, the reflections may contain incorrect objects, spatial arrangements that do not match the surrounding scene, or temporal inconsistencies across frames. These observations suggest that mirror reflection generation remains a challenging setting for video diffusion models. Recent studies have explored mirror reflection generation in image diffusion models (Dhiman et al. 2025; Dhiman, Shah, and Babu 2025) by formulating the task as an image inpainting problem. While these approaches show promising results for image synthesis, they are designed for singleimage generation and do not explicitly model consistency across video frames. When applied independently to each frame, reflected appearances may vary over time. In addition, existing methods focus on object-centric scenarios in which one or a few foreground objects are placed in front of a mirror. Such settings only partially capture real-world videos, where mirrors often reflect complex scenes containing multiple objects, background structures, camera motion, and interactions. Extending reflection generation from im-

ages to videos therefore requires producing plausible reflections in individual frames while maintaining consistent scene-to-mirror relationships a video . Mirror reflection generation differs from conventional video inpainting in two ways. First, mirror content is not arbitrary in our setting and should be inferred from reflectionrelevant scene content. Second, reflected content should be arranged consistently within the mirror region. This requires addressing two complementary questions: what should be reflected, and how it should be arranged. Motivated by this observation, we propose MirrorWorld, a reflection-aware video inpainting framework that models scene-to-mirror relationships. Given a video with masked mirror regions, MirrorWorld reconstructs missing reflections while preserving non-mirror content by aligning scene features and mirror-region features. MirrorWorld consists of two components. Semantic Relation Distillation (SRD) transfers relational information from a frozen visual foundation model to encourage mirrorregion representations to associate with reflection-relevant content in the visible scene. Geometric Transformation Alignment (GTA) learns a transformation that guides the spatial arrangement of reflected content. Intuitively, SRD addresses what should be reflected, while GTA addresses how it should be arranged, together encouraging scene-consistent reflection generation. To facilitate future research, we construct a benchmark for video mirror reflection generation by repurposing four existing video mirror datasets into a unified reflection reconstruction task. The benchmark includes diverse scenes, object configurations, camera motions, and mirror appearances, enabling systematic evaluation. Experiments demonstrate improved reflection reconstruction quality over representative image-based reflection generation methods and strong video inpainting baselines. Ablation studies show complementary benefits from semantic relation learning and transformation alignment, supporting the proposed what-to-reflect and howto-reflect decomposition. Our contributions are: • We identify two key challenges in video mirror reflection generation: determining what scene content should be reflected and how reflected content should be spatially arranged within mirror regions. • We propose MirrorWorld, a reflection-aware video inpainting framework that combines Semantic Relation Distillation (SRD) and Geometric Transformation Alignment (GTA) to align visible scene content with mirror regions. • We formulate video mirror reflection generation as a reflection-aware video reconstruction task and construct a benchmark for systematic evaluation. • Experiments on the constructed benchmark demonstrate improved reflection reconstruction quality over representative image-based reflection generation methods and strong video inpainting baselines.

Related Work Video Diffusion Models Diffusion models have become a dominant paradigm for video generation. Video Diffusion Models (Ho et al. 2022) extend image diffusion models to jointly model spatial appearance and temporal dynamics, while VideoLDM (Blattmann et al. 2023) enables high-resolution video generation in the latent space. Recent methods further improve visual quality and scalability through motion modules and diffusion transformers (Guo et al. 2024; Chen et al. 2024; Yang et al. 2025b). Recent work has also investigated physical plausibility in generated videos. VideoPhy (Bansal et al. 2025) and PhyGenBench (Meng et al. 2025) evaluate physical commonsense, while VLIPP (Yang et al. 2025a) and PhysHPC (Chen et al. 2025) improve physical fidelity using physical priors and preference alignment. Inspired by REPA (Yu et al. 2025), VideoREPA (Zhang et al. 2026) transfers relational representations from video understanding to generation models. However, these methods focus on motion quality, physical plausibility, or representation alignment for video generation. In contrast, mirror reflection generation requires reasoning about relationships between visible scene content and mirror regions. MirrorWorld addresses this challenge through semantic relation learning and transformation alignment for reflection generation.

Video Inpainting Methods Video inpainting reconstructs missing regions from visible frames and text prompts. Early data-driven methods exploit information across frames through optical flow, feature propagation, and transformers (Li et al. 2022; Zhang, Fu, and Liu 2022; Zhou et al. 2023). Recent generative methods leverage video diffusion priors to improve visual quality and diversity (Zhang et al. 2024; Zi et al. 2025; Bian et al. 2025; Jiang et al. 2025; Litman et al. 2026). However, these methods are not designed to model relationships between mirror regions and the surrounding scene and, when applied to mirrors, may rely on local context and generative priors. In mirror region inpainting, generated content should remain consistent with visible scene content. MirrorWorld instead explicitly models the dependency between visible context and mirror regions.

Mirror Understanding and Reflection Generation Mirror-related vision research has primarily focused on mirror detection and segmentation. Image-based methods distinguish mirrors using contextual similarity (Lin, Wang, and Lau 2020), depth discontinuities (Mei et al. 2021), and semantic associations (Guan, Lin, and Lau 2022). Video-based methods further leverage spatial-temporal relationships (Lin, Tan, and Lau 2023), weak supervision (Xu, Siu, and Lau 2024), and motion cues (Warren et al. 2024) to improve mirror localization across frames. While these methods exploit mirror-scene relationships for localization and segmentation, they do not address reflection generation. More recently, MirrorFusion (Dhiman et al. 2025) introduces a diffusion-based mirror reflection generation method using image inpainting and depth conditioning for single-object reflection synthesis. MirrorVerse (Dhiman, Shah, and Babu 2025) extends this

(1) Semantic Relation Distillation (SRD)

...

Projector

Diffusion Features FD

Relation Matrix

(2) Geometric Transformation Alignment (GTA) Projector

Condition Encoder

Local Temporal Window t-r t t+r

...

Source Map St

Text Encoder

Ct

Average

Transformation Regressor

Geometric Features FG

Text Prompt T The mirror region in the scene is naturally completed, ...

Outside Mirror Tokens

...

VAE Encoder

Masked Video Vm

DiT Block

Mirror Mask M

Inside Mirror Tokens

...

Diffusion Loss

Relation Matrix

Outside Mirror Tokens

...

Reference Features FR= (V)

Inside Mirror Tokens

...

Visual Foundation Model

...

Target Video V

Warped Visible Features

Mirror-Region Features

Figure 2: Overview of MirrorWorld. Given a masked video, a mirror mask, and a text prompt, the conditional VDM reconstructs the missing mirror regions. During training, Semantic Relation Distillation (SRD) transfers cross-region relations from a frozen visual foundation model to constrain what visible content should be reflected, while Geometric Transformation Alignment (GTA) learns temporally conditioned feature-space transformations to constrain how this content should be arranged. setting to multi-object configurations. However, existing reflection generation methods remain limited to image-based and object-centric settings. Existing mirror datasets (Lin, Tan, and Lau 2023; Xu, Siu, and Lau 2024; Warren et al. 2024; Xu et al. 2025) are designed for detection and segmentation rather than reflection reconstruction. In contrast, MirrorWorld addresses reflection generation in videos by leveraging relationships between visible content and mirror regions, handling scene-level reflections in real-world environments, and providing a benchmark for systematic evaluation.

Method Overview Our goal is to develop a video mirror reflection generation framework, denoted as MirrorWorld, which reconstructs reflection-consistent mirror content while preserving the visible non-mirror regions. Formally, given an input video Vm with its mirror regions masked, the corresponding binary mirror mask M, and a text prompt T, MirrorWorld generates b = G(Vm , M, T), where G denotes the conditional video V b is expected video generation model. The generated video V to preserve the observed scene content while reconstructing mirror regions that remain semantically compatible and geometrically consistent with the visible scene. To achieve this, MirrorWorld integrates two complementary components. (1) Semantic Relation Distillation (SRD). We transfer relational knowledge from a frozen visual foun-

dation model to establish semantic correspondences between mirror regions and reflection-relevant content in the visible scene. (2) Geometric Transformation Alignment (GTA). To determine how the associated scene content should appear within the mirror, we learn its spatial transformation from the visible scene to the reflected regions. Figure 2 illustrates the overall architecture.

Semantic Relation Distillation (SRD) Semantic Relation Distillation (SRD) aims to establish semantic correspondences between mirror regions and the visible scene. A plausible reflection should contain elements that are semantically related to the surrounding visible content. However, such scene-to-mirror semantic relations are not explicitly encoded in the pretrained VDMs. Hence, we distill semantic relational knowledge from a frozen visual foundation model (Wang et al. 2023) into the diffusion representation. We employ a frozen visual foundation model (Wang et al. 2023) E to extract reference semantic features FR = E(V) from the target video V. Given the noisy video latent zs at diffusion timestep s, we extract a hidden representation Hθ = fθ (zs , s, Vm , M, T) from the VDM fθ , which is then passed through a light weight MLP projection head ϕSRD to obtain the diffusion features FD = ϕSRD (Hθ ). The diffusion features and mirror mask are resampled onto the reference token grid. For each valid temporal slice τ , let Iτ and Oτ denote the sampled token indices inside and outside the mirror, re-

spectively. Following relational representation alignment (Yu et al. 2025), we measure the relation between two feature vectors using cosine similarity: sim(a, b) = a⊤ b/(∥a∥2 ∥b∥2 ). The scene-to-mirror semantic relations obtained from the reference features FR and diffusion features are defined as:
R R RR τ (i, j) = sim Fτ,i , Fτ,j , (1)
D D RD τ (i, j) = sim Fτ,i , Fτ,j .

where W denotes differentiable bilinear warping. Thus, the K-step context is used only to estimate At , while the transformation itself is applied to the current source feature map St . GTA measures the cosine distance between the warped visible features and the corresponding mirror-region features:

The SRD objective aligns the scene-to-mirror semantic relation as X 1 X 1 D LSRD = RR τ (i, j) − Rτ (i, j) , |VS | |Iτ ||Oτ |

(6) Here, VG contains the temporal indices with sufficient mirrorregion and outside-region support. This objective encourages the transformed visible features to match the mirror-region representation.

(2) where VS contains the temporal slices with sufficient tokens in both regions. Since visual foundation model E is frozen, gradients are propagated only through the diffusion representation.

Training Objective

τ ∈VS

i∈Iτ j∈Oτ

Geometric Transformation Alignment (GTA) SRD identifies semantic relationships between mirror regions and visible scene content, but it does not explicitly determine how the associated content should be spatially arranged within the reflection. We therefore introduce Geometric Transformation Alignment (GTA) to learn a feature-space mapping from visible scene content to its reflected regions. We obtain the geometric features FG = ϕGTA (Hθ ) using a separate trainable MLP projection head ϕGTA . For each temporal index t, we construct a source map containing only visible scene information: ( FG q ∈ Ot , t (q), P St (q) = (3) 1 G (p), q ∈ It . F p∈Ot t |Ot | where It and Ot denote the sets of locations inside and outside the mirror, respectively. Replacing mirror-region features with the mean feature of the visible region prevents the transformation estimator from directly accessing the target features inside the mirror. Estimating the transformation independently at each temporal step may produce unstable geometric mappings. GTA therefore aggregates a local temporal window of K = 2r + 1 feature steps, where r is the temporal radius. Specifically, we compute the local context as r

1 X Sclip(t+k) , Ct = K

(4)

k=−r

where clip(·) replicates boundary indices. The window is defined over the temporal grid of the video features. A shared transformation regressor gη predicts an affine matrix from the temporal context At = gη (Ct ). The regressor predicts an affine transformation initialized from the identity mapping. The predicted transformation is then applied to the source feature at the current temporal index to obtain the warped visible features b G = W (St , At ) , F t

(5)

LGTA =

 i 1 X 1 Xh b G (q), FG (q) . 1 − sim F t t |VG | |It | t∈VG

q∈It

Given a target video V, the video VAE encodes it into a clean latent z0 . We sample a diffusion timestep s and Gaussian noise ϵ ∼ N (0, I) to construct the noisy latent zs = qs (z0 , ϵ). Let c = (Vm , M, T) denote the masked video, mirror mask, and text condition. The base diffusion objective is Ldiff = Ez0 ,s,ϵ [w(s) MSE (uθ (zs , s, c), u∗s )] ,

(7)

where uθ is the model prediction, u∗s is the scheduler-defined training target, and w(s) is the timestep-dependent weighting function. The final objective combines the diffusion generation loss with SRD and GTA: L = Ldiff + λSRD LSRD + λGTA LGTA ,

(8)

where λSRD and λGTA control the strength of the two objectives. This strategy enables the VDM to learn what visible scene content is associated with the mirror region from SRD and how this content is spatially transformed within the reflection in GTA. Benchmark We construct a video mirror reflection generation benchmark from four existing video mirror segmentation and detection datasets: VMD-D (Lin, Tan, and Lau 2023), ZOOM (Xu, Siu, and Lau 2024), MMD (Warren et al. 2024), and DVMD-D (Xu et al. 2025). These datasets provide realworld videos together with annotated mirror masks, covering diverse scenes, camera motions, and mirror appearances. By integrating the four datasets into a shared evaluation environment, our benchmark defines a unified reflection-aware video reconstruction task, in which all methods receive the same masked videos, mirror masks, and text prompts. Since the original videos vary substantially in length, we divide them into clips of at most 49 frames and discard clips shorter than 10 frames to satisfy the input-length constraints of VDMs. This process produces 1,242 video clips. We perform the train-test split at the source-video level, ensuring that the clips from the same source video never appear in both the training and testing sets. This results in 1,142 training clips and 100 testing clips. For each clip, the original video serves as the target, while the conditional input is constructed by masking the annotated mirror regions.

Experiments

Method

PSNR ↑

MirrorFusion MirrorVerse

9.508 9.666

Evaluation Protocol Implementation Details Our model is built upon Wan2.1VACE-14B (Jiang et al. 2025), which is fine-tuned using LoRA (Hu et al. 2022) with rank 32. For SRD, we employ the frozen VideoMAEv2-Base (Wang et al. 2023) as the visual foundation model. The GTA transformation estimator uses a local window of K = 5 to predict the affine transformation. The weights of the SRD and GTA objectives are set to λSRD = 0.05 and λGTA = 0.01, respectively. We train MirrorWorld on 4 NVIDIA A100 80GB GPUs with a global batch size of 4, using the AdamW (Loshchilov and Hutter 2019) optimizer with a learning rate of 10−4 . Baselines We compare MirrorWorld with two categories of baseline methods. Image mirror reflection generation methods include MirrorFusion (Dhiman et al. 2025) and MirrorVerse (Dhiman, Shah, and Babu 2025), which are applied independently to each video frame. Video inpainting models include VideoPainter (Bian et al. 2025) and VACE (Jiang et al. 2025). VACE adopts Wan2.1-14B as the base model. To ensure a fair comparison, all baseline methods are finetuned on our training set using their official implementations. During evaluation, all methods use the same masked videos, mirror masks, and text prompts. Metrics We evaluate the generated mirror reflections using Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS). These metrics, computed within mirror regions, quantify how well each method reconstructs groundtruth reflections that are consistent with the observed scene. We further report Fréchet Video Distance (FVD) to measure video-level generation quality.

Comparison with Baseline Methods Quantitative Evaluation Table 1 compares MirrorWorld with representative image mirror reflection generation and video inpainting methods. MirrorWorld achieves the strongest overall performance across both mirror-region reconstruction and video-level evaluation. Compared with the VACE backbone, MirrorWorld produces reflections that more closely match the target videos in pixel values, structural appearance, and perceptual features. The improved FVD score further indicates that the generated videos are closer to the distribution of real videos. Together, these results support the benefit of introducing reflectionspecific semantic and geometric supervision beyond conventional video inpainting. The video-based methods generally perform better than the image-based approaches. Since MirrorFusion and MirrorVerse process each frame independently, they do not explicitly preserve reflection consistency across frames. Their weaker video-level performance is consistent with this limitation. Qualitative Evaluation Figure 3 presents qualitative comparisons on two test videos. The image-based methods exhibit clear semantic and temporal inconsistencies. In the first example, MirrorFusion generates unrelated shelves and people inside the mirror, while MirrorVerse produces dark or

SSIM ↑

LPIPS ↓

FVD ↓

Image-based Methods 0.293 0.312

0.699 0.680

513.062 416.146

0.606 0.493 0.488

229.558 191.617 184.868

Video-based Methods VideoPainter VACE Ours

11.282 13.537 14.005

0.399 0.489 0.504

Table 1: Quantitative comparison with baseline methods. PSNR, SSIM, and LPIPS are evaluated within the mirror regions. The best and second-best results are highlighted in bold and underlined. structurally unrelated content. These reflections also change substantially across frames. VideoPainter generates more coherent local appearance, but the reflected content remains weakly aligned with the surrounding scene and contains inconsistent boundaries and spatial arrangements. The video-based VACE baseline better preserves the overall scene structure, yet still fails to maintain the scene-tomirror relationship. In the first example, it duplicates the wall painting and places two framed objects inside the mirror. In the second example, VACE incorrectly generates clothing reflections in the regions highlighted by the orange boxes, although the visible clothing positions indicate that no such reflections should appear there. MirrorWorld avoids these duplicated and misplaced reflections. These comparisons are consistent with our design of learning both what should be reflected and how it should appear in the mirror. Specifically, SRD learns to associate the mirror region with relevant visible content, while GTA learns how this content should be spatially arranged within the reflection.

Ablation Study Key Components We conduct ablation studies to examine the contributions of SRD and GTA. The fine-tuned VACE model serves as the baseline. As shown in Table 2, using either SRD or GTA alone provides only limited improvements in mirror-region reconstruction. SRD models the semantic association between the visible scene and mirror regions, but does not explicitly constrain how the associated content should be spatially arranged. In contrast, GTA models the geometric transformation into the mirror region, but lacks explicit semantic supervision regarding the content being transformed. Each component alone therefore addresses only one aspect of reflection generation. Notably, GTA alone obtains the lowest FVD among the evaluated variants. GTA estimates the geometric transformation using local temporal context, providing consistent spatial guidance across neighboring frames. This design is consistent with its stronger video-level distribution quality. Combining SRD and GTA produces the strongest mirror-region reconstruction, supporting their complementary roles. SRD models what visible scene content should be reflected, while GTA models how this content should be arranged within the mirror. Their combination jointly constrains the semantic content and spatial

Input MirrorFusion MirrorVerse VideoPainter VACE Ours

Figure 3: Qualitative comparisons on video mirror reflection generation. Existing methods introduce unrelated content or place reflected objects in implausible regions, whereas MirrorWorld generates reflections consistent with the visible scene and its spatial arrangement. Orange boxes highlight incorrectly placed reflections. SRD

GTA

PSNR ↑

SSIM ↑

LPIPS ↓

FVD ↓

Model

PSNR ↑

SSIM ↑

LPIPS ↓

FVD ↓

0.489 0.500 0.495 0.504

0.493 0.489 0.490 0.488

191.617 185.708 174.444 184.868

DINOv3 VideoMAEv2

13.503 14.006

0.494 0.504

0.494 0.488

183.304 184.868

✓ ✓

13.537 13.551 13.542 14.005

✓ ✓

Table 2: Ablation study on key components. organization of the generated reflection. Figure 4 further illustrates the complementary roles of SRD and GTA. Without SRD, the model generates two adjacent framed artworks inside the mirror, resulting in incorrect reflected content. Without GTA, only part of the metal ring is reconstructed, as highlighted by the orange box. Since only half of the ring is visible outside the mirror, this incomplete reflection indicates that the model directly copies the visible fragment without learning its geometric transformation. In contrast, the full model reconstructs the complete ring with a coherent spatial arrangement. These comparisons are consistent with our design that SRD learns what visible con-

Table 3: Ablation on the visual foundation model in SRD. tent should be reflected, while GTA learns how this content should appear within the reflection. Visual Foundation Model for SRD We investigate the choice of visual foundation model for SRD by replacing VideoMAEv2 (Wang et al. 2023) with DINOv3 (Siméoni et al. 2025). As shown in Table 3, VideoMAEv2 performs better across all three mirror-region reconstruction metrics, indicating more accurate reflected content. DINOv3 obtains a slightly lower FVD, suggesting a small advantage in videolevel distribution quality that does not translate into improved mirror-region reconstruction. These results suggest that VideoMAEv2 provides more suitable semantic relational supervision for SRD.

(a) Input

(b) with all

(c) w/o SRD

(d) w/o GTA

Frame 1

Frame 25

Frame 49

Figure 5: Failure case analysis. The reflected person is located outside the camera’s field of view. Although the generated reflection remains temporally consistent, its exact content cannot be determined from the visible scene.

Figure 4: Visual ablation of key components. Without SRD, the model produces incorrect reflected content. Without GTA, it copies the partially visible metal ring and produces an incomplete reflection, as highlighted by the orange box. The full model preserves both semantic and geometric consistency.

Temporal Context

PSNR ↑

SSIM ↑

LPIPS ↓

FVD ↓

Per-frame (K = 1) K=5 K=7 All frames

13.543 14.005 13.541 13.528

0.494 0.504 0.495 0.497

0.499 0.488 0.494 0.490

175.789 184.868 181.303 189.899

Table 5: Ablation on the temporal window size in GTA. cues required for reflection alignment.

Transformation

PSNR ↑

SSIM ↑

LPIPS ↓

FVD ↓

Flip Geometric

13.802 14.005

0.501 0.504

0.490 0.488

189.707 184.868

Table 4: Ablation on the transformation strategy in GTA. Transformation Strategy for GTA We investigate the transformation strategy for GTA by comparing our geometric transformation with a learnable flip-based variant. The Flip variant first horizontally flips the visible-scene features and then applies a trainable residual MLP to refine the feature at each spatial location. Although the MLP learns feature corrections, the underlying spatial correspondence remains determined by horizontal flipping. As shown in Table 4, the geometric transformation performs better across all three mirror-region reconstruction metrics and obtains a lower FVD. This suggests that correcting horizontally flipped features alone is insufficient to capture scene-dependent reflection geometry. By estimating a spatial transformation from local temporal context, GTA more effectively learns how visible content should be arranged within the reflection. Temporal Window Size for GTA We investigate the temporal window size used by GTA by comparing per-frame transformation estimation, local windows of different sizes, and aggregation over all frames. As shown in Table 5, using K = 5 performs best across all three mirror-region reconstruction metrics, indicating more accurate reflected content. Although the per-frame setting obtains the lowest FVD, its weaker reconstruction metrics suggest that this video-level advantage does not translate into improved mirror-region quality. Increasing the window size to K = 7 or aggregating features from all frames does not further improve mirrorregion reconstruction. These results suggest that GTA benefits most from a moderate local context. Frame-wise estimation provides insufficient temporal evidence, whereas an overly broad window may dilute the frame-specific geometric

Failure Case Analysis Our formulation reconstructs mirror content by establishing semantic and geometric correspondences with the visible scene. It therefore assumes that reflection-relevant content is at least partially visible in the input video. When a mirror reflects content located entirely outside the camera’s field of view, SRD and GTA cannot directly constrain its semantic identity or spatial configuration. Figure 5 illustrates this ambiguity. The person reflected in the mirror is located behind the camera and does not appear in the visible non-mirror region. Although the model produces a temporally consistent reflection, its exact content cannot be determined from the observed scene. In such cases, the model can only synthesize reflections that are consistent with observed evidence and learned priors, rather than recover the actual unseen content.

Conclusion We presented MirrorWorld, a reflection-aware video inpainting framework that models the relationship between visible scene content and mirror regions. MirrorWorld decomposes reflection generation into two complementary questions: what content should be reflected and how it should be spatially arranged. SRD transfers semantic relations from a frozen visual foundation model, while GTA aligns visible scene features with mirror-region representations through learned geometric transformations. We also constructed a unified video mirror reflection generation benchmark from four existing mirror datasets. Experiments show that MirrorWorld improves reflection reconstruction and video-level quality over representative image-based reflection generation and video inpainting methods, while ablation studies support the complementary roles of SRD and GTA. While the current framework focuses on feature-space transformations, modeling more complex reflection geometry remains an important direction for future work.

References Bansal, H.; Lin, Z.; Xie, T.; Zong, Z.; Yarom, M.; Bitton, Y.; Jiang, C.; Sun, Y.; Chang, K.-W.; and Grover, A. 2025. VideoPhy: Evaluating Physical Commonsense for Video Generation. In Yue, Y.; Garg, A.; Peng, N.; Sha, F.; and Yu, R., eds., International Conference on Learning Representations, volume 2025, 102075–102121. Bian, Y.; Zhang, Z.; Ju, X.; Cao, M.; Xie, L.; Shan, Y.; and Xu, Q. 2025. VideoPainter: Any-length Video Inpainting and Editing with Plug-and-Play Context Control. In Proceedings of the Special Interest Group on Computer Graphics and Interactive Techniques Conference Conference Papers, SIGGRAPH Conference Papers ’25. New York, NY, USA: Association for Computing Machinery. ISBN 9798400715402. Blattmann, A.; Rombach, R.; Ling, H.; Dockhorn, T.; Kim, S. W.; Fidler, S.; and Kreis, K. 2023. Align Your Latents: High-Resolution Video Synthesis With Latent Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 22563– 22575. Bruce, J.; Dennis, M. D.; Edwards, A.; Parker-Holder, J.; Shi, Y.; Hughes, E.; Lai, M.; Mavalankar, A.; Steigerwald, R.; Apps, C.; Aytar, Y.; Bechtle, S. M. E.; Behbahani, F.; Chan, S. C.; Heess, N.; Gonzalez, L.; Osindero, S.; Ozair, S.; Reed, S.; Zhang, J.; Zolna, K.; Clune, J.; Freitas, N. D.; Singh, S.; and Rocktäschel, T. 2024. Genie: Generative Interactive Environments. In Salakhutdinov, R.; Kolter, Z.; Heller, K.; Weller, A.; Oliver, N.; Scarlett, J.; and Berkenkamp, F., eds., Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, 4603–4623. PMLR. Carreira, J.; and Zisserman, A. 2017. Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Chen, H.; Zhang, Y.; Cun, X.; Xia, M.; Wang, X.; Weng, C.; and Shan, Y. 2024. VideoCrafter2: Overcoming Data Limitations for High-Quality Video Diffusion Models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 7310–7320. Chen, H. H.; Huang, H.; Chen, Q.; Yang, H.; and Lim, S. N. 2025. Hierarchical Fine-grained Preference Optimization for Physically Plausible Video Generation. In Belgrave, D.; Zhang, C.; Lin, H.; Pascanu, R.; Koniusz, P.; Ghassemi, M.; and Chen, N., eds., Advances in Neural Information Processing Systems, volume 38, 133919–133951. Curran Associates, Inc. Dhiman, A.; Shah, M.; and Babu, R. V. 2025. MirrorVerse: Pushing Diffusion Models to Realistically Reflect the World. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 11239–11249. Dhiman, A.; Shah, M.; Parihar, R.; Bhalgat, Y.; Boregowda, L. R.; and Babu, R. V. 2025. Reflecting Reality: Enabling Diffusion Models to Produce Faithful Mirror Reflections. In 2025 International Conference on 3D Vision (3DV), 824– 834.

Guan, H.; Lin, J.; and Lau, R. W. 2022. Learning Semantic Associations for Mirror Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 5941–5950. Guo, Y.; Yang, C.; Rao, A.; Liang, Z.; Wang, Y.; Qiao, Y.; Agrawala, M.; Lin, D.; and Dai, B. 2024. AnimateDiff: Animate Your Personalized Text-to-Image Diffusion Models without Specific Tuning. International Conference on Learning Representations. HaCohen, Y.; Chiprut, N.; Brazowski, B.; Shalem, D.; Moshe, D.; Richardson, E.; Levin, E.; Shiran, G.; Zabari, N.; Gordon, O.; Panet, P.; Weissbuch, S.; Kulikov, V.; Bitterman, Y.; Melumian, Z.; and Bibi, O. 2024. LTX-Video: Realtime Video Latent Diffusion. arXiv preprint arXiv:2501.00103. Ho, J.; Salimans, T.; Gritsenko, A.; Chan, W.; Norouzi, M.; and Fleet, D. 2022. Video Diffusion Models. In Koyejo, S.; Mohamed, S.; Agarwal, A.; Belgrave, D.; Cho, K.; and Oh, A., eds., Advances in Neural Information Processing Systems, volume 35, 8633–8646. Curran Associates, Inc. Hu, E. J.; yelong shen; Wallis, P.; Allen-Zhu, Z.; Li, Y.; Wang, S.; Wang, L.; and Chen, W. 2022. LoRA: Low-Rank Adaptation of Large Language Models. In International Conference on Learning Representations. Jiang, Z.; Han, Z.; Mao, C.; Zhang, J.; Pan, Y.; and Liu, Y. 2025. VACE: All-in-One Video Creation and Editing. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 17191–17202. Ju, X.; Liu, X.; Wang, X.; Bian, Y.; Shan, Y.; and Xu, Q. 2025. BrushNet: A Plug-and-Play Image Inpainting Model with Decomposed Dual-Branch Diffusion. In Leonardis, A.; Ricci, E.; Roth, S.; Russakovsky, O.; Sattler, T.; and Varol, G., eds., Computer Vision – ECCV 2024, 150–168. Cham: Springer Nature Switzerland. ISBN 978-3-031-72661-3. Ke, B.; Obukhov, A.; Huang, S.; Metzger, N.; Daudt, R. C.; and Schindler, K. 2024. Repurposing Diffusion-Based Image Generators for Monocular Depth Estimation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 9492–9502. Kong, W.; Tian, Q.; Zhang, Z.; Min, R.; Dai, Z.; Zhou, J.; Xiong, J.; Li, X.; Wu, B.; Zhang, J.; et al. 2024. Hunyuanvideo: A systematic framework for large video generative models. arXiv preprint arXiv:2412.03603. Labs, B. F. 2024. FLUX. https://github.com/black-forestlabs/flux. Lai, W.-S.; Huang, J.-B.; Wang, O.; Shechtman, E.; Yumer, E.; and Yang, M.-H. 2018. Learning Blind Video Temporal Consistency. In Proceedings of the European Conference on Computer Vision (ECCV). Li, Z.; Lu, C.-Z.; Qin, J.; Guo, C.-L.; and Cheng, M.-M. 2022. Towards an End-to-End Framework for Flow-Guided Video Inpainting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 17562– 17571. Lin, J.; Tan, X.; and Lau, R. W. 2023. Learning To Detect Mirrors From Videos via Dual Correspondences. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 9109–9118.

Lin, J.; Wang, G.; and Lau, R. W. 2020. Progressive Mirror Detection. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). Litman, Y.; Liu, S.; Seyb, D.; Milef, N.; Zhou, Y.; Marshall, C.; Tulsiani, S.; and Leak, C. 2026. EditCtrl: Disentangled Local and Global Control for Real-Time Generative Video Editing. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 8965– 8975. Loshchilov, I.; and Hutter, F. 2019. Decoupled Weight Decay Regularization. In International Conference on Learning Representations. Mei, H.; Dong, B.; Dong, W.; Peers, P.; Yang, X.; Zhang, Q.; and Wei, X. 2021. Depth-Aware Mirror Segmentation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 3044–3053. Meng, F.; Liao, J.; Tan, X.; Lu, Q.; Shao, W.; Zhang, K.; Cheng, Y.; Li, D.; and Luo, P. 2025. Towards World Simulator: Crafting Physical Commonsense-Based Benchmark for Video Generation. In Singh, A.; Fazel, M.; Hsu, D.; LacosteJulien, S.; Berkenkamp, F.; Maharaj, T.; Wagstaff, K.; and Zhu, J., eds., Proceedings of the 42nd International Conference on Machine Learning, volume 267 of Proceedings of Machine Learning Research, 43781–43806. PMLR. Ronneberger, O.; Fischer, P.; and Brox, T. 2015. U-Net: Convolutional Networks for Biomedical Image Segmentation. In Navab, N.; Hornegger, J.; Wells, W. M.; and Frangi, A. F., eds., Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, 234–241. Cham: Springer International Publishing. ISBN 978-3-319-24574-4. Siméoni, O.; Vo, H. V.; Seitzer, M.; Baldassarre, F.; Oquab, M.; Jose, C.; Khalidov, V.; Szafraniec, M.; Yi, S.; Ramamonjisoa, M.; Massa, F.; Haziza, D.; Wehrstedt, L.; Wang, J.; Darcet, T.; Moutakanni, T.; Sentana, L.; Roberts, C.; Vedaldi, A.; Tolan, J.; Brandt, J.; Couprie, C.; Mairal, J.; Jégou, H.; Labatut, P.; and Bojanowski, P. 2025. DINOv3. arXiv:2508.10104. Skorokhodov, I.; Tulyakov, S.; and Elhoseiny, M. 2022. StyleGAN-V: A Continuous Video Generator With the Price, Image Quality and Perks of StyleGAN2. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 3626–3636. Teed, Z.; and Deng, J. 2020. RAFT: Recurrent All-Pairs Field Transforms for Optical Flow. In Vedaldi, A.; Bischof, H.; Brox, T.; and Frahm, J.-M., eds., Computer Vision – ECCV 2020, 402–419. Cham: Springer International Publishing. ISBN 978-3-030-58536-5. Unterthiner, T.; Van Steenkiste, S.; Kurach, K.; Marinier, R.; Michalski, M.; and Gelly, S. 2018. Towards accurate generative models of video: A new metric & challenges. arXiv preprint arXiv:1812.01717. Veo Team. 2025. Veo: A Text-to-Video Generation System. Technical report, Google DeepMind. Wan, T.; Wang, A.; Ai, B.; Wen, B.; Mao, C.; Xie, C.-W.; Chen, D.; Yu, F.; Zhao, H.; Yang, J.; Zeng, J.; Wang, J.; Zhang, J.; Zhou, J.; Wang, J.; Chen, J.; Zhu, K.; Zhao, K.; Yan, K.; Huang, L.; Feng, M.; Zhang, N.; Li, P.; Wu, P.;

Chu, R.; Feng, R.; Zhang, S.; Sun, S.; Fang, T.; Wang, T.; Gui, T.; Weng, T.; Shen, T.; Lin, W.; Wang, W.; Wang, W.; Zhou, W.; Wang, W.; Shen, W.; Yu, W.; Shi, X.; Huang, X.; Xu, X.; Kou, Y.; Lv, Y.; Li, Y.; Liu, Y.; Wang, Y.; Zhang, Y.; Huang, Y.; Li, Y.; Wu, Y.; Liu, Y.; Pan, Y.; Zheng, Y.; Hong, Y.; Shi, Y.; Feng, Y.; Jiang, Z.; Han, Z.; Wu, Z.-F.; and Liu, Z. 2025. Wan: Open and Advanced Large-Scale Video Generative Models. arXiv preprint arXiv:2503.20314. Wang, L.; Huang, B.; Zhao, Z.; Tong, Z.; He, Y.; Wang, Y.; Wang, Y.; and Qiao, Y. 2023. VideoMAE V2: Scaling Video Masked Autoencoders With Dual Masking. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 14549–14560. Wang, Z.; Bovik, A.; Sheikh, H.; and Simoncelli, E. 2004. Image quality assessment: from error visibility to structural similarity. IEEE Transactions on Image Processing, 13(4): 600–612. Warren, A.; Xu, K.; Lin, J.; Tam, G. K.; and Lau, R. W. 2024. Effective Video Mirror Detection with Inconsistent Motion Cues. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 17244–17252. Wen, Y.; Zhao, Y.; Liu, Y.; Jia, F.; Wang, Y.; Luo, C.; Zhang, C.; Wang, T.; Sun, X.; and Zhang, X. 2024. Panacea: Panoramic and controllable video generation for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 6902–6912. Xiao, J.; Yang, C.; Zhang, L.; Cai, S.; Zhao, Y.; Guo, Y.; Wetzstein, G.; Agrawala, M.; Yuille, A.; and Jiang, L. 2026. Captain Cinema: Towards Short Movie Generation. In The Fourteenth International Conference on Learning Representations. Xu, K.; Siu, T. W.; and Lau, R. W. 2024. ZOOM: Learning Video Mirror Detection with Extremely-Weak Supervision. Proceedings of the AAAI Conference on Artificial Intelligence, 38(6): 6315–6323. Xu, M.; Herbert, P.; Lai, Y.-K.; Ji, Z.; and Wu, J. 2025. RGB-D Video Mirror Detection. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 9640–9649. Yang, X.; Li, B.; Zhang, Y.; Yin, Z.; Bai, L.; Ma, L.; Wang, Z.; Cai, J.; Wong, T.-T.; Lu, H.; and Jia, X. 2025a. VLIPP: Towards Physically Plausible Video Generation with Vision and Language Informed Physical Prior. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 12360–12370. Yang, Z.; Teng, J.; Zheng, W.; Ding, M.; Huang, S.; Xu, J.; Yang, Y.; Hong, W.; Zhang, X.; Feng, G.; Yin, D.; Yuxuan.Zhang; Wang, W.; Cheng, Y.; Xu, B.; Gu, X.; Dong, Y.; and Tang, J. 2025b. CogVideoX: Text-to-Video Diffusion Models with An Expert Transformer. In The Thirteenth International Conference on Learning Representations. Yu, S.; Kwak, S.; Jang, H.; Jeong, J.; Huang, J.; Shin, J.; and Xie, S. 2025. Representation Alignment for Generation: Training Diffusion Transformers Is Easier Than You Think. In International Conference on Learning Representations.

Zhang, K.; Fu, J.; and Liu, D. 2022. Flow-guided transformer for video inpainting. In European conference on computer vision, 74–90. Springer. Zhang, R.; Isola, P.; Efros, A. A.; Shechtman, E.; and Wang, O. 2018. The Unreasonable Effectiveness of Deep Features as a Perceptual Metric. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Zhang, X.; Liao, J.; Zhang, S.; Meng, F.; Wan, X.; Yan, J.; and Cheng, Y. 2026. Videorepa: Learning physics for video generation through relational alignment with foundation models. Advances in Neural Information Processing Systems, 38: 122647–122676. Zhang, Z.; Wu, B.; Wang, X.; Luo, Y.; Zhang, L.; Zhao, Y.; Vajda, P.; Metaxas, D.; and Yu, L. 2024. AVID: Any-Length Video Inpainting with Diffusion Model. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 7162–7172. Zhou, S.; Li, C.; Chan, K. C.; and Loy, C. C. 2023. ProPainter: Improving Propagation and Transformer for Video Inpainting. In Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 10477–10486. Zi, B.; Zhao, S.; Qi, X.; Wang, J.; Shi, Y.; Chen, Q.; Liang, B.; Xiao, R.; Wong, K.-F.; and Zhang, L. 2025. CoCoCo: Improving Text-Guided Video Inpainting for Better Consistency, Controllability and Compatibility. Proceedings of the AAAI Conference on Artificial Intelligence, 39(10): 11067–11076.

Text Prompt To ensure a consistent evaluation setting, we use the same task-level text prompt for all video clips and baseline methods. As shown in Figure A, the prompt describes the desired properties of the reconstructed reflection. We avoid objectspecific descriptions so that the generated mirror content must be inferred from the visible scene and mirror mask rather than explicit semantic cues in the prompt. This unified prompt also prevents differences in prompt design from affecting the comparison between methods. Since the prompt contains only task-level instructions and no scene-specific object identities, the reconstruction is driven primarily by the visual conditions rather than scene-specific prompt engineering.

Text Prompt “The mirror region in the scene is naturally completed, with reflections consistent with the surroundings, correct structure, realistic details, and temporally coherent motion.”

MirrorWorld achieves the lowest flow warping error among the evaluated methods, indicating that its reflected content remains more consistent after motion compensation. This result is consistent with the design of GTA, which uses local temporal context to provide stable geometric guidance across neighboring frames. VACE (Jiang et al. 2025) also obtains a low warping error despite producing duplicated or incorrectly placed reflections in Figure 3 of the main paper. This distinction highlights that Ewarp measures temporal consistency on the video mirror reflection, complementing the mirror-region reconstruction metrics and qualitative comparisons. Ewarp ↓

Method Image-based Methods MirrorFusion (Dhiman et al. 2025) MirrorVerse (Dhiman, Shah, and Babu 2025)

0.192 0.188

Video-based Methods VideoPainter (Bian et al. 2025) VACE (Jiang et al. 2025) Ours

0.076 0.026 0.025

Table A: Temporal consistency comparison in mirror regions. Figure A: Text prompt for video mirror reflection generation.

Additional Ablation Study Temporal Consistency Evaluation We further evaluate temporal consistency using the flow warping error, denoted as Ewarp . Ewarp is evaluated exclusively within the mirror masks. Specifically, we measure the residual difference between adjacent generated frames after flow-based warping and average the error only over valid mirror pixels, excluding non-mirror and invalid correspondences. A lower value indicates stronger motioncompensated temporal consistency in the generated reflections. Importantly, a low Ewarp should not be interpreted as evidence of reflection correctness. A temporally stable but semantically incorrect or spatially misplaced reflection can also obtain a low warping error, as illustrated by VACE (Jiang et al. 2025) in Figure 3 of the main paper. We therefore use Ewarp as a complementary measure of temporal stability rather than a replacement for mirror-region reconstruction metrics and qualitative evaluation. As shown in Table A, the image-based methods produce the largest warping errors. This result provides quantitative support for the temporal instability observed in Figure 3 of the main paper. In the first example, MirrorFusion (Dhiman et al. 2025) generates substantially different objects and people across frames, while the content produced by MirrorVerse (Dhiman, Shah, and Babu 2025) changes between dark regions and unrelated scene structures. When these methods are applied frame by frame, the generated reflections lack explicit cross-frame constraints, resulting in unstable content and appearance. VideoPainter (Bian et al. 2025) reduces the warping error through video-level modeling, but its reflected boundaries and local structures remain less consistent across frames.

Loss Weight We study the effects of the loss weights λSRD and λGTA by varying one weight while keeping the other fixed. As shown in Figure B, λSRD = 0.05 provides the strongest mirrorregion reconstruction across PSNR, SSIM, and LPIPS. Similarly, λGTA = 0.01 performs best on all three reconstruction metrics. Smaller or larger values lead to less accurate reflected content, indicating that an intermediate strength is more suitable for both auxiliary objectives. The lowest FVD occurs at different weights from the mirror-region metrics. This difference is expected because FVD evaluates the full-frame video distribution, whereas PSNR, SSIM, and LPIPS focus on the reconstructed mirror regions. Therefore, improvements targeted at reflection correctness may not produce proportional gains in this global distribution metric. We therefore treat FVD as complementary to the mirror-region reconstruction metrics. This distinction also explains the component ablation in Table 2 of the main paper. Although GTA alone obtains the lowest FVD, combining SRD and GTA achieves the strongest mirror-region reconstruction results. Since FVD evaluates the distribution of complete frames rather than reflection correctness specifically, the lowest FVD does not imply that GTA alone is preferable for video mirror reflection generation.

Token Sharing We investigate whether SRD and GTA should operate on the same token subset. In our framework, SRD randomly samples mirror and visible-scene tokens to model their semantic relations, while GTA uses all tokens to learn the geometric

PSNR ↑

PSNR ↑

Fixed λGTA = 0.01

14.8

LPIPS ↓

0.50

200

0.50

13.6

0.49

0.48

13.2

FVD ↓

220

0.52 0.51

14.0

180

0.48

12.8

0.47

0.46 0

0.025

0.05

0.075

λSRD

0.1

0

0.025

PSNR ↑

14.8

0.05

λSRD

0.075

0.1

160 0

SSIM ↑

0.025

0.05

λSRD

0.075

0.1

LPIPS ↓

0.52

14.4

0.50

0.025

0.05

λSRD

0.075

0.1

0.015

0.02

FVD ↓

200

0.50

13.6

0.49

0.48

13.2

0

220

0.52 0.51

14.0

180

0.48

12.8 12.4

FVD ↓

LPIPS ↓

0.52

14.4

12.4

Fixed λSRD = 0.05

SSIM ↑

SSIM ↑

0.47

0.46 0

0.005

0.01

λGTA

0.015

0.02

0

0.005

0.01

λGTA

0.015

0.02

160 0

0.005

0.01

λGTA

0.015

0.02

0

0.005

0.01

λGTA

Figure B: Ablation on loss weights. We vary λSRD while fixing λGTA = 0.01 (top), and vary λGTA while fixing λSRD = 0.05 (bottom). PSNR, SSIM, and LPIPS are evaluated within mirror regions, whereas FVD uses full-frame video features. transformation. The shared variant instead reuses the token locations sampled by SRD for both components. As shown in Table B, sharing tokens slightly improves LPIPS and preserves SSIM, but reduces PSNR and substantially increases FVD. This suggests that the sparse token pairs used for semantic relation distillation do not provide sufficient spatial coverage for learning the geometric transformation. Using all tokens allows GTA to better capture how visible content should be spatially arranged, while SRD focuses on learning what content should be reflected. The LPIPS improvement is relatively small, suggesting that sharing tokens may preserve certain local perceptual properties. However, the reduced PSNR and substantially worse FVD indicate that the sparse shared token support is less suitable for accurate geometric alignment and global video-level quality.

scores RR τ (i, j). The selected index pairs are then shared with the diffusion branch to compute the corresponding relations RD τ (i, j). Thus, the two strategies differ only in how the visible-region tokens are selected, while all other settings remain unchanged. In our experiments, random sampling uses up to 64 tokens from each region, while Top-N sampling uses N = 16. As shown in Table C, random sampling performs better across all mirror-region reconstruction metrics. This suggests that covering a broader range of scene-to-mirror relations is more effective for reconstructing complex reflections than focusing only on the strongest reference relations. Although Top-N sampling achieves a lower FVD, this improvement in video-level distribution quality does not benefit the reconstruction within the mirror regions.

GTA Token

PSNR ↑

SSIM ↑

LPIPS ↓

FVD ↓

SRD Token Sampling

PSNR ↑

SSIM ↑

LPIPS ↓

FVD ↓

Shared with SRD All tokens (Ours)

13.773 14.005

0.504 0.504

0.484 0.488

196.814 184.868

Top-N Random (Ours)

13.591 14.005

0.494 0.504

0.491 0.488

173.294 184.868

Table B: Ablation on token sharing between SRD and GTA.

Token Sampling We investigate the token sampling strategy used to construct the scene-to-mirror relations in SRD. For random sampling, we independently sample token indices from both the mirror region Iτ and the visible region Oτ . The same sampled index pairs are used to construct the reference and diffusion relation D matrices RR τ and Rτ , respectively. For Top-N sampling, the mirror indices are sampled in the same manner. For each sampled mirror token i ∈ Iτ , we compute its reference relations with all visible-region tokens and select the N indices with the highest relation

Table C: Ablation on token sampling strategy in SRD.

Benchmark Statistics Our benchmark integrates four video mirror detection and segmentation datasets, including VMD-D (Lin, Tan, and Lau 2023), DVMD-D (Xu et al. 2025), MMD (Warren et al. 2024), and ZOOM (Xu, Siu, and Lau 2024), resulting in 1,242 video clips comprising 45,373 frames. As shown in Figure C, these videos cover a wide range of spatial resolutions. By using these datasets under a unified reflection reconstruction setting, our benchmark provides a larger and more diverse data resource for training and systematically evaluating video mirror reflection generation methods.

500

459

400 300 200 100

88

75

55

36

0

12

12

6

2

40 20 80 60 20 80 80 40 68 40 38 19 10 21 12 ×7 ×4 ×6 ×5 ×6 0× 0× 0× 0× 80 0× 40 12 20 60 6 8 2 4 2 2 6 5 3 3 1 7 21 10 19 38

Resolution (a)

20000

497

400 300 200

17,814

459

198 88

100 0

D-D D-D MMDZOOM VM DVM

Dataset (b)

Number of Frames

497

Number of Videos

Number of Videos

500

15000

10000

14,947

9,590

5000

3,022

0

D-D D-D MMDZOOM VM DVM

Dataset (c)

Figure C: Statistics of our benchmark. (a) Distribution of video resolutions. (b) Number of videos and (c) number of frames contributed by each dataset. In our benchmark, split assignments are determined using the identities of the original source videos. All clips derived from the same source video are assigned exclusively to either the training or testing partition, and no source video contributes clips to both partitions. This source-level partitioning prevents content leakage caused by placing different clips from the same video in separate splits. Since the source datasets contain real videos with visible mirrors, the original unmasked clips directly provide the ground-truth reflections. During benchmark construction, mirror regions are masked only in the conditional inputs, while the original clips are retained as the supervision and evaluation targets.

Additional Implementation Details Training and Inference Details of MirrorWorld During training, we preserve the original aspect ratio of each video while limiting the spatial resolution of each frame to at most 399,360 pixels. We adapt the backbone using rank-32 LoRA, applied to the query, key, value, output, and feedforward projections of the transformer. All pretrained backbone parameters remain frozen throughout training. SRD and GTA operate on the same diffusion features, which are separately projected to 256 dimensions for their respective objectives. At each temporal step, we randomly sample up to 64 tokens from the mirror and visible regions for SRD. During inference, the auxiliary SRD and GTA branches are removed, introducing no additional inference overhead. We generate videos of up to 49 frames using 50 denoising steps and a classifier-free guidance scale of 5.0. The original aspect ratio is preserved, with the spatial resolution limited to at most 399,360 pixels per frame. No post-processing or explicit copy-and-paste operation is applied to the visible regions.

Baseline Details All baselines are fine-tuned using their officially released implementations. Original architectures and training settings are preserved wherever possible, with only minimal modifications required to support the unified benchmark protocol. All methods are fine-tuned for the same number of epochs as our method on the same training split and evaluated using the same masked videos, mirror masks, text prompts, and testing protocol. MirrorFusion and MirrorVerse We initialize MirrorFusion (Dhiman et al. 2025) and MirrorVerse (Dhiman, Shah, and Babu 2025) from the released v1 and v2 checkpoints, respectively. Each video frame is processed independently at a resolution of 512 × 512, with depth estimated from the masked frame using Marigold (Ke et al. 2024). Both models are trained with a learning rate of 10−5 and a global batch size of 16. MirrorFusion updates only BrushNet (Ju et al. 2025), whereas MirrorVerse jointly updates BrushNet and the U-Net (Ronneberger, Fischer, and Brox 2015). During inference, we use 50 denoising steps and a guidance scale of 7.5. VideoPainter We initialize VideoPainter (Bian et al. 2025) from CogVideoX-5B-I2V (Yang et al. 2025b) and its released context branch, and fine-tune only the context branch. Training uses clips of up to 49 frames at 480×720, with a learning rate of 10−5 . Following the official protocol, the first frame is generated using FLUX-Fill (Labs 2024) from the masked input without access to the ground-truth video. During inference, we use 50 denoising steps and a guidance scale of 6.0. VACE For VACE (Jiang et al. 2025), we fine-tune Wan2.1-VACE-14B using rank-32 LoRA. Training uses AdamW (Loshchilov and Hutter 2019) with a learning rate of 10−4 . During inference, we use 50 denoising steps, a guidance scale of 5.0, and a VACE conditioning scale of 1.0.

Metric Definitions We evaluate mirror-region reconstruction, overall video quality, and temporal consistency using five metrics. The PSNR, SSIM, and LPIPS values reported in all quantitative comparison and ablation tables are computed only within the annotated mirror regions. The flow warping error Ewarp is also evaluated within mirror regions, whereas FVD is computed over complete spatial frames without applying the mirror masks.

Peak Signal-to-Noise Ratio (PSNR) PSNR measures pixel-level reconstruction accuracy. We follow its standard implementation but compute the mean squared error using only pixels inside the mirror mask. Higher PSNR indicates more accurate reconstruction of the reflected content.

Structural Similarity Index Measure (SSIM) SSIM (Wang et al. 2004) evaluates the structural similarity between generated and ground-truth reflections. We compute the standard full-frame SSIM map and average it over the mirror mask, which excludes local windows extending beyond the mirror boundary. Higher SSIM indicates better preservation of reflection structure.

Learned Perceptual Image Patch Similarity (LPIPS) LPIPS (Zhang et al. 2018) measures perceptual differences using deep visual features. We use the standard spatial LPIPS implementation with an AlexNet backbone. The mirror mask is resized to the resolution of the spatial LPIPS map, and the distances are averaged only within the masked region. Lower LPIPS indicates greater perceptual similarity to the ground truth.

Fréchet Video Distance (FVD) FVD (Unterthiner et al. 2018) measures the distributional discrepancy between generated and ground-truth videos. Unlike the reconstruction metrics above, FVD is computed over the complete frames without applying the mirror masks. Since the benchmark videos vary from 10 to 49 frames, we use a fixed 10-frame prefix for every video to ensure identical temporal support across all methods and samples. These frames are resized to 224×224, and FVD is computed from their I3D (Carreira and Zisserman 2017) features using the StyleGAN-V (Skorokhodov, Tulyakov, and Elhoseiny 2022) implementation. Lower FVD indicates better overall video-level quality.

Flow Warping Error (Ewarp ) Ewarp (Lai et al. 2018) evaluates temporal consistency after motion compensation. We estimate optical flow between adjacent ground-truth frames using RAFT-Large (Teed and Deng 2020) and use it to warp the corresponding generated frames. The photometric warping error is computed only over valid mirror-region pixels, excluding out-of-bound and forward–backward inconsistent locations. Lower Ewarp indicates more temporally consistent reflections.

Additional Qualitative Results Qualitative Comparisons Figures D and E provide additional comparisons across four sampled frames. MirrorFusion (Dhiman et al. 2025) and MirrorVerse (Dhiman, Shah, and Babu 2025) generate unrelated content that changes substantially across frames. VideoPainter (Bian et al. 2025) improves temporal coherence, but its generated reflections remain weakly related to the visible scene. In Figure D, the visible wall contains both a dispenser and an electrical outlet. VACE (Jiang et al. 2025) reconstructs the dispenser but misses the reflection of the outlet in the regions highlighted by the red boxes. MirrorWorld preserves both elements and maintains their spatial relationship across frames. In Figure E, the image-based methods generate unrelated indoor structures, while VideoPainter produces a temporally coherent but semantically unrelated interior. VACE recovers scene-related content, but the reflected structure near the mirror boundary changes inconsistently across frames. MirrorWorld maintains a more coherent mirror reflection layout. These comparisons are consistent with our design of learning both what should be reflected and how it should appear within the mirror. SRD associates the mirror with relevant visible content, while GTA guides its reflected arrangement.

Qualitative Evaluation Figures F and G present additional qualitative results of MirrorWorld. Each example shows four sampled frames and covers different mirror shapes, aspect ratios, scene layouts, and motion patterns. As shown in Figure F, the reconstructed reflections remain spatially consistent with the surrounding scene as the viewpoint changes. In particular, the doorways and wall boundaries in the bathroom scenes remain stable within the mirror, while the reflected sofa and carpet preserve their relative layout across frames. Figure G further presents results on irregular mirror shapes and foreground motion. MirrorWorld adapts the generated content to polygonal and oval mirror boundaries while preserving the reflected room structure across frames. In the final example, the reflection follows the changes in the person’s pose and position. These results are consistent with our design of learning both what should be reflected and how it should appear within the mirror.

Input MirrorFusion MirrorVerse VideoPainter VACE Ours Figure D: Additional qualitative comparisons. Red boxes highlight the missing reflection of the electrical outlet in VACE. MirrorWorld reconstructs the relevant visible elements and maintains their spatial arrangement across frames.

Input MirrorFusion MirrorVerse VideoPainter VACE Ours Figure E: Additional qualitative comparisons. Red boxes highlight spatial inconsistencies in VACE. MirrorWorld maintains more coherent reflected content and spatial organization.

Input Ours Input Ours Input Ours Figure F: Additional qualitative results of MirrorWorld. MirrorWorld preserves the spatial arrangement of reflected scene structures across changing viewpoints.

Input Ours Input Ours Input Ours Figure G: Additional qualitative results under diverse mirror geometry and motion. MirrorWorld reconstructs coherent reflections for polygonal and oval mirrors and maintains the reflected subject as its pose changes across frames.

Related documents

• LiveVVT: High-Fidelity Video Virtual Try-On in Real Time#780759
• LiveVVT: High-Fidelity Video Virtual Try-On in Real Time#779384
• ViRDM: Taming Representation Distribution Matching for Few-Step Causal Video Generation#1076579
• AnyFlow: Any-Step Video Diffusion Model with On-Policy Flow Map Distillation#182755
• AnyFlow: Any-Step Video Diffusion Model with On-Policy Flow Map Distillation#180694