Paper deep dive
VGGT-Diff: Visual Geometry Meets Diffusion for Sparse-View Novel View Synthesis
Kangjie Chen, Xiangyu Li, Dongbin Zhang, Chaoda Zheng, Shijia Chen, Jinhao Deng, Hongbin Lin, Choo Sin Wai, Minqi Wang, Minghao Yang, Dake Zhong, Guorui Song, Yu Zhang, Xianming Liu, Boyang Wang
Intelligence
Status: not_run | Model: - | Prompt: - | Confidence: 0%
Entities (0)
Relation Signals (0)
No relation signals yet.
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:We present VGGT-Diff, a geometry-routed multi-view diffusion model for sparse-view novel view synthesis. Existing novel view synthesis (NVS) methods face a fundamental trade-off: reconstruction-based approaches preserve observed geometry but struggle to synthesize unseen regions, while diffusion-based methods provide strong generative priors yet rely on implicit source-to-query correspondence. VGGT-Diff bridges these regimes by routing visual geometry latents from VGGT-{\Omega} into a pretrained video diffusion model. Each visual token is associated with a 3D point and confidence, then transformed into query-aligned latent conditions through a confidence-aware Visual Geometry Router (VGR) that preserves front and back surface evidence. These conditions guide joint target-view denoising, while Point-Track Residual Consistency (PTRC) regularizes predicted-clean residuals along reliable 3D tracks, improving multi-view stability. We further introduce robust geometry conditioning, combining training-time regularization with inference-time guidance for improved robustness. Experiments show competitive or state-of-the-art performance across interpolation and extrapolation under different viewpoint difficulties. Our code is available at this https URL.
Tags
Links
- Source: https://arxiv.org/abs/2609.33253v1
- Canonical: https://arxiv.org/abs/2609.33253v1
Trouble viewing inline? Open PDF directly →
Full Text
92,206 characters extracted from source content.
Expand or collapse full text
Visual Geometry Meets Diffusion for Sparse-View Novel View Synthesis Kangjie Chen Xiangyu Li11footnotemark: 1 Dongbin Zhang Chaoda Zheng Shijia Chen Jinhao Deng †thanks: Equal contribution. Hongbin Lin Choo Sin Wai Minqi Wang Minghao Yang Dake Zhong Guorui Song Yu Zhang Xianming Liu Boyang Wang †thanks: Corresponding author. XPeng Motors The Chinese University of Hong Kong Tsinghua University Abstract We present Vggt-Diff, a geometry-routed multi-view diffusion model for sparse-view novel view synthesis. Existing novel view synthesis (NVS) methods face a fundamental trade-off: reconstruction-based approaches preserve observed geometry but struggle to synthesize unseen regions, while diffusion-based methods provide strong generative priors yet rely on implicit source-to-query correspondence. Vggt-Diff bridges these regimes by routing visual geometry latents from VGGT-Ω into a pretrained video diffusion model. Each visual token is associated with a 3D point and confidence, then transformed into query-aligned latent conditions through a confidence-aware Visual Geometry Router (VGR) that preserves front and back surface evidence. These conditions guide joint target-view denoising, while Point-Track Residual Consistency (PTRC) regularizes predicted-clean residuals along reliable 3D tracks, improving multi-view stability. We further introduce robust geometry conditioning, combining training-time regularization with inference-time guidance for improved robustness. Experiments show competitive or state-of-the-art performance across interpolation and extrapolation under different viewpoint difficulties. Our code is available at https://github.com/chenkangjie1123/VGGT-Diff. Figure 1: Vggt-Diff grounds a pretrained video diffusion prior with geometry-routed visual features for sparse-view novel-view synthesis. It maintains strong target-view fidelity across varying pose difficulties while better preserving source-observed details than diffusion-based baselines. 1 Introduction Novel view synthesis (NVS) aims to render previously unseen viewpoints of a scene from sparse observations. Existing generalizable methods broadly follow two complementary paradigms. Reconstruction-oriented approaches infer explicit or implicit scene representations, including neural radiance fields, Gaussian primitives, and feed-forward 3D representations (Mildenhall et al., 2021; Kerbl et al., 2023; Charatan et al., 2024; Chen et al., 2024a), while Transformer-based models such as LVSM (Jin et al., 2024), RayZer (Jiang et al., 2025a), LagerNVS (Szymanowicz et al., 2026), and SVSM (Kim et al., 2026) learn view synthesis with reduced explicit 3D inductive bias. These approaches provide strong geometric fidelity when target views are well supported by observations, but sparse inputs inevitably leave parts of the scene unobserved, making deterministic reconstruction increasingly under-constrained under large viewpoint changes. In unseen regions, predictions can degenerate into patch-like artifacts biased toward colors observed in nearby source views. A complementary line formulates NVS as conditional generation and exploits image or video diffusion priors to complete unseen content (Watson et al., 2022; Kong et al., 2024; Zheng and Vedaldi, 2024; Müller et al., 2024; Yu et al., 2024; Zhou et al., 2025; Wu et al., 2026a). Such models provide strong generative priors, but often lack explicit geometric conditioning to anchor generation to observed scene structure. As a result, generative priors may override geometry-supported evidence and hallucinate plausible yet inconsistent content, leading to structural drift, unstable occlusions, and cross-view inconsistency under wide-baseline interpolation and extrapolation. Meanwhile, visual geometry foundation models such as VGGT (Wang et al., 2025) and VGGT-Ω (Wang et al., 2026a) learn rich multi-view representations encoding appearance, 3D structure, correspondence, and confidence, offering a natural way to ground generative NVS with explicit geometric evidence. These limitations suggest that generative NVS should combine strong completion priors with explicit yet uncertainty-aware geometric guidance. We introduce Vggt-Diff, a geometry-routed multi-view diffusion framework that injects visual geometry latents from VGGT-Ω into a pretrained video diffusion model. Rather than rediscovering geometry from RGB and camera rays alone, Vggt-Diff associates visual tokens with 3D locations and confidence to construct spatially aligned conditions for source and query views. Since hard projection is brittle near occlusions and uncertain geometry, we propose a confidence-aware Visual Geometry Router (VGR) that preserves front and back surface evidence together with visibility uncertainty, providing query-aligned guidance without treating reconstructed geometry as a complete scene explanation. A lightweight input projection maps the concatenated source RGB latents, Plücker ray maps, noisy query latents, and routed geometry conditions into the pretrained DiT latent space, allowing Vggt-Diff to reuse the video diffusion prior with minimal architectural modification. Vggt-Diff then jointly denoises multiple query views within the same DiT sequence, enabling direct information exchange across targets. Joint generation alone, however, does not guarantee consistency with a shared 3D scene. We therefore introduce Point-Track Residual Consistency (PTRC), which uses VGGT-derived 3D correspondences to align denoising residual errors of the same physical point across generated views, rather than forcing view-dependent features or VAE latents to match. Since geometry recovered from sparse observations is inevitably imperfect, we randomly retain, attenuate, or remove the routed geometry condition during training to prevent over-reliance on uncertain projections. This regularization also enables optional Geometry-Prior CFG at inference, which strengthens geometry-aware denoising and further improves novel-view synthesis quality. In this way, visual geometry both conditions the diffusion process and regularizes multi-view generation. Together, Vggt-Diff combines geometry foundation priors with video diffusion for faithful and generative sparse-view NVS. Our contributions are threefold: (1) Geometry-routed generation. We introduce a multi-view diffusion framework whose confidence-aware Visual Geometry Router (VGR) transforms VGGT-Ω features and 3D locations into query-aligned conditions, grounding generative completion in observed scene structure; (2) Geometry-grounded consistency. We propose Point-Track Residual Consistency (PTRC), which aligns predicted-clean residuals along reliable 3D correspondences to reduce cross-view drift without suppressing valid view-dependent appearance; and (3) Geometry-condition regularization. We stochastically attenuate or drop routed geometry during training to prevent over-reliance on imperfect projections, improving robustness under sparse or inaccurate geometry. The dropped-condition branch also supports optional matched guidance at inference. Extensive experiments demonstrate competitive or state-of-the-art performance across pose difficulties and improved geometric reconstructability of jointly generated views. 2 Related Work Generalizable and Feed-Forward View Synthesis. Novel view synthesis has progressed from scene-specific representations such as NeRF and Gaussian Splatting (Mildenhall et al., 2021; Barron et al., 2021; Kerbl et al., 2023; Chen et al., 2025; Chen et al., 2026) toward generalizable models that amortize reconstruction across scenes. Early approaches infer radiance fields or image-based representations through pixel-aligned features, multi-view stereo, epipolar reasoning, or learned ray aggregation (Yu et al., 2021; Chen et al., 2021; Wang et al., 2021; Wang et al., 2022). More recent methods reduce explicit 3D inductive bias and learn view synthesis with Transformers, including SRT (Sajjadi et al., 2022), ViewFormer (Kulhánek et al., 2022), LVSM (Jin et al., 2024), Efficient-LVSM (Jia et al., 2026), and RayZer (Jiang et al., 2025a). In parallel, feed-forward models directly predict renderable 3D representations (Charatan et al., 2024; Chen et al., 2024a; Szymanowicz et al., 2024; Hong et al., 2024; Zhang et al., 2024), while recent works further improve geometry-aware scaling and efficiency through LagerNVS (Szymanowicz et al., 2026), SVSM (Kim et al., 2026), projective conditioning (Wu et al., 2026b), and SHARP (Mescheder et al., 2026). These approaches provide strong geometric fidelity and rendering, but deterministic reconstruction remains under-constrained when sparse observations do not cover content revealed by distant query views. Generative and Geometry-Guided Novel View Synthesis. Generative NVS uses image or video priors to complete unobserved regions. Early diffusion methods perform pose-conditioned synthesis or model joint multi-view distributions (Watson et al., 2022; Liu et al., 2023; Shi et al., 2023; Shi et al., 2024; Liu et al., 2024; Ye et al., 2024; Kong et al., 2024; Zheng and Vedaldi, 2024), while later methods exploit video priors for stronger cross-view coherence (Kwak et al., 2024; Voleti et al., 2024; Müller et al., 2024; Yu et al., 2024; Zhou et al., 2025). Stronger pixel-space backbones improve end-to-end NVS (Elata et al., 2025), and FrameCrafter (Wu et al., 2026a) adapts pretrained video diffusion to complete unordered posed views. Related work studies hybrid deterministic and generative modeling (Le et al., 2026), test-time video completion (Xu et al., 2026), dynamic arbitrary-view generation (Van Hoorick et al., 2026), and correspondence-supervised diffusion (Kwon et al., 2026). However, source-to-query correspondence often remains implicit, limiting robustness to large viewpoint changes and cross-view consistency. Geometry foundation models such as DUSt3R (Wang et al., 2024b), MASt3R (Leroy et al., 2024), VGGT (Wang et al., 2025), and VGGT-Ω (Wang et al., 2026a) encode multi-view geometry, correspondence, and cameras. Their geometric evidence has been combined with Gaussian reconstruction and latent video diffusion (Chen et al., 2024b), sparse-view 3D optimization (Wang et al., 2024a), joint image and geometry diffusion (Kwak et al., 2026), latent generative refinement (Hirschorn et al., 2026), geometry-conditioned video diffusion (Kang et al., 2026), and wide-baseline guidance (Zhou et al., 2026). Unlike methods that construct complete Gaussian or radiance-field scenes, Vggt-Diff routes visual geometry features, 3D points, and confidence into query-aligned diffusion conditions and reuses their correspondences to regularize jointly generated views. Geometry therefore guides and constrains the generative prior as uncertainty-aware evidence rather than replacing it with deterministic reconstruction. 3 Method Given sparse observations, novel-view synthesis must preserve the scene evidence visible in the source images while completing unobserved regions. Reconstruction models encode correspondence explicitly but have limited support for such completion; video diffusion models provide strong generative priors but leave source-to-query correspondence largely implicit. Vggt-Diff uses geometry to connect these two capabilities. Rather than treating the recovered geometry as a complete scene representation, we use it to route visual evidence into a pretrained multi-view diffusion model and to regularize the generated views along shared 3D point tracks. The pipeline is shown in Fig. 2. 3.1 Camera-Conditioned Multi-View Diffusion We consider sparse-view novel-view synthesis from M posed source images to N prescribed query cameras. The observed views and requested cameras are represented as =(is,is,is)i=1M,=(jq,jq)j=1N.S= \(I_i^s,K_i^s,T_i^s) \_i=1^M, = \(K_j^q,T_j^q) \_j=1^N. (1) Our goal is to synthesize query views ^jqj=1N\ I_j^q\_j=1^N. We place source and query views in one diffusion sequence, enabling information exchange across slots while preserving view-specific camera controls. A frozen VAE encodes training views into the clean latent volume 0=[0s;0q]Z_0=[Z_0^s;Z_0^q]. Following Wan’s linear flow formulation (Wan et al., 2025), we sample a flow time σ and Gaussian noise ϵ∼(,) ε (0,I): σ=(1−σ)0+σϵ,=ϵ−0,ℒFM=1N∑j=1N‖θ,j−j‖22.X_σ=(1-σ)Z_0+σ ε, = ε-Z_0, _FM= 1N _j=1^N _θ,j-U_j _2^2. (2) Here, σX_σ is the diffusion state and θ,jV_θ,j is the predicted velocity for query view j. The loss applies only to query slots. Query images define training targets but are never exposed as conditions. The model receives three complementary conditions in addition to σX_σ. First, Wan’s image-conditioning stream contains clean source latents and blank query slots, providing observed appearance without leaking query RGB. Second, dense Plücker ray maps encode the camera associated with every source and query pixel. Third, the geometry-routed features introduced in Sec. 3.2 provide scene-specific source-to-query correspondence. These signals are concatenated and mapped into the pretrained Wan token space by an expanded input projection. The pretrained image pathway is preserved, while newly introduced camera and geometry inputs are initialized to have no effect at the start of fine-tuning. Source and query tokens are then jointly processed by the Wan DiT. We keep spatial positional encoding but remove temporal ordering from the view axis, since the inputs form a set of camera observations rather than a video timeline. No scene-specific text is used: training and inference share the same fixed empty textual context. After joint denoising, query latents are decoded independently into the requested views. 3.2 Geometry-Routed Visual Conditioning Figure 2: Overview of Vggt-Diff. The Visual Geometry Conditioner uses VGGT-Ω points and confidence to route appearance-bearing source features into query-aligned conditions. The diffusion state, clean-source image condition, Plücker rays, and routed visual condition are concatenated channel-wise, embedded independently within each view, and jointly processed by a Wan-initialized multi-view DiT. Geometry conditioning is stochastically regularized during training, while flow matching and PTRC supervise target fidelity and cross-view consistency, respectively. Feature association and projection. Camera rays specify where a view is sampled, but not which source observation supports each query location. We establish this correspondence using a frozen VGGT-Ω (Wang et al., 2026a), which associates each source feature nf_n with a 3D point np_n and confidence cnc_n. A lightweight adapter maps these features into the diffusion conditioning space, while the points and confidence determine where each feature is routed and how strongly it contributes. Source features retain their native image correspondence and are resampled directly onto the diffusion grid. Query views instead require geometric alignment. For query camera j, let (j,j)(R_j,t_j) denote its extrinsics and (fx,j,fy,j,cx,j,cy,j)(f_x,j,f_y,j,c_x,j,c_y,j) its focal lengths and principal point. The camera-space coordinates of np_n and its projected query location are njc=(xnjc,ynjc,znjc)⊤=jn+j,nj=πj(n)=(fx,jxnjcznjc+cx,j,fy,jynjcznjc+cy,j)⊤.x_nj^c=(x_nj^c,y_nj^c,z_nj^c) =R_jp_n+t_j, _nj= _j(p_n)= (f_x,j x_nj^cz_nj^c+c_x,j,f_y,j y_nj^cz_nj^c+c_y,j ) . (3) Here, znjcz_nj^c is the query-camera depth and πj _j is perspective projection. We route nf_n to the resulting query-grid location, transferring source-observed appearance evidence rather than pre-rendered RGB values or explicit geometric predictions. Figure 3: Visual Geometry Conditioning. Our conditioner constructs a query-aligned hard anchor from z-buffered near-front source features. A lightweight layered router forms confidence-weighted front and secondary hypotheses and predicts a residual correction from them and routing statistics. Hard-anchor routing. The primary geometry condition is a depth-selected hard anchor. After projection, each source feature is assigned to its nearest query-grid cell. For every cell, a hard z-buffer retains features close to the nearest valid camera-space depth and averages them according to their normalized VGGT-Ω confidence, producing jhardG_j^hard. This route is sharp, simple, and already provides a strong source-to-query correspondence. Layered residual refinement. The hard anchor provides the primary condition, but discrete visibility may discard valid evidence near occlusion boundaries or under small geometric errors. We therefore construct confidence-weighted front and secondary hypotheses only as residual cues. For a projected feature nf_n at continuous grid coordinate (un,vn)(u_n,v_n) with normalized confidence c~n c_n, its weight and aggregated feature at cell q, with grid coordinate =(qx,qy)q=(q_x,q_y), are wnqℓ=c~nk(un−qx)k(vn−qy)νnqℓ, w_nq = c_n\,k(u_n-q_x)\,k(v_n-q_y)\, _nq , (4) qℓ _q =∑n∈qℓwnqℓn∑n∈qℓwnqℓ+ε,k(t)=[1−|t|]+. = _n _q w_nq f_n _n _q w_nq + , k(t)=[1-|t|]_+. Here, k provides bilinear support and νnqℓ _nq encodes relative-depth visibility. The front hypothesis is anchored at the nearest splatted depth zqfrontz_q^front, with νnqfront=exp(−[zn/zqfront−1]+/τ) _nq^front= (-[z_n/z_q^front-1]_+/τ ). The secondary hypothesis retains points satisfying zn>(1+δ)zqfrontz_n>(1+δ)z_q^front, anchors them at their nearest retained depth, and applies the same attenuation. A zero-initialized residual refiner combines both hypotheses with statistics jS_j describing their support, confidence, and relative depth separation: j=jhard+ℛ(jfront,jback,j).G_j=G_j^hard+R\! (G_j^front,G_j^back,S_j ). (5) Thus, the hard route remains the initial and primary condition, while layered evidence supplies learned corrections where useful. Unsupported locations receive no source-specific condition and are completed by the diffusion prior. 3.3 Point-Track Residual Consistency The flow-matching objective in Eq. (2) decomposes over query views, optimizing their individual fidelity. Such per-view supervision does not enforce geometric coherence among the remaining errors: where sparse observations provide weak constraints, predictions of the same physical point can drift differently across jointly generated views even when each appears plausible and incurs a small per-view loss. A cross-view constraint is therefore needed, but directly matching RGB values, VAE latents, or velocities would be overly restrictive because viewpoint-dependent illumination, visibility, and local context legitimately alter these representations. We instead couple each prediction through its residual to its own target, enforcing consistency only in the error that should be removed. For a sampled flow time, the predicted clean latent and its residual in query view j are ^0,j=σ,j−σθ,j,j=^0,j−0,j. Z_0,j=X_σ,j- _θ,j, _j= Z_0,j-Z_0,j. (6) We reuse the VGGT-Ω point predictions to establish tracks across every pair of query views. A point is retained only when it projects inside both views, lies in front of both cameras, and passes a per-view z-buffer visibility test. For a valid track γ, its projections are mapped to the nearest latent-grid locations γ,au_γ,a and γ,bu_γ,b. We define Δγ=a(γ,a)−b(γ,b) _γ=E_a(u_γ,a)-E_b(u_γ,b), and let ρ¯β ρ_β denote the channel average of the scalar Smooth-L1 penalty ρβ _β. With normalized VGGT-Ω confidence wγw_γ, PTRC is ℒPTRC=∑γ∈wγρ¯β(Δγ)∑γ∈wγ,ρβ(x)=x2/(2β),|x|<β,|x|−β/2,|x|≥β. _PTRC= _γ w_γ\, ρ_β( _γ) _γ w_γ, _β(x)= casesx^2/(2β),&|x|<β,\\ |x|-β/2,&|x|≥β. cases (7) The quadratic region provides precise alignment for small discrepancies, while the linear region, together with confidence weighting, limits the influence of unreliable tracks. Since PTRC aligns errors rather than predictions, corresponding views retain valid appearance changes without drifting independently from their respective targets. 3.4 Training and Inference All camera poses are expressed in a gauge defined solely by the source cameras. Consequently, changing the number or subset of query cameras does not alter the conditioning coordinate frame. We freeze the VAE and VGGT-Ω , and fine-tune the diffusion transformer together with the new input and routing modules. The number of jointly generated query views is varied during training to support both single-view synthesis and longer view trajectories. Training loss. The full objective combines query-wise flow matching with point-track regularization as ℒ=ℒFM+λPTRCℒPTRCL=L_FM+ _PTRCL_PTRC, where λPTRC=0.1 _PTRC=0.1 in all experiments. The two terms serve complementary roles: ℒFML_FM learns accurate denoising for each query view, while ℒPTRCL_PTRC couples prediction errors along reliable 3D point tracks. Geometry condition dropout. Geometry reconstructed from sparse observations is informative but inevitably imperfect. During training, we therefore randomly retain, attenuate, or remove the routed geometry condition, while leaving source appearance and camera rays unchanged. This prevents the diffusion model from treating projected geometry as an infallible scene reconstruction and improves its robustness when routed evidence is sparse, uncertain, or locally missing. Geometry-prior CFG. Geometry-condition dropout also provides a reference for selectively strengthening the geometry prior at inference. Let GV_G denote the fully conditioned velocity and refV_ref the reference prediction obtained by removing only query-routed geometry. During early denoising, we apply GeoCFG=ref+sg(G−ref)V_GeoCFG=V_ref+s_g(V_G-V_ref), and otherwise retain GV_G. The two branches share the diffusion state, source appearance, source-aligned visual features, camera conditions, timestep, and empty textual context. Thus, sgs_g strengthens query geometry without removing the observations that define the scene. 4 Experiments 4.1 Experimental Setup Data and training. We train on only the 1K-scene split of DL3DV-10K (960P) (Ling et al., 2024). Removing one unavailable scene and 19 overlapping with the 140-scene DL3DV-Benchmark leaves 980 training scenes. Vggt-Diff initializes from Wan2.1-I2V-14B (Wan et al., 2025). We freeze the VAE and VGGT-Ω (Wang et al., 2026a), while training the Wan DiT, expanded input projection, visual-feature adapter, and Visual Geometry Router. Training runs for 147 epochs at 192×336192×336, then 60 at 480×832480×832, using BF16 and a global batch size of eight. We adopt a 66-to-variable-N curriculum: each sample uses six sources and jointly denoises N targets, progressing from N∈1,2,4N∈\1,2,4\ to N∈4,8,12,16N∈\4,8,12,16\. The Wan DiT uses learning rates of 10−510^-5 and 5×10−65×10^-6 across the two stages; new modules use 10−410^-4 throughout. Additional half-resolution experiments on the full DL3DV-10K yield substantial gains from scaling data and optimization; see the Appendix. Evaluation and baselines. We evaluate 6,188 DL3DV-Benchmark targets and test zero-shot transfer on Mip-NeRF 360 (Barron et al., 2022). All methods receive identical six-view sources and target cameras, with each target generated independently; Vggt-Diff uses 50 flow-sampling steps. Our evaluation follows FrameCrafter (Wu et al., 2026a), with native outputs resized or center-cropped to 480×480480×480. We report PSNR, SSIM (Wang et al., 2004), LPIPS (Zhang et al., 2018), and DreamSim (Fu et al., 2023). Baselines span regression-based and diffusion-based models in the tables. Additional diagnostics use 192 pose-stratified targets and assess eight-view geometric consistency with two frozen reconstructors, VGGT-Ω and Pi3 (Wang et al., 2026b). 4.2 Comparison with Baselines Table 1: Quantitative comparison of 6-view NVS on DL3DV and Mip-NeRF 360. All methods follow the same protocol within each benchmark; #Scenes excludes generic foundation-model pretraining. LPIPS uses AlexNet features, and DS denotes DreamSim. Best and second-best results are highlighted by first and second. Method Backbone #Scenes DL3DV Mip-NeRF 360 PSNR ↑ SSIM ↑ LPIPS ↓ DS ↓ PSNR ↑ SSIM ↑ LPIPS ↓ DS ↓ Regression-based Models AnySplat (Jiang et al., 2025b) VGGT (Wang et al., 2025) 254K 12.388 0.298 0.510 0.214 11.009 0.232 0.594 0.252 E-RayZer (Zhao et al., 2026) – 10K 16.850 0.442 0.455 0.254 16.560 0.343 0.621 0.340 LVSM (Jin et al., 2024) – 67.5K 17.090 0.478 0.333 0.204 15.250 0.317 0.609 0.577 DepthSplat (Xu et al., 2025) Depth Anything V2 (Yang et al., 2024) 77.5K 17.284 0.566 0.307 0.158 15.977 0.370 0.410 0.215 Diffusion-based Models EscherNet (Kong et al., 2024) SD1.5 (Rombach et al., 2022) 10K 12.070 0.251 0.484 0.227 11.140 0.126 0.540 0.315 Aether (Zhu et al., 2025) CogVideoX-5B (Yang et al., 2025) – 12.660 0.258 0.469 0.140 12.600 0.220 0.651 0.334 MVSplat360 (Chen et al., 2024b) SVD (Blattmann et al., 2023) 69.5K 14.150 0.358 0.513 0.174 13.859 0.292 0.636 0.278 SEVA (Zhou et al., 2025) SD2.1 (Rombach et al., 2022) 80K 16.150 0.470 0.253 0.088 14.590 0.294 0.372 0.137 FrameCrafter (Wu et al., 2026a) Wan2.1-I2V-14B (Wan et al., 2025) 1K 17.180 0.445 0.223 0.066 15.640 0.279 0.365 0.111 Vggt-Diff Wan2.1-I2V-14B (Wan et al., 2025) 1K 18.104 0.508 0.222 0.062 16.296 0.317 0.315 0.091 (1) Overall performance. Table 1 compares all methods under the common protocol. On 6,188 DL3DV targets, Vggt-Diff ranks first in PSNR, LPIPS, and DreamSim and second in SSIM. Notably, these results are achieved using only ∼ 1k training scenes, substantially fewer than many competing methods, demonstrating strong data efficiency. Against FrameCrafter with the same Wan2.1-I2V-14B prior and a comparable training scale, Vggt-Diff gains 0.9240.924 dB PSNR and 0.0630.063 SSIM while retaining comparable perceptual quality, demonstrating the benefit of geometry-routed adaptation. On zero-shot Mip-NeRF 360, Vggt-Diff ranks first in LPIPS and DreamSim and second in PSNR. Although E-RayZer and DepthSplat lead PSNR and SSIM, respectively, their substantially higher perceptual errors indicate that these gains come at the cost of perceptual quality. (2) Qualitative comparison. Fig. 4 and Fig. 5 localize these gains. On DL3DV, Vggt-Diff preserves chair legs, desk boundaries, facade geometry, and legible signage while recovering target layout and occlusion order; competing methods blur thin structures, distort geometry, or drift from the requested camera. On zero-shot Mip-NeRF 360, Vggt-Diff likewise better preserves object boundaries and source-observed appearance, indicating scene-grounded rather than unsupported synthesis. Figure 4: Qualitative comparison on diverse DL3DV scenes. Vggt-Diff better matches the requested viewpoints while preserving fine geometry, legible text, and object identity; previous methods exhibit blur, structural distortion, or content drift. Figure 5: Qualitative comparison on Mip-NeRF 360. Without dataset-specific fine-tuning, Vggt-Diff better preserves fine structures, occlusion boundaries, and source-observed appearance while more accurately matching the target view. Table 2: PSNR across target-pose difficulty on DL3DV. All methods use independent 66-to-11 inference over 32 targets per bin. Method Interpolation Extrapolation Near Mid Far Near Mid Far Regression-based Models AnySplat (Jiang et al., 2025b) 14.196 12.415 10.862 13.977 12.566 11.827 E-RayZer (Zhao et al., 2026) 17.388 14.878 13.824 16.833 14.477 14.712 LVSM (Jin et al., 2024) 20.276 16.599 13.969 20.790 16.633 15.579 DepthSplat (Xu et al., 2025) 20.376 17.424 14.820 20.653 17.943 16.374 Diffusion-based Models Aether (Zhu et al., 2025) 14.030 12.167 11.065 15.739 12.295 11.795 GEN3C (Ren et al., 2025) 13.620 12.356 11.548 14.288 13.050 12.288 MVSplat360 (Chen et al., 2024b) 15.282 14.420 13.389 15.542 14.479 14.176 FrameCrafter (Wu et al., 2026a) 20.338 17.485 14.649 21.168 17.888 16.594 SEVA (Zhou et al., 2025) 20.940 17.861 14.703 22.311 18.050 16.575 Vggt-Diff 20.920 18.391 15.853 22.396 18.793 17.793 Table 3: Ablations on DL3DV. All variants use independent 66-to-11 inference on the 192 pose-stratified targets; LPIPS uses VGG features. Configuration PSNR ↑ SSIM ↑ LPIPS ↓ Model design w/o VGGT-Ω prior 17.294 0.501 0.357 w/ point-rendered RGB 18.553 0.536 0.316 w/o PTRC 18.586 0.530 0.312 w/o layered residual refinement 18.895 0.548 0.303 Geometry-conditioning strategy w/o condition regularization 18.722 0.538 0.309 w/o CFG (full method) 19.024 0.557 0.301 w/ geometry-prior CFG 19.139 0.563 0.296 (3) Performance across pose difficulties. To expose failures hidden by aggregate metrics, we stratify 192 DL3DV targets by interpolation/extrapolation and near/mid/far difficulty (32 per bin). Table 3 shows that Vggt-Diff leads five bins and trails SEVA by only 0.0200.020 dB in interpolation-near. Its margin over FrameCrafter ranges from 0.5820.582 to 1.2281.228 dB and exceeds 0.90.9 dB in five bins, confirming broad effectiveness across pose difficulties. (4) Cross-view consistency. Single-target metrics cannot determine whether jointly generated views support a coherent scene. We therefore reconstruct the same eight views with two frozen geometry models: VGGT-Ω as the primary probe and Pi3, unused in training or conditioning, as an independent check for shared-backbone evaluator bias. If generated views are mutually consistent and match GT content and detail, each model should recover a coherent point cloud close to its GT-view reconstruction. Across 52 DL3DV scenes, all methods use the same six sources and eight target cameras, with generated and GT views processed identically by each reconstructor. Table 4 shows that Vggt-Diff ranks first under both models. Relative to FrameCrafter, it reduces Chamfer-L1 by 26.7%26.7\% with VGGT-Ω and 16.4%16.4\% with Pi3, while improving F-score@1% and F-score@2% by 11.4511.45/12.4812.48 and 9.499.49/11.4111.41 percentage points, respectively. This agreement indicates that the gains do not arise solely from reusing VGGT-Ω as the evaluator. Appendix Fig. 10 provides corresponding visualizations. Since the references are reconstructed from target RGB rather than physical scans, these metrics measure relative consistency and reconstructability, not absolute 3D accuracy. Table 4: Cross-reconstructor evaluation on DL3DV. Eight jointly generated views are reconstructed with frozen VGGT-Ω and Pi3 and compared with GT-view references produced by the same reconstructor. Pi3 is not used for training or conditioning Vggt-Diff. Metrics are trajectory-normalized and should be compared within each reconstructor. Method VGGT-Ω (Wang et al., 2026a) Pi3 (Wang et al., 2026b) Chamfer ↓ F@1% ↑ F@2% ↑ Chamfer ↓ F@1% ↑ F@2% ↑ LVSM (Jin et al., 2024) 0.2387 0.1116 0.2218 0.3610 0.0673 0.1525 SEVA (Zhou et al., 2025) 0.1944 0.2050 0.3244 0.2514 0.1640 0.2751 FrameCrafter (Wu et al., 2026a) 0.1000 0.3086 0.4793 0.1404 0.2140 0.3880 Vggt-Diff 0.0733 0.4231 0.6041 0.1174 0.3089 0.5021 4.3 Ablation Studies (1) Protocol. All variants reported in Table 3 follow the same full-resolution controlled protocol and use independent 66-to-11 inference on 192 pose-stratified targets. Training-time variants share the same Wan initialization, change only the listed component, and are evaluated without Geometry-Prior CFG. The final row applies CFG to the full checkpoint and changes only inference. Unless stated otherwise, differences are measured against the full checkpoint without CFG. (2) Model design. Table 3 shows that removing all VGGT-Ω -dependent components causes the largest degradation: PSNR falls by 1.7301.730 dB and SSIM by 0.0560.056, while LPIPS rises by 0.0560.056. Camera rays and the Wan prior alone therefore cannot replace visual geometry. Replacing routed features with point-rendered RGB, while retaining the geometry, router, and PTRC, costs 0.4710.471 dB PSNR and raises LPIPS by 0.0150.015, confirming that appearance-bearing features provide richer evidence than incomplete RGB renderings. Removing PTRC costs 0.4380.438 dB PSNR and 0.0270.027 SSIM, the largest SSIM loss among single-component ablations, supporting residual alignment over shared 3D tracks. Removing layered refinement while retaining the hard anchor causes smaller drops of 0.1290.129 dB PSNR and 0.0090.009 SSIM, indicating that hard routing captures most of the benefit while layered refinement corrects ambiguous locations near occlusions and imperfect projections. (3) Geometry-conditioning strategy. Removing condition regularization costs 0.3020.302 dB PSNR and 0.0190.019 SSIM while increasing LPIPS by 0.0080.008, confirming that stochastic attenuation and dropout reduce over-reliance on imperfect geometry. The no-CFG row evaluates the full checkpoint without guidance; applying Geometry-Prior CFG to the same checkpoint improves PSNR by 0.1150.115 dB and SSIM by 0.0060.006 while reducing LPIPS by 0.0050.005. The larger gain from regularization identifies it as the primary mechanism, while matched CFG provides a modest complementary improvement at inference without retraining or changing the learned model parameters or training objective. Additional experiments, protocol details, results and analyses are provided in the Appendix. 5 Conclusion We presented Vggt-Diff, a geometry-routed multi-view framework combining visual geometry with pretrained video diffusion for sparse-view NVS. Its confidence-aware VGR maps source features into query-aligned conditions using 3D points and confidence, while PTRC aligns predicted-clean residuals along reliable correspondences. Stochastic training-time weakening and selective inference-time strengthening improve robustness to imperfect geometry. Across diverse pose difficulties, Vggt-Diff achieves strong target-view fidelity and cross-view coherence, combining geometry-grounded scene structure with the generative prior’s ability to complete unseen content. AI use statement We used generative AI tools to assist with language polishing, improving clarity and concision, and refining overleaf formatting and presentation. We also used these tools to support literature discovery and track recent developments relevant to our work. All suggested references were manually verified against their original sources, and all AI-assisted text and formatting changes were reviewed and revised by us. Generative AI was not used to conduct experiments, or determine the scientific claims and conclusions of this work. We take full responsibility for the final content, including all text, citations, and artifacts produced with the aid of generative AI. Reproducibility statement We provide detailed data construction, view sampling, training schedules, conditioning hyperparameters, evaluation protocols, and additional controlled experiments in the Appendix. All reported comparisons use fixed source views, target cameras, evaluation cases, and metric backbones as specified in the paper. We will release the training and evaluation code, model checkpoints, and configuration files to support reproduction of our results. References Barron et al. (2021) J. T. Barron, B. Mildenhall, M. Tancik, P. Hedman, R. Martin-Brualla, and P. P. Srinivasan Mip-nerf: a multiscale representation for anti-aliasing neural radiance fields. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), p. 5835–5844. Cited by: §2. Barron et al. (2022) J. T. Barron, B. Mildenhall, D. Verbin, P. P. Srinivasan, and P. Hedman Mip-nerf 360: unbounded anti-aliased neural radiance fields. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 5460–5469. Cited by: §4.1. Blattmann et al. (2023) A. Blattmann, T. Dockhorn, S. Kulal, D. Mendelevitch, M. Kilian, D. Lorenz, Y. Levi, Z. English, V. Voleti, A. Letts, et al. Stable video diffusion: scaling latent video diffusion models to large datasets. arXiv preprint arXiv:2311.15127. Cited by: Table 1. Charatan et al. (2024) D. Charatan, S. L. Li, A. Tagliasacchi, and V. Sitzmann Pixelsplat: 3d gaussian splats from image pairs for scalable generalizable 3d reconstruction. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 19457–19467. Cited by: §1, §2. Chen et al. (2021) A. Chen, Z. Xu, F. Zhao, X. Zhang, F. Xiang, J. Yu, and H. Su Mvsnerf: fast generalizable radiance field reconstruction from multi-view stereo. In 2021 IEEE/CVF international conference on computer vision (ICCV), p. 14104–14113. Cited by: §2. Chen et al. (2025) K. Chen, B. Dai, M. Qin, D. Zhang, P. Li, Y. Zou, and H. Wang Slgaussian: fast language gaussian splatting in sparse views. In Proceedings of the 33rd ACM International Conference on Multimedia, p. 3047–3056. Cited by: §2. Chen et al. (2026) K. Chen, Y. Zhong, Z. Li, J. Lin, Y. Chen, M. Qin, and H. Wang Quantifying and alleviating co-adaptation in sparse-view 3d gaussian splatting. Advances in Neural Information Processing Systems 38, p. 115939–115968. Cited by: §2. Chen et al. (2024a) Y. Chen, H. Xu, C. Zheng, B. Zhuang, M. Pollefeys, A. Geiger, T. Cham, and J. Cai Mvsplat: efficient 3d gaussian splatting from sparse multi-view images. In European conference on computer vision, p. 370–386. Cited by: §1, §2. Chen et al. (2024b) Y. Chen, C. Zheng, H. Xu, B. Zhuang, A. Vedaldi, T. Cham, and J. Cai Mvsplat360: feed-forward 360 scene synthesis from sparse views. Advances in Neural Information Processing Systems 37, p. 107064–107086. Cited by: Table 10, Table 9, §2, Table 1, Table 3. Elata et al. (2025) N. Elata, B. Kawar, Y. Ostrovsky-Berman, M. Farber, and R. Sokolovsky Novel view synthesis with pixel-space diffusion models. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 26756–26766. Cited by: §2. Fu et al. (2023) S. Fu, N. Tamir, S. Sundaram, L. Chai, R. Zhang, T. Dekel, and P. Isola Dreamsim: learning new dimensions of human visual similarity using synthetic data. arXiv preprint arXiv:2306.09344. Cited by: §4.1. Hirschorn et al. (2026) O. Hirschorn, O. Sela, I. Huberman-Spiegelglas, N. Efrat, E. Alshan, I. Ideses, F. Devernay, Y. Zvik, and L. Fritz Splatent: splatting diffusion latents for novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 8319–8330. Cited by: §2. Hong et al. (2024) Y. Hong, K. Zhang, J. Gu, S. Bi, Y. Zhou, D. Liu, F. Liu, K. Sunkavalli, T. Bui, and H. Tan Lrm: large reconstruction model for single image to 3d. In International Conference on Learning Representations, Vol. 2024, p. 50678–50702. Cited by: §2. Jia et al. (2026) X. Jia, Y. Sun, J. You, Z. Zou, J. Yan, Z. Wu, Y. Jiang, et al. Efficient-lvsm: faster, cheaper, and better large view synthesis model via decoupled co-refinement attention. In International Conference on Learning Representations, Vol. 2026, p. 40029–40046. Cited by: §2. Jiang et al. (2025a) H. Jiang, H. Tan, P. Wang, H. Jin, Y. Zhao, S. Bi, K. Zhang, F. Luan, K. Sunkavalli, Q. Huang, et al. Rayzer: a self-supervised large view synthesis model. In 2025 IEEE/CVF International Conference on Computer Vision (ICCV), p. 4918–4929. Cited by: §1, §2. Jiang et al. (2025b) L. Jiang, Y. Mao, L. Xu, T. Lu, K. Ren, Y. Jin, X. Xu, M. Yu, J. Pang, F. Zhao, et al. Anysplat: feed-forward 3d gaussian splatting from unconstrained views. ACM Transactions on Graphics (TOG) 44 (6), p. 1–16. Cited by: Table 10, Table 9, Table 1, Table 3. Jin et al. (2024) H. Jin, H. Jiang, H. Tan, K. Zhang, S. Bi, T. Zhang, F. Luan, N. Snavely, and Z. Xu Lvsm: a large view synthesis model with minimal 3d inductive bias. arXiv preprint arXiv:2410.17242. Cited by: Table 10, Table 9, §1, §2, Table 1, Table 3, Table 4. Kang et al. (2026) M. Kang, I. Shin, T. Lee, M. Kim, I. S. Kweon, and K. Yoon GeoNVS: geometry grounded video diffusion for novel view synthesis. arXiv preprint arXiv:2603.14965. Cited by: §2. Kerbl et al. (2023) B. Kerbl, G. Kopanas, T. Leimkühler, G. Drettakis, et al. 3d gaussian splatting for real-time radiance field rendering.. ACM Trans. Graph. 42 (4), p. 139–1. Cited by: §1, §2. Kim et al. (2026) E. Kim, H. Ryu, T. W. Mitchel, and V. Sitzmann Scaling view synthesis transformers. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 28893–28902. Cited by: §1, §2. Kong et al. (2024) X. Kong, S. Liu, X. Lyu, M. Taher, X. Qi, and A. J. Davison Eschernet: a generative model for scalable view synthesis. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 9503–9513. Cited by: §1, §2, Table 1. Kulhánek et al. (2022) J. Kulhánek, E. Derner, T. Sattler, and R. Babuška Viewformer: nerf-free neural rendering from few images using transformers. In European Conference on Computer Vision, p. 198–216. Cited by: §2. Kwak et al. (2024) J. Kwak, E. Dong, Y. Jin, H. Ko, S. Mahajan, and K. M. Yi Vivid-1-to-3: novel view synthesis with video diffusion models. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 6775–6785. Cited by: §2. Kwak et al. (2026) M. Kwak, J. Kim, S. Yun, D. Han, T. Kim, S. Kim, and J. Kim Aligned novel view image and geometry synthesis via cross-modal attention instillation. In International Conference on Learning Representations, Vol. 2026, p. 16029–16050. Cited by: §2. Kwon et al. (2026) M. Kwon, J. Choi, J. Park, S. Jeon, J. Jang, J. Seo, M. Kwak, J. Kim, and S. Kim Correspondence-attention alignment for multi-view diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 2316–2326. Cited by: §2. Le et al. (2026) T. Le, T. Pham, T. Nguyen, D. Kong, X. Xie, and S. Mandt Umami: unifying masked autoregressive models and deterministic rendering for view synthesis. Advances in Neural Information Processing Systems 38, p. 175038–175065. Cited by: §2. Leroy et al. (2024) V. Leroy, Y. Cabon, and J. Revaud Grounding image matching in 3d with mast3r. In European conference on computer vision, p. 71–91. Cited by: §2. Ling et al. (2024) L. Ling, Y. Sheng, Z. Tu, W. Zhao, C. Xin, K. Wan, L. Yu, Q. Guo, Z. Yu, Y. Lu, et al. Dl3dv-10k: a large-scale scene dataset for deep learning-based 3d vision. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 22160–22169. Cited by: §A.1, §4.1. Liu et al. (2023) R. Liu, R. Wu, B. Van Hoorick, P. Tokmakov, S. Zakharov, and C. Vondrick Zero-1-to-3: zero-shot one image to 3d object. In 2023 IEEE/CVF International Conference on Computer Vision (ICCV), p. 9264–9275. Cited by: §2. Liu et al. (2024) Y. Liu, C. Lin, Z. Zeng, X. Long, L. Liu, T. Komura, and W. Wang Syncdreamer: generating multiview-consistent images from a single-view image. In International conference on learning representations, Vol. 2024, p. 27676–27697. Cited by: §2. Mescheder et al. (2026) L. Mescheder, W. Dong, S. Li, X. Bai, M. Santos, P. Hu, B. Lecouat, M. Zhen, A. Delaunoy, T. Fang, et al. Sharp monocular view synthesis in less than a second. In International Conference on Learning Representations, Vol. 2026, p. 24192–24230. Cited by: §2. Mildenhall et al. (2021) B. Mildenhall, P. P. Srinivasan, M. Tancik, J. T. Barron, R. Ramamoorthi, and R. Ng Nerf: representing scenes as neural radiance fields for view synthesis. Communications of the ACM 65 (1), p. 99–106. Cited by: §1, §2. Müller et al. (2024) N. Müller, K. Schwarz, B. Rössle, L. Porzi, S. R. Bulò, M. Nießner, and P. Kontschieder Multidiff: consistent novel view synthesis from a single image. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 10258–10268. Cited by: §1, §2. Ren et al. (2025) X. Ren, T. Shen, J. Huang, H. Ling, Y. Lu, M. Nimier-David, T. Müller, A. Keller, S. Fidler, and J. Gao Gen3c: 3d-informed world-consistent video generation with precise camera control. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 6121–6132. Cited by: Table 10, Table 9, Table 3. Rombach et al. (2022) R. Rombach, A. Blattmann, D. Lorenz, P. Esser, and B. Ommer High-resolution image synthesis with latent diffusion models. In 2022 IEEE/CVF conference on computer vision and pattern recognition (CVPR), p. 10674–10685. Cited by: Table 1, Table 1. Sajjadi et al. (2022) M. S. Sajjadi, H. Meyer, E. Pot, U. Bergmann, K. Greff, N. Radwan, S. Vora, M. Lučić, D. Duckworth, A. Dosovitskiy, et al. Scene representation transformer: geometry-free novel view synthesis through set-latent scene representations. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 6219–6228. Cited by: §2. Shi et al. (2023) R. Shi, H. Chen, Z. Zhang, M. Liu, C. Xu, X. Wei, L. Chen, C. Zeng, and H. Su Zero123++: a single image to consistent multi-view diffusion base model. arXiv preprint arXiv:2310.15110. Cited by: §2. Shi et al. (2024) Y. Shi, P. Wang, J. Ye, L. Mai, K. Li, and X. Yang Mvdream: multi-view diffusion for 3d generation. In International conference on learning representations, Vol. 2024, p. 39838–39859. Cited by: §2. Szymanowicz et al. (2026) S. Szymanowicz, M. Chen, J. Wang, C. Rupprecht, and A. Vedaldi LagerNVS: latent geometry for fully neural real-time novel view synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 15443–15453. Cited by: §1, §2. Szymanowicz et al. (2024) S. Szymanowicz, C. Rupprecht, and A. Vedaldi Splatter image: ultra-fast single-view 3d reconstruction. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 10208–10217. Cited by: §2. Van Hoorick et al. (2026) B. Van Hoorick, D. Chen, S. Iwase, P. Tokmakov, M. Z. Irshad, I. Vasiljevic, S. Gupta, F. Cheng, S. Zakharov, and V. C. Guizilini Anyview: synthesizing any novel view in dynamic scenes. arXiv preprint arXiv:2601.16982. Cited by: §2. Voleti et al. (2024) V. Voleti, C. Yao, M. Boss, A. Letts, D. Pankratz, D. Tochilkin, C. Laforte, R. Rombach, and V. Jampani Sv3d: novel multi-view synthesis and 3d generation from a single image using latent video diffusion. In European Conference on Computer Vision, p. 439–457. Cited by: §2. Wan et al. (2025) T. Wan, A. Wang, B. Ai, B. Wen, C. Mao, C. Xie, D. Chen, F. Yu, H. Zhao, J. Yang, et al. Wan: open and advanced large-scale video generative models. arXiv preprint arXiv:2503.20314. Cited by: §A.2, §3.1, §4.1, Table 1, Table 1. Wang et al. (2025) J. Wang, M. Chen, N. Karaev, A. Vedaldi, C. Rupprecht, and D. Novotny Vggt: visual geometry grounded transformer. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 5294–5306. Cited by: §1, §2, Table 1. Wang et al. (2026a) J. Wang, M. Chen, S. Zhang, N. Karaev, J. Schönberger, P. Labatut, P. Bojanowski, D. Novotny, A. Vedaldi, and C. Rupprecht VGGT-omega. arXiv preprint arXiv:2605.15195. Cited by: Appendix A, §1, §2, §3.2, §4.1, Table 4. Wang et al. (2022) P. Wang, X. Chen, T. Chen, S. Venugopalan, Z. Wang, et al. Is attention all that nerf needs?. arXiv preprint arXiv:2207.13298. Cited by: §2. Wang et al. (2021) Q. Wang, Z. Wang, K. Genova, P. P. Srinivasan, H. Zhou, J. T. Barron, R. Martin-Brualla, N. Snavely, and T. Funkhouser Ibrnet: learning multi-view image-based rendering. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, p. 4690–4699. Cited by: §2. Wang et al. (2024a) Q. Wang, Y. Zhao, J. Ma, and J. Li How to use diffusion priors under sparse views?. Advances in Neural Information Processing Systems 37, p. 30394–30424. Cited by: §2. Wang et al. (2024b) S. Wang, V. Leroy, Y. Cabon, B. Chidlovskii, and J. Revaud Dust3r: geometric 3d vision made easy. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 20697–20709. Cited by: §2. Wang et al. (2026b) Y. Wang, J. Zhou, H. Zhu, W. Chang, Y. Zhou, Z. Li, J. Chen, J. Pang, C. Shen, and T. He π3π^3: permutation-equivariant visual geometry learning. In International Conference on Learning Representations, Vol. 2026, p. 10481–10497. Cited by: §A.7, §4.1, Table 4. Wang et al. (2004) Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli Image quality assessment: from error visibility to structural similarity. IEEE transactions on image processing 13 (4), p. 600–612. Cited by: §4.1. Watson et al. (2022) D. Watson, W. Chan, R. Martin-Brualla, J. Ho, A. Tagliasacchi, and M. Norouzi Novel view synthesis with diffusion models. arXiv preprint arXiv:2210.04628. Cited by: §1, §2. Wu et al. (2026a) Q. Wu, K. Vuong, M. Jeon, S. Narasimhan, and D. Ramanan Novel view synthesis as video completion. arXiv preprint arXiv:2604.08500. Cited by: §A.1, Table 10, Table 13, Table 13, Table 9, §1, §2, §4.1, Table 1, Table 3, Table 4. Wu et al. (2026b) Z. Wu, Z. Jiang, M. R. Oswald, and J. Song From rays to projections: better inputs for feed-forward view synthesis. arXiv preprint arXiv:2601.05116. Cited by: §2. Xu et al. (2025) H. Xu, S. Peng, F. Wang, H. Blum, D. Barath, A. Geiger, and M. Pollefeys Depthsplat: connecting gaussian splatting and depth. In 2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 16453–16463. Cited by: Table 10, Table 9, Table 1, Table 3. Xu et al. (2026) Y. Xu, Y. Wang, and S. X. Yu Novel view synthesis from a few glimpses via test-time natural video completion. Advances in Neural Information Processing Systems 38, p. 46541–46561. Cited by: §2. Yang et al. (2024) L. Yang, B. Kang, Z. Huang, Z. Zhao, X. Xu, J. Feng, and H. Zhao Depth anything v2. Advances in neural information processing systems 37, p. 21875–21911. Cited by: Table 1. Yang et al. (2025) Z. Yang, J. Teng, W. Zheng, M. Ding, S. Huang, J. Xu, Y. Yang, W. Hong, X. Zhang, G. Feng, et al. Cogvideox: text-to-video diffusion models with an expert transformer. In International Conference on Learning Representations, Vol. 2025, p. 83048–83077. Cited by: Table 1. Ye et al. (2024) J. Ye, P. Wang, K. Li, Y. Shi, and H. Wang Consistent-1-to-3: consistent image to 3d view synthesis via geometry-aware diffusion models. In 2024 International Conference on 3D Vision (3DV), p. 664–674. Cited by: §2. Yu et al. (2021) A. Yu, V. Ye, M. Tancik, and A. Kanazawa Pixelnerf: neural radiance fields from one or few images. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 4576–4585. Cited by: §2. Yu et al. (2024) W. Yu, J. Xing, L. Yuan, W. Hu, X. Li, Z. Huang, X. Gao, T. Wong, Y. Shan, and Y. Tian Viewcrafter: taming video diffusion models for high-fidelity novel view synthesis. arXiv preprint arXiv:2409.02048. Cited by: §1, §2. Zhang et al. (2024) K. Zhang, S. Bi, H. Tan, Y. Xiangli, N. Zhao, K. Sunkavalli, and Z. Xu Gs-lrm: large reconstruction model for 3d gaussian splatting. In European Conference on Computer Vision, p. 1–19. Cited by: §2. Zhang et al. (2018) R. Zhang, P. Isola, A. A. Efros, E. Shechtman, and O. Wang The unreasonable effectiveness of deep features as a perceptual metric. In 2018 IEEE/CVF conference on computer vision and pattern recognition, p. 586–595. Cited by: §4.1. Zhao et al. (2026) Q. Zhao, H. Tan, Q. Wang, S. Bi, K. Zhang, K. Sunkavalli, S. Tulsiani, and H. Jiang E-rayzer: self-supervised 3d reconstruction as spatial visual pre-training. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, p. 7525–7535. Cited by: Table 10, Table 9, Table 1, Table 3. Zheng and Vedaldi (2024) C. Zheng and A. Vedaldi Free3d: consistent novel view synthesis without 3d representation. In 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), p. 9720–9731. Cited by: §1, §2. Zhou et al. (2026) H. Zhou, W. Yu, C. Feng, X. Zhou, Y. Tian, and L. Yuan UniWorld-view: large-baseline view synthesis via video diffusion models. arXiv preprint arXiv:2608.04701. Cited by: §2. Zhou et al. (2025) J. Zhou, H. Gao, V. Voleti, A. Vasishta, C. Yao, M. Boss, P. Torr, C. Rupprecht, and V. Jampani Stable virtual camera: generative view synthesis with diffusion models. In 2025 IEEE/CVF International Conference on Computer Vision (ICCV), p. 12405–12414. Cited by: §A.7, Table 10, Table 9, §1, §2, Table 1, Table 3, Table 4. Zhu et al. (2025) H. Zhu, Y. Wang, J. Zhou, W. Chang, Y. Zhou, Z. Li, J. Chen, C. Shen, J. Pang, and T. He Aether: geometric-aware unified world modeling. In 2025 IEEE/CVF International Conference on Computer Vision (ICCV), p. 8535–8546. Cited by: Table 10, Table 9, Table 1, Table 3. Appendix A Appendix Metric convention. The cross-method comparison in Table 1 uses LPIPS-AlexNet to follow the evaluation protocol of prior work. Controlled ablations use LPIPS-VGG consistently within the fixed 192-case diagnostic protocol. We label the backbone explicitly and do not compare absolute LPIPS values across these protocols. Appendix organization. The supplementary material is organized as follows. Secs. A.1 and A.2. We specify the view-sampling protocol, camera gauge, pose-difficulty definition, and fixed training and conditioning hyperparameters. Secs. A.3, A.4, and A.5. We analyze data and optimization scaling, joint multi-target generation, and the feature depth and optimization scope of VGGT-Ω (Wang et al., 2026a). Secs. A.6 and A.7. We report complete pose-difficulty results and evaluate cross-view consistency using two independent point-cloud reconstructors. Secs. A.8 and A.9. We isolate the residual formulation of PTRC and provide additional optimization and scene-level statistical diagnostics. Secs. A.10 and A.11. We examine robustness to fewer source views and clarify the VAE protocols used for quantitative evaluation and continuous-trajectory generation. Sec. A.12. We conclude with promising directions for extending geometry-routed multi-view generation. A.1 View Sampling and Pose-Difficulty Protocol Training view sampling. For each training scene, we precompute 16 reproducible six-view source bundles that rotate deterministically across epochs. Thirteen bundles use global sampling: the ordered camera trajectory is divided into six contiguous segments, and one frame is sampled from each segment. The remaining three use the same stratification within a randomly selected local window. The six selected views are shuffled before entering the model, exposing it to both broad trajectory coverage and locally concentrated observations. Targets are sampled online after excluding the active source frames. Local bundles preferentially draw targets from their window and fall back to the full scene when necessary. The target-count curriculum progresses through 1,2,4→2,4,8→4,8,12→4,8,12,16\1,2,4\\!→\!\2,4,8\\!→\!\4,8,12\\!→\!\4,8,12,16\ in the first training stage and 4,8→4,8,12→4,8,12,16\4,8\\!→\!\4,8,12\\!→\!\4,8,12,16\ in the second. For eight or more targets, we always sample an ordered trajectory by selecting evenly spaced frames within a random temporal span. Smaller target sets use this mode with probability 0.50.5 and otherwise sample views uniformly without replacement. Training does not explicitly balance the pose-difficulty bins used for evaluation. Source-defined camera gauge. Camera normalization depends only on the six source views. We choose the source camera nearest their centroid as the anchor, align all cameras to its rotation and origin, and normalize translation by the mean source-to-anchor distance. Query cameras do not affect this transformation, so single-target and joint inference share the same coordinate frame for a fixed source set. Pose-difficulty definition. We construct the diagnostic set from the official six-view DL3DV-Benchmark (Ling et al., 2024) split, using its fixed source views and candidate test frames. Let imini_ and imaxi_ be the minimum and maximum source frame indices. A target is trajectory interpolation when imin<it<imaxi_ <i_t<i_ and trajectory extrapolation when it lies outside this interval. This definition follows the ordered capture trajectory and does not imply containment within the 3D convex hull of the source cameras. Difficulty is measured by the target’s distance from its nearest source pose. Let t,tc_t,f_t denote the target camera center and unit forward direction, and j,jc_j,f_j those of source j. We define the source-trajectory diameter as D=maxj,k∈∥j−k∥2D_S= _j,k _j-c_k _2 and its stabilized value as D¯=max(D,10−6) D_S= (D_S,10^-6). The angular difference to source j is θj=arccos[clip(t⊤j,−1,1)] _j= [clip(f_t f_j,-1,1) ]. Pose novelty is then νt=minj∈∥t−j∥22D¯2+θj2π2. _t= _j _t-c_j _2^2 D_S^2+ _j^2π^2. (8) This combines scene-normalized translation and viewing-direction change. Near, mid, and far bins are defined by global pose-novelty tertiles computed separately for interpolation and extrapolation. Their two thresholds are (0.1165,0.2147)(0.1165,0.2147) for interpolation and (0.0560,0.1272)(0.0560,0.1272) for extrapolation. Figure 6 visualizes the two complementary criteria. The row depends on the target’s position along the ordered capture trajectory, whereas the column reflects its translation-and-rotation novelty νt _t. A target can therefore be extrapolative but pose-near, or interpolative yet pose-far. Figure 6: Pose-difficulty protocol and representative predictions. Blue frustums denote the six source cameras and purple denotes the query. Rows distinguish trajectory interpolation from extrapolation; columns increase pose novelty from near to far. Frozen VGGT-Ω point clouds, fused from multiple continuous GT-view windows, provide scene context only. Insets compare Vggt-Diff with GT and do not determine the bins. Frozen evaluation cases. The candidate pool contains 5,549 interpolation and 638 extrapolation targets. From each of the six bins, we deterministically select 32 targets from 32 different scenes. The resulting protocol contains 192 unique targets from 109 scenes; scenes may recur across bins, but no scene-target pair is duplicated. Every method receives the same six source views and predicts each target independently using fixed per-case noise, without additional target slots, rollout, or prediction feedback. This pose-stratified diagnostic is separate from the aggregate FrameCrafter-aligned (Wu et al., 2026a) comparison in Table 1, and its cases remain fixed across model resolutions. A.2 Fixed Hyperparameters and Conditioning Regularization Table 5 collects the constants shared by the main experiments. The router thresholds operate on relative depth and are therefore dimensionless; PTRC is evaluated directly in the Wan (Wan et al., 2025) latent space. Visual Geometry Router. We retain the notation of Eq. 4. The raw confidence cnc_n is normalized per scene as c~n=clip(cn/maxmcm,0,1) c_n=clip(c_n/ _mc_m,0,1). For query-grid cell q, the front anchor is the nearest splatted depth zqfrontz_q^front, and νnqfront=exp[−(zn/zqfront−1)+/τ] _nq^front= [-(z_n/z_q^front-1)_+/τ], with τ=0.02τ=0.02. The back (secondary) hypothesis retains points satisfying zn>(1+δ)zqfrontz_n>(1+δ)z_q^front, where δ=0.04δ=0.04; their nearest depth defines zqbackz_q^back, around which the same attenuation is applied. Besides qfrontG_q^front and qbackG_q^back, the router outputs four per-cell statistics: clipped front and back support, bilinear-weighted mean confidence, and the clipped relative gap (zqback/zqfront−1)(z_q^back/z_q^front-1). PTRC optimization. Following Eq. 6, query view j uses ^0,j=σ,j−σθ,j Z_0,j=X_σ,j- _θ,j and residual j=^0,j−0,jE_j= Z_0,j-Z_0,j. We compute Eq. 7 in the Wan latent space, using the confidence-weighted Smooth-L1 penalty with β=0.1β=0.1 and no VAE decoding. Within each training stage, utr=e/(E−1)u_tr=e/(E-1) denotes normalized epoch progress and λeff(utr)=λPTRC0,utr<0.1,(utr−0.1)/0.2,0.1≤utr<0.3,1,utr≥0.3,λPTRC=0.1. _eff(u_tr)= _PTRC cases0,&u_tr<0.1,\\ (u_tr-0.1)/0.2,&0.1≤ u_tr<0.3,\\ 1,&u_tr≥ 0.3, cases _PTRC=0.1. (9) Thus, PTRC is disabled for the first 10%10\% of each stage, linearly activated during the next 20%20\%, and fully weighted thereafter. The latent-space implementation uses no additional timestep gate. Geometry-condition regularization. Each training sample draws one shared geometry-condition scale a∼0,Pr=0.10,(0.2,0.8),Pr=0.20,1,Pr=0.70.a cases0,& =0.10,\\ U(0.2,0.8),& =0.20,\\ 1,& =0.70. cases (10) The same a scales the global Ω context and dense routed geometry over both source and query slots. Source-image latents, Plücker rays, diffusion state, and timestep remain unchanged; this operation is not text dropout. Geometry-Prior CFG. The formal guided result uses a target-only reference branch. The conditional prediction GV_G retains all routed geometry, whereas refV_ref zeros only query-slot dense geometry while preserving source-aligned geometry, global Ω context, image latents, Plücker rays, diffusion state, timestep, and empty-text context. For denoising step k, with uden=k/(K−1)u_den=k/(K-1), we use GeoCFG=ref+sgeff(uden)(G−ref),sgeff(uden)=1+(sg−1)cos2[π2min(uden0.6,1)].V_GeoCFG=V_ref+s_g^eff(u_den)(V_G-V_ref), s_g^eff(u_den)=1+(s_g-1) ^2\! [ π2 \! ( u_den0.6,1 ) ]. (11) Hence guidance decays smoothly from 22 to 11 over the first 60%60\% of the 50-step trajectory and is inactive thereafter; it is neither a hard switch nor constant-scale guidance. The full-Ω CFG ablation instead removes the global context and all source/query routed geometry from its reference branch. Table 5: Fixed geometry-conditioning hyperparameters. The relative-depth router constants are shared across scenes and resolutions. Component Setting Value Visual Geometry Router Secondary-layer relative margin δ=0.04δ=0.04 Visual Geometry Router Relative-depth visibility temperature τ=0.02τ=0.02 PTRC Full loss weight; Smooth-L1 transition 0.10.1; 0.10.1 PTRC schedule Stage-wise off / linear ramp / full-weight fractions 10%10\% / 20%20\% / 70%70\% Condition regularization Dropped / weakened / full-condition probabilities 0.100.10 / 0.200.20 / 0.700.70 Geometry-Prior CFG Guidance scale; denoising steps; guided interval 2.02.0; 5050; early 60%60\% Text conditioning Prompt; text-CFG scale empty; 11 A.3 Scaling with Data and Optimization Budget Protocol. We extend the full VGGT-Diff model trained on DL3DV clean-10K, containing 9,510 available scenes, from the original 18K fixed-compute budget to 100K effective steps. Every checkpoint is evaluated on the same 192 targets, with 32 cases in each interpolation/extrapolation near, mid, and far bin. All evaluations use six source views, 192×336192×336 inference, a 192×192192×192 center crop for metric computation, seed 20260823, and 50 sampling steps. We report LPIPS-VGG throughout this diagnostic; these values should not be compared directly with the LPIPS-AlexNet results in Table 1. The losses in Table 6 average the ten log entries from the 1K effective steps preceding each checkpoint. Flow loss is the original flow-matching objective, while PTRC denotes the raw unweighted point-track loss before applying its training weight or activation scale. Table 6: Checkpoint scaling on DL3DV clean-10K. Image metrics use the fixed 192-case half-resolution protocol. Losses average the preceding 1K effective steps; PTRC is reported before weighting. Effective steps PSNR ↑ SSIM ↑ LPIPS ↓ Flow loss ↓ PTRC loss ↓ 18K 18.7861 0.5272 0.2793 0.074406 0.087436 24K 19.1529 0.5509 0.2621 0.057234 0.086353 30K 19.1980 0.5579 0.2593 0.056245 0.085368 36K 19.6120 0.5770 0.2507 0.055986 0.084214 42K 19.5829 0.5762 0.2541 0.053832 0.080950 48K 19.6875 0.5830 0.2490 0.053431 0.080535 54K 19.8138 0.5889 0.2467 0.053542 0.079932 60K 19.4351 0.5725 0.2534 0.052149 0.077931 72K 19.9368 0.5973 0.2411 0.052154 0.077268 84K 20.1706 0.6088 0.2373 0.052251 0.076402 88K 20.1350 0.6059 0.2388 0.052218 0.076123 96K 20.2644 0.6146 0.2342 0.050399 0.074175 100K 20.2780 0.6166 0.2335 0.050209 0.073912 Figure 7: Evaluation metrics across the clean-10K optimization budget. Under the 192-case half-resolution protocol, PSNR, SSIM, and LPIPS-VGG are measured as training extends from 18K to 100K effective steps. Despite local fluctuations, all metrics improve substantially, with the best results at 100K. The open square denotes the clean-1K model evaluated at the matched 18K budget. Evaluation scaling. Table 6 and Fig. 7 show that the clean-10K model remains substantially under-optimized at 18K steps. Extending training to 100K raises PSNR from 18.786118.7861 to 20.278020.2780 dB and SSIM from 0.52720.5272 to 0.61660.6166, while reducing LPIPS-VGG from 0.27930.2793 to 0.23350.2335. These correspond to gains of 1.49181.4918 dB PSNR and 0.08940.0894 SSIM, together with a 0.04580.0458 reduction in LPIPS-VGG. Checkpoint quality is not strictly monotonic: 42K, 60K, and 88K exhibit local regressions, but each is followed by recovery and the long-term trend remains consistently favorable. The 100K checkpoint is best on all three image metrics. Its PSNR gain over 96K narrows to 0.01350.0135 dB, suggesting that optimization is approaching a plateau; continued SSIM and LPIPS-VGG improvements prevent concluding that performance has fully saturated. Figure 8: Training dynamics on DL3DV clean-10K. Curves connect non-overlapping 1K-step averages across the full training trajectory. The dashed line marks the original 18K fixed-compute budget. PTRC is reported as the raw, unweighted consistency loss. Its initial rise reflects scheduled activation and the increasingly difficult target curriculum rather than optimization divergence. Training dynamics. Figure 8 shows that optimization continues well beyond the original fixed-compute boundary. Between the 18K and 100K checkpoints, the preceding-1K Flow loss decreases from 0.0744060.074406 to 0.0502090.050209, a 32.5%32.5\% reduction, while the unweighted PTRC loss decreases from 0.0874360.087436 to 0.0739120.073912, a 15.5%15.5\% reduction. PTRC rises early because the objective is progressively activated as the target curriculum introduces more difficult multi-view correspondences. Once fully active, its sustained decline accompanies the continued improvement in image quality. Together, the evaluation and loss curves show that scaling the dataset requires a commensurate optimization budget to realize its benefit. A.4 Benefits of Joint Multi-Target Generation We compare generating eight targets in one diffusion sequence (Joint-8) with eight strictly independent 66-to-11 runs. Across 52 DL3DV scenes and 416 targets, both modes use the same checkpoint, six sources, eight target cameras, seed, per-view noise, and 50-step sampler. Independent runs cannot access other target cameras, latents, or predictions. Table 7: Joint versus independent eight-view generation. Image metrics use 416 targets; geometry proxies are computed from the same outputs with frozen VGGT-Ω . LPIPS uses VGG features, and Chamfer-L1 is trajectory-normalized. Mode PSNR ↑ SSIM ↑ LPIPS ↓ Chamfer ↓ F@1% ↑ F@2% ↑ Joint-8 17.6401 0.4787 0.3568 0.0483 0.3438 0.5341 Independent-8 17.1784 0.4453 0.3744 0.0741 0.2432 0.4047 Joint generation improves PSNR by 0.4620.462 dB and SSIM by 0.03340.0334, reduces LPIPS-VGG by 0.01760.0176, lowers normalized Chamfer-L1 by 0.02590.0259, and raises F-score@1% and F-score@2% by 0.10060.1006 and 0.12940.1294. Scene-bootstrap 95% confidence intervals exclude zero for all three image metrics and all three geometry metrics. Joint target slots therefore provide useful mutual context rather than merely batching otherwise independent predictions. A.5 VGGT-Ω Design Choices for Geometry Conditioning Setup. We examine feature depth and optimization scope under the same half-resolution 192-case protocol. First, we vary the extraction layer while keeping VGGT-Ω frozen. Second, using Layer 23, we jointly optimize its visual aggregator and encoder at a learning rate of 10−610^-6 while retaining frozen camera and depth heads. The Wan DiT, adapter, and Visual Geometry Router are trained normally in all variants; all remaining inputs and objectives are fixed. Figure 9: VGGT-Ω feature depth and optimization. Intermediate frozen features perform best, while jointly optimizing the Layer 23 visual pathway degrades PSNR. Table 8: VGGT-Ω feature and optimization diagnostics. All variants use the same half-resolution protocol; LPIPS uses VGG features. Configuration PSNR ↑ SSIM ↑ LPIPS ↓ Frozen VGGT-Ω : feature depth Layer 4 18.588 0.5160 0.2828 Layer 11 18.734 0.5253 0.2763 Layer 17 18.670 0.5197 0.2788 Layer 23 18.581 0.5155 0.2803 Additional optimization diagnostic Layer 23, jointly optimized 17.934 0.4803 0.3021 Feature depth. Table 8 shows a modest preference for intermediate features. Layer 11 performs best across all metrics, improving over Layer 23 by 0.1530.153 dB PSNR and 0.00980.0098 SSIM while reducing LPIPS-VGG by 0.00400.0040. Intermediate representations may better balance cross-view correspondence with local appearance: shallower features contain less aggregated multi-view context, whereas deeper features may trade fine detail for more abstract scene information. The limited variation among frozen layers also indicates robustness to feature depth. Frozen versus jointly optimized features. Joint optimization reduces PSNR by 0.6470.647 dB and SSIM by 0.03520.0352, while increasing LPIPS-VGG by 0.02180.0218 relative to the frozen Layer 23 setting. Under the current clean-980 data, learning rate, and frozen-head configuration, additional trainable capacity therefore does not improve NVS. One plausible explanation is that flow-matching gradients alter the pretrained visual representation while the frozen geometry heads retain their original feature organization, producing internal drift. These half-resolution results are not directly comparable with the full-resolution main results. All main-paper experiments retain the pre-specified frozen Layer 23 configuration to avoid retrospective model selection on the fixed evaluation set. The complementary behavior across depths motivates multi-level feature fusion as a structured alternative that preserves the pretrained geometry representation while combining local appearance with broader multi-view context. A.6 Complete Pose-Difficulty Results We provide complete PSNR and LPIPS results across interpolation and extrapolation difficulty. All methods use six fixed source views and independent 66-to-11 inference on a common 480×480480×480 image plane, with bins determined solely by camera pose. Each DL3DV bin contains 32 targets; Mip-NeRF 360 contains 66 targets per interpolation bin and 16 per extrapolation bin. Ranking highlights follow Table 1, and dashes denote unavailable results. Unlike the aggregate comparison in Table 1, this diagnostic separates trajectory relation from pose novelty. Interpolation and extrapolation are each divided into near, mid, and far bins, revealing whether performance degrades smoothly as target cameras move away from the available observations. The same frozen cases are used for every method, so differences across bins reflect pose robustness rather than sampling changes. Figure 6 provides the corresponding camera layouts and representative Vggt-Diff predictions for the six categories. Table 9: PSNR across target-pose difficulty (↑ ). Results are reported separately for DL3DV and Mip-NeRF 360. Method DL3DV Mip-NeRF 360 Interpolation Extrapolation Interpolation Extrapolation Near Mid Far Near Mid Far Near Mid Far Near Mid Far Regression-based Models AnySplat (Jiang et al., 2025b) 14.196 12.415 10.862 13.977 12.566 11.827 12.424 11.040 9.851 12.594 10.273 8.972 E-RayZer (Zhao et al., 2026) 17.388 14.878 13.824 16.833 14.477 14.712 15.964 15.085 14.028 15.404 14.760 12.609 LVSM (Jin et al., 2024) 20.276 16.599 13.969 20.790 16.633 15.579 16.117 14.488 13.722 15.296 14.040 13.016 DepthSplat (Xu et al., 2025) 20.376 17.424 14.820 20.653 17.943 16.374 17.478 16.046 15.303 16.208 15.518 12.511 Diffusion-based Models Aether (Zhu et al., 2025) 14.030 12.167 11.065 15.739 12.295 11.795 12.977 12.994 12.867 11.841 10.873 10.894 GEN3C (Ren et al., 2025) 13.620 12.356 11.548 14.288 13.050 12.288 – – – – – – MVSplat360 (Chen et al., 2024b) 15.282 14.420 13.389 15.542 14.479 14.176 14.283 13.928 13.585 14.407 13.463 12.794 FrameCrafter (Wu et al., 2026a) 20.338 17.485 14.649 21.168 17.888 16.594 16.933 15.700 15.016 16.222 15.330 13.017 SEVA (Zhou et al., 2025) 20.940 17.861 14.703 22.311 18.050 16.575 17.911 16.467 15.427 17.448 15.925 13.933 Vggt-Diff 20.920 18.391 15.853 22.396 18.793 17.793 17.254 16.630 15.611 16.336 16.061 13.989 Analysis. On DL3DV, Vggt-Diff achieves the best PSNR in five of six bins and the best LPIPS in all six, while trailing SEVA by only 0.0200.020 dB in interpolation-near. Its margin over FrameCrafter ranges from 0.582 to 1.228 dB and exceeds 0.9 dB in five of six bins. On zero-shot Mip-NeRF 360, Vggt-Diff obtains the best LPIPS in every bin and the best PSNR in four of six bins. The consistent perceptual gains across both datasets suggest that the geometry-routed diffusion prior remains scene-grounded under increasing viewpoint change. Table 10: LPIPS across target-pose difficulty (↓ ). We use VGG features on DL3DV and AlexNet features on Mip-NeRF 360. Method DL3DV Mip-NeRF 360 Interpolation Extrapolation Interpolation Extrapolation Near Mid Far Near Mid Far Near Mid Far Near Mid Far Regression-based Models AnySplat (Jiang et al., 2025b) 0.5052 0.5609 0.6214 0.5057 0.5483 0.5828 0.5513 0.5854 0.6135 0.5920 0.6349 0.6830 E-RayZer (Zhao et al., 2026) 0.5496 0.6357 0.6925 0.5527 0.6433 0.6508 0.7270 0.7978 0.8284 0.7494 0.7975 0.7859 LVSM (Jin et al., 2024) 0.2911 0.4413 0.5656 0.2375 0.3913 0.4835 0.4980 0.6461 0.6897 0.5315 0.6589 0.7058 DepthSplat (Xu et al., 2025) 0.2688 0.3592 0.4643 0.2600 0.3141 0.4077 0.3164 0.4058 0.4555 0.3965 0.4842 0.5703 Diffusion-based Models Aether (Zhu et al., 2025) 0.4949 0.5891 0.6405 0.3952 0.5579 0.6098 0.5731 0.5855 0.5978 0.5965 0.6857 0.6980 GEN3C (Ren et al., 2025) 0.5194 0.5808 0.6269 0.4773 0.5578 0.5990 – – – – – – MVSplat360 (Chen et al., 2024b) 0.5009 0.5374 0.5952 0.4855 0.5223 0.5724 0.5820 0.6453 0.6610 0.6404 0.6683 0.6743 FrameCrafter (Wu et al., 2026a) 0.2489 0.3558 0.4689 0.2150 0.3011 0.3996 0.2712 0.3639 0.3996 0.3148 0.4204 0.5152 SEVA (Zhou et al., 2025) 0.2460 0.3390 0.4546 0.2122 0.2959 0.3929 0.2564 0.3267 0.3854 0.2983 0.4013 0.4744 Vggt-Diff 0.2423 0.3227 0.4163 0.1922 0.2732 0.3572 0.2474 0.3085 0.3504 0.2922 0.3685 0.4424 A.7 Point-Cloud Reconstruction Figure 10 complements the quantitative geometry evaluation with point clouds reconstructed from jointly generated views. We use two frozen geometry models: VGGT-Ω as the primary probe and Pi3 (Wang et al., 2026b) as an independent reconstructor that is not used by Vggt-Diff for training or conditioning. Their agreement helps distinguish genuine cross-view improvements from potential shared-backbone evaluator bias. A coherent set of generated views provides mutually compatible evidence and should produce a stable point cloud close to the corresponding GT-view reconstruction. Cross-view drift instead produces duplicated, displaced, or fragmented structures. All methods use the same six source views and eight target cameras. Within each reconstructor, generated and GT views follow identical reconstruction settings and the same downstream alignment, filtering, and sampling procedure. Because VGGT-Ω and Pi3 produce different point-cloud representations, comparisons are made within each reconstructor rather than across them. Their GT-view reconstructions serve as model-specific references, not absolute 3D scans. Figure 10: Point-cloud reconstruction from jointly generated views. The upper and lower blocks use VGGT-Ω and Pi3, respectively. Within each block, normalized point-to-reference errors are shown above the corresponding reconstructions using the displayed color scale. Both reconstructors show fewer displaced points and more coherent geometry for Vggt-Diff, indicating stronger cross-view consistency. Both reconstructors produce the same overall ordering summarized in Table 4. SEVA (Zhou et al., 2025) and FrameCrafter exhibit larger high-error regions, displaced foreground points, and less stable facade structure. In contrast, Vggt-Diff produces more compact reconstructions with fewer structural offsets and better preserves the shared layout across views. The agreement between VGGT-Ω and Pi3 supports the Chamfer-L1 and F-score gains while reducing dependence on any single geometry evaluator. A.8 Residual versus Raw-Velocity Consistency Motivation. PTRC regularizes predicted-clean residuals rather than directly tying raw model outputs across views. To isolate this design choice, we replace PTRC with a naive objective that applies the same point tracks, confidence weights, and robust penalty directly to predicted velocities: ℒVel=∑γ∈wγρ¯β(θ,a(γ,a)−θ,b(γ,b))∑γ∈wγ.L_Vel= _γ w_γ\, ρ_β (V_θ,a(u_γ,a)-V_θ,b(u_γ,b) ) _γ w_γ. (12) This alternative assumes that corresponding locations should share the same velocity. Under linear flow matching, however, each view has its own target j=ϵj−0,jU_j= ε_j-Z_0,j; view-specific clean latents and noise states therefore make raw velocity equality generally invalid. Since j=−σ(θ,j−j)E_j=-σ(V_θ,j-U_j), PTRC instead aligns errors relative to each view’s own flow target, preserving valid differences between their velocities. This comparison is stricter than simply removing PTRC. The direct-matching baseline retains the same 3D correspondences and explicitly couples the same query locations, but changes what is required to agree. It therefore tests whether the benefit comes from generic point-track smoothing or specifically from aligning residual errors relative to each view’s supervision target. Controlled comparison. Both variants share the data, initialization, geometry conditioning, point-track construction, and optimization, differing only in the consistency objective. For efficiency, they follow the same half-resolution protocol and are evaluated on the same 192 pose-stratified targets. Geometry conditioning, condition regularization, target curriculum, and robust penalty remain unchanged; both objectives use the stage-wise schedule in Eq. 9, with full weight 0.10.1. Table 11: PTRC versus direct velocity matching across pose difficulty. Both use the same half-resolution protocol; PTRC values are highlighted and LPIPS uses VGG features. Interpolation Difficulty PSNR ↑ SSIM ↑ LPIPS ↓ Direct PTRC Direct PTRC Direct PTRC Near 19.008 20.133 0.531 0.602 0.274 0.221 Mid 17.520 18.103 0.435 0.467 0.341 0.300 Far 15.210 15.670 0.308 0.352 0.440 0.397 Extrapolation Difficulty PSNR ↑ SSIM ↑ LPIPS ↓ Direct PTRC Direct PTRC Direct PTRC Near 20.021 21.326 0.591 0.674 0.238 0.173 Mid 17.863 18.613 0.478 0.535 0.308 0.260 Far 16.994 17.641 0.415 0.463 0.377 0.331 Analysis. Across all 192 targets, PTRC improves PSNR from 17.76917.769 to 18.58118.581 dB and SSIM from 0.4600.460 to 0.5160.516, while reducing LPIPS from 0.3300.330 to 0.2800.280. Table 11 further shows consistent gains on every metric across all six pose bins, with PSNR improvements ranging from 0.4600.460 to 1.3051.305 dB. Because both variants use identical correspondences and weighting, these gains isolate the residual-relative formulation rather than the point tracks themselves. Direct velocity matching penalizes valid differences between view-specific flow targets, whereas PTRC couples only the prediction errors that should be removed. The improvement appears in both interpolation and extrapolation, showing that the residual formulation is not tied to one camera relation or difficulty range. A.9 Additional Optimization and Statistical Diagnostics Wan adaptation. We compare full DiT fine-tuning with a rank-32 LoRA adaptation while retaining the same geometry conditioner, data, initialization, objectives, and half-resolution 192-case evaluation. Full fine-tuning improves PSNR by 1.3241.324 dB and SSIM by 0.04780.0478, while reducing LPIPS-VGG by 0.07510.0751. The large gap indicates that sparse low-rank updates are insufficient to integrate routed geometry into the pretrained video prior under this setup. Table 12: Effect of Wan adaptation strategy. Both variants use the same half-resolution controlled protocol; LPIPS uses VGG features. Wan adaptation PSNR ↑ SSIM ↑ LPIPS ↓ LoRA, rank 32 17.2563 0.4677 0.3554 Full fine-tuning 18.5807 0.5155 0.2803 Paired scene-level statistics. We additionally compare locally generated Vggt-Diff and FrameCrafter outputs under the same 140-scene, 6,188-target protocol. The scene-mean PSNR gain is 0.78930.7893 dB, with a 92.14% scene win rate and a scene-bootstrap 95% confidence interval of [0.7003,0.8769][0.7003,0.8769] dB. The corresponding scene-mean changes are +0.0330+0.0330 SSIM, −0.0365-0.0365 LPIPS-AlexNet, and −0.0118-0.0118 DreamSim. These paired results show that the aggregate improvement is distributed across scenes rather than driven by a small subset. They are reported separately from Table 1, whose baseline rows follow the published cross-method protocol. A.10 Robustness to Sparse Input Views Protocol. We evaluate whether the frozen full-resolution checkpoint trained with six source views transfers to reduced source coverage without retraining or test-time optimization. Both Vggt-Diff and FrameCrafter use independent M-to-11 inference on the same 192 pose-stratified DL3DV targets, with M∈3,4M∈\3,4\. Inference runs at 480×832480×832 for 50 sampling steps, and metrics are computed on the center 480×480480×480 crop. Both methods use identical source images, target cameras, and per-case noise seeds. For each target, we select M views from the six sources fixed by the main protocol using camera poses only. A source s is ranked by (∥t−s∥2/D)2+(θts/π)2 ( _t-c_s _2/D)^2+( _ts/π)^2, where D is the maximum pairwise distance among the six source cameras and θts _ts is their viewing direction difference. The selected views are restored to their registered trajectory order, producing nested source subsets independent of image content, predictions, and method identity. Table 13: Robustness to sparse input views on DL3DV. Both methods use full-resolution independent M-to-11 inference on the same 192 targets. Source subsets are selected identically using only camera poses; LPIPS uses VGG features. Sources Method PSNR ↑ SSIM ↑ LPIPS ↓ DreamSim ↓ 3 FrameCrafter (Wu et al., 2026a) 16.9559 0.4647 0.3709 0.0807 Vggt-Diff 17.7372 0.4966 0.3391 0.0680 4 FrameCrafter (Wu et al., 2026a) 17.8545 0.5076 0.3390 0.0685 Vggt-Diff 18.4380 0.5307 0.3135 0.0592 Results. Despite being trained with six sources, Vggt-Diff remains stronger under both reduced-view settings. With three inputs, it improves over FrameCrafter by 0.7810.781 dB PSNR and 0.03190.0319 SSIM while reducing LPIPS-VGG by 0.03180.0318 and DreamSim by 0.01270.0127. With four inputs, the corresponding gains are 0.5840.584 dB, 0.02310.0231, 0.02550.0255, and 0.00930.0093. The consistent improvement across all four metrics shows that geometry-routed conditioning remains effective when the available observations are reduced to three or four views, without adapting the six-view-trained checkpoint. Figure 11: Target-view alignment and pixel error. Predicted RGB views are shown above their absolute pixel differences from GT under a shared error scale. Vggt-Diff produces lower and more localized errors than the baselines, indicating closer alignment with the GT viewpoint and fewer spatial or structural mismatches. A.11 VAE Encoding Protocols All Vggt-Diff variants use the same frozen Wan2.1 Video VAE, which produces 16-channel latents with an 8×8× spatial downsampling factor (192×336→16×24×42192×336→ 16×24×42 and 480×832→16×60×104480×832→ 16×60×104). All quantitative experiments adopt an independent-view latent layout. During training, each source image and ground-truth target image is independently encoded, with the latter providing the clean latent for denoising supervision. During inference, only source images are encoded; target latent slots are initialized from noise, and no target RGB image is provided to the model. Consequently, six sources and N targets produce 6+N6+N view-time latent slots, each aligned one-to-one with a physical view, camera, Plücker map, and routed geometry condition. Thus, none of our reported quantitative results uses temporal VAE compression. To accelerate experimentation and reduce the training cost of long sequences, we use a hybrid protocol for the continuous 80-frame trajectory demo and fine-tune it at the half resolution of 192×336192×336. The six sources remain independently encoded, while the ordered target sequence is jointly encoded through the causal temporal pathway of the same frozen VAE, yielding LT=1+⌈(T−1)/4⌉L_T=1+ (T-1)/4 target slots. For T=80T=80, we repeat the final frame once to form an 81-frame sequence, obtaining LT=21L_T=21 target slots; the repeated output is discarded after decoding. This reduces the number of view-time latent slots processed by the DiT from 6+80=866+80=86 to 6+21=276+21=27. Camera conditions bypass the VAE. Each six-channel Plücker map is spatially rearranged into 384384 channels at latent resolution using an 8×8× PixelUnshuffle. To avoid temporally discarding intermediate target poses, we group these maps according to the causal VAE windows: [0][0], [1,2,3,4][1,2,3,4], …, and [77,78,79,79][77,78,79,79]. Within each four-frame window, the ordered 384384-channel maps are concatenated into 15361536 channels and projected back to 384384 by a lightweight 1×1×11×1×1 Plücker adapter, making all four poses available to the conditioning pathway while preserving the DiT interface. The first target slot retains its single-frame conditioning path. We initialize the adapter to select only the fourth map, making the augmented model exactly equivalent to the previous endpoint-only checkpoint at initialization and enabling checkpoint-compatible fine-tuning without modifying the DiT backbone. In the current adaptation, target camera matrices and routed geometry conditions remain aligned with the endpoint indices T=0,4,8,…,76,79A_T=\0,4,8,…,76,79\; only the Plücker pathway aggregates all poses within each temporal window. Because independently encoded and temporally compressed target latents have different temporal semantics, we fine-tune the initialized model for this 80-frame regime rather than treating the two representations as interchangeable. This half-resolution extension is used only for the continuous demo and does not affect any reported quantitative result. A.12 Future Directions Our results suggest several directions for extending geometry-routed multi-view generation. First, the complementary behavior observed across VGGT-Ω feature depths motivates adaptive multi-level fusion that combines fine appearance cues with increasingly global geometric context. Second, although directly optimizing the VGGT-Ω visual pathway is ineffective under our current setting, parameter-efficient adaptation, geometry-preserving objectives, or staged optimization may better specialize the encoder without disrupting its pretrained representations. The Visual Geometry Router could also be extended beyond its hard anchor and layered residual refinement to model richer visibility distributions, multiple occlusion layers, and better-calibrated geometric uncertainty. At the generation level, memory-efficient joint denoising and chunked generation may scale the benefits of shared multi-view context to more views and longer camera trajectories while maintaining consistency across chunks. Finally, extending 3D point tracks to spatiotemporal correspondences could generalize the framework from static scenes to dynamic and non-rigid content with moving objects and time-varying occlusions.