LYHDEMO · EVALUATION

Camera-controlled novel view synthesis

Trend on a fixed twelve-case subset of the evaluation: Mundo against video-generation and feed-forward baselines that render a ground-truth camera trajectory from posed context views.

Setup

Cases. A fixed subset of the test set: the first two test scenes of ScanNet++ [7], DL3DV [8] and CA-1M [9] in identifier order, each with a near and a far trajectory, twelve cases in total. A trajectory is a real capture segment of thirteen keyframes; the far trajectory leaves the coverage of the context views.
Inputs. Every method receives ground-truth camera poses and intrinsics and no depth. Methods that accept several images receive the same nine context views; FramePack receives the first one, Holo4D a 49-frame video that contains the nine views. FramePack and MiniMax H3 take no camera input and follow a text description of the trajectory.
Metrics. Frames are cropped to a shared 4:3 canvas of 512×384 with identical rays for every method. PSNR, SSIM and LPIPS (VGG) on keyframes 1 to 12, over the full image and over the pixels that the context views see. Camera following: Pi3X estimates the cameras of the thirteen keyframes of each video, aligned to the ground truth with a similarity transform; ATE is divided by the extent of the trajectory.
Training overlap. The DL3DV test scenes belong to DL3DV-10K, which is part of the training data of LagerNVS and Gen3C; whether these scenes were held out is not documented. Mundo was not trained on DL3DV or ScanNet++, and CA-1M test scenes are absent from its training data.
Status. A method with fewer than twelve cases is still running; its means cover the cases listed and are not ranked.

Results

best in columnsecond best

Method	Cases	Full image			Covisible pixels			Camera following
Method	Cases	PSNR ↑	SSIM ↑	LPIPS ↓	PSNR ↑	SSIM ↑	LPIPS ↓	ATE ↓	RPE-R ↓	RotErr ↓
Mundo 9 views	12/12	15.181	0.528	0.452	15.390	0.540	0.442	0.028	1.316	3.268
Stable Virtual Camera [1] 9 views	8/12	14.409	0.583	0.506	14.672	0.602	0.487	0.030	1.185	2.928
Gen3C [2] 9 views	0/12	–	–	–	–	–	–	–	–	–
Holo4D [3] 49-frame video	0/12	–	–	–	–	–	–	–	–	–
LagerNVS [4] 9 views	12/12	19.126	0.686	0.389	19.712	0.704	0.370	0.029	1.233	2.749
FramePack [5] 1 view + text	12/12	9.989	0.368	0.673	10.046	0.369	0.668	0.231	10.415	47.273
MiniMax H3 [6] 9 views + text	8/12	10.433	0.336	0.640	10.368	0.343	0.641	0.126	12.773	60.559
Ground-truth video estimator floor	12/12	–	–	–	–	–	–	0.012	0.418	1.113

ScanNet++ [7]

Method	Cases	Full image			Covisible pixels			Camera following
Method	Cases	PSNR ↑	SSIM ↑	LPIPS ↓	PSNR ↑	SSIM ↑	LPIPS ↓	ATE ↓	RPE-R ↓	RotErr ↓
Mundo 9 views	4/4	14.429	0.583	0.468	14.822	0.612	0.457	0.023	1.351	3.151
Stable Virtual Camera [1] 9 views	4/4	13.127	0.559	0.508	13.271	0.581	0.498	0.038	1.107	2.752
Gen3C [2] 9 views	0/4	–	–	–	–	–	–	–	–	–
Holo4D [3] 49-frame video	0/4	–	–	–	–	–	–	–	–	–
LagerNVS [4] 9 views	4/4	17.361	0.689	0.442	17.932	0.720	0.423	0.031	0.982	2.276
FramePack [5] 1 view + text	4/4	8.215	0.392	0.690	8.284	0.403	0.687	0.212	12.526	51.413
MiniMax H3 [6] 9 views + text	4/4	9.206	0.444	0.667	9.157	0.463	0.667	0.147	18.518	88.795
Ground-truth video estimator floor	4/4	–	–	–	–	–	–	0.023	0.638	1.535

DL3DV [8]

Method	Cases	Full image			Covisible pixels			Camera following
Method	Cases	PSNR ↑	SSIM ↑	LPIPS ↓	PSNR ↑	SSIM ↑	LPIPS ↓	ATE ↓	RPE-R ↓	RotErr ↓
Mundo 9 views	4/4	15.865	0.394	0.377	15.894	0.393	0.372	0.011	0.548	1.585
Stable Virtual Camera [1] 9 views	0/4	–	–	–	–	–	–	–	–	–
Gen3C [2] 9 views	0/4	–	–	–	–	–	–	–	–	–
Holo4D [3] 49-frame video	0/4	–	–	–	–	–	–	–	–	–
LagerNVS [4] 9 views	4/4	21.139	0.658	0.274	21.402	0.667	0.262	0.005	0.161	0.578
FramePack [5] 1 view + text	4/4	11.397	0.236	0.626	11.425	0.234	0.626	0.256	5.565	32.054
MiniMax H3 [6] 9 views + text	4/4	11.661	0.229	0.612	11.580	0.224	0.615	0.105	7.029	32.323
Ground-truth video estimator floor	4/4	–	–	–	–	–	–	0.003	0.122	0.443

CA-1M [9]

Method	Cases	Full image			Covisible pixels			Camera following
Method	Cases	PSNR ↑	SSIM ↑	LPIPS ↓	PSNR ↑	SSIM ↑	LPIPS ↓	ATE ↓	RPE-R ↓	RotErr ↓
Mundo 9 views	4/4	15.249	0.606	0.511	15.454	0.615	0.498	0.049	2.049	5.069
Stable Virtual Camera [1] 9 views	4/4	15.691	0.607	0.504	16.073	0.623	0.476	0.022	1.264	3.103
Gen3C [2] 9 views	0/4	–	–	–	–	–	–	–	–	–
Holo4D [3] 49-frame video	0/4	–	–	–	–	–	–	–	–	–
LagerNVS [4] 9 views	4/4	18.879	0.712	0.452	19.802	0.725	0.424	0.052	2.557	5.391
FramePack [5] 1 view + text	4/4	10.355	0.475	0.703	10.428	0.469	0.689	0.226	13.154	58.351
MiniMax H3 [6] 9 views + text	0/4	–	–	–	–	–	–	–	–	–
Ground-truth video estimator floor	4/4	–	–	–	–	–	–	0.011	0.495	1.363

Cases

ScanNet++ · 0529d56cce · near trajectorycovisibility 0.87, 2.8 m, 82°

Context views (the first is the trajectory start)

Ground-truth keyframes 0, 3, 6, 9, 12

MundoPSNR 17.88 · LPIPS 0.394 · ATE 0.008

Stable Virtual CameraPSNR 12.36 · LPIPS 0.545 · ATE 0.012

LagerNVSPSNR 18.76 · LPIPS 0.402 · ATE 0.007

FramePackPSNR 7.56 · LPIPS 0.745 · ATE 0.208

MiniMax H3PSNR 8.99 · LPIPS 0.715 · ATE 0.052

ScanNet++ · 0529d56cce · far trajectorycovisibility 0.55, 1.8 m, 68°

Context views (the first is the trajectory start)

Ground-truth keyframes 0, 3, 6, 9, 12

MundoPSNR 13.17 · LPIPS 0.527 · ATE 0.022

Stable Virtual CameraPSNR 14.01 · LPIPS 0.516 · ATE 0.023

LagerNVSPSNR 15.53 · LPIPS 0.523 · ATE 0.026

FramePackPSNR 8.12 · LPIPS 0.683 · ATE 0.274

MiniMax H3PSNR 8.71 · LPIPS 0.691 · ATE 0.195

ScanNet++ · 1ada7a0617 · near trajectorycovisibility 0.86, 0.8 m, 170°

Context views (the first is the trajectory start)

Ground-truth keyframes 0, 3, 6, 9, 12

MundoPSNR 15.47 · LPIPS 0.381 · ATE 0.041

Stable Virtual CameraPSNR 14.75 · LPIPS 0.388 · ATE 0.075

LagerNVSPSNR 20.09 · LPIPS 0.324 · ATE 0.040

FramePackPSNR 10.13 · LPIPS 0.635 · ATE 0.231

MiniMax H3PSNR 10.71 · LPIPS 0.590 · ATE 0.151

ScanNet++ · 1ada7a0617 · far trajectorycovisibility 0.72, 1.0 m, 117°

Context views (the first is the trajectory start)

Ground-truth keyframes 0, 3, 6, 9, 12

MundoPSNR 11.20 · LPIPS 0.570 · ATE 0.022

Stable Virtual CameraPSNR 11.38 · LPIPS 0.582 · ATE 0.043

LagerNVSPSNR 15.05 · LPIPS 0.519 · ATE 0.052

FramePackPSNR 7.05 · LPIPS 0.697 · ATE 0.133

MiniMax H3PSNR 8.42 · LPIPS 0.673 · ATE 0.191

CA-1M · 42897521 · near trajectorycovisibility 0.89, 1.1 m, 95°

Context views (the first is the trajectory start)

Ground-truth keyframes 0, 3, 6, 9, 12

MundoPSNR 16.36 · LPIPS 0.444 · ATE 0.021

Stable Virtual CameraPSNR 16.31 · LPIPS 0.435 · ATE 0.013

LagerNVSPSNR 20.07 · LPIPS 0.371 · ATE 0.007

FramePackPSNR 9.41 · LPIPS 0.711 · ATE 0.294

CA-1M · 42897521 · far trajectorycovisibility 0.66, 1.3 m, 179°

Context views (the first is the trajectory start)

Ground-truth keyframes 0, 3, 6, 9, 12

MundoPSNR 12.40 · LPIPS 0.558 · ATE 0.028

Stable Virtual CameraPSNR 13.88 · LPIPS 0.547 · ATE 0.027

LagerNVSPSNR 16.66 · LPIPS 0.499 · ATE 0.016

FramePackPSNR 9.47 · LPIPS 0.749 · ATE 0.233

CA-1M · 42897545 · near trajectorycovisibility 0.92, 2.2 m, 57°

Context views (the first is the trajectory start)

Ground-truth keyframes 0, 3, 6, 9, 12

MundoPSNR 16.92 · LPIPS 0.470 · ATE 0.034

Stable Virtual CameraPSNR 15.52 · LPIPS 0.502 · ATE 0.022

LagerNVSPSNR 19.14 · LPIPS 0.443 · ATE 0.038

FramePackPSNR 11.93 · LPIPS 0.644 · ATE 0.237

CA-1M · 42897545 · far trajectorycovisibility 0.49, 2.1 m, 106°

Context views (the first is the trajectory start)

Ground-truth keyframes 0, 3, 6, 9, 12

MundoPSNR 15.31 · LPIPS 0.572 · ATE 0.114

Stable Virtual CameraPSNR 17.06 · LPIPS 0.534 · ATE 0.028

LagerNVSPSNR 19.65 · LPIPS 0.497 · ATE 0.148

FramePackPSNR 10.61 · LPIPS 0.706 · ATE 0.139

DL3DV · 165f5af8bf · near trajectorycovisibility 0.96, 5.2 m, 75°

Context views (the first is the trajectory start)

Ground-truth keyframes 0, 3, 6, 9, 12

MundoPSNR 15.53 · LPIPS 0.372 · ATE 0.010

LagerNVSPSNR 22.22 · LPIPS 0.244 · ATE 0.005

FramePackPSNR 11.54 · LPIPS 0.621 · ATE 0.554

MiniMax H3PSNR 12.37 · LPIPS 0.565 · ATE 0.141

DL3DV · 165f5af8bf · far trajectorycovisibility 0.88, 2.6 m, 31°

Context views (the first is the trajectory start)

Ground-truth keyframes 0, 3, 6, 9, 12

MundoPSNR 16.77 · LPIPS 0.387 · ATE 0.018

LagerNVSPSNR 21.02 · LPIPS 0.335 · ATE 0.009

FramePackPSNR 11.41 · LPIPS 0.672 · ATE 0.149

MiniMax H3PSNR 11.57 · LPIPS 0.654 · ATE 0.160

DL3DV · 341b4ff3df · near trajectorycovisibility 1.00, 4.5 m, 63°

Context views (the first is the trajectory start)

Ground-truth keyframes 0, 3, 6, 9, 12

MundoPSNR 15.80 · LPIPS 0.362 · ATE 0.005

LagerNVSPSNR 22.28 · LPIPS 0.236 · ATE 0.002

FramePackPSNR 10.79 · LPIPS 0.627 · ATE 0.160

MiniMax H3PSNR 10.96 · LPIPS 0.629 · ATE 0.022

DL3DV · 341b4ff3df · far trajectorycovisibility 0.80, 3.7 m, 35°

Context views (the first is the trajectory start)

Ground-truth keyframes 0, 3, 6, 9, 12

MundoPSNR 15.36 · LPIPS 0.387 · ATE 0.009

LagerNVSPSNR 19.03 · LPIPS 0.280 · ATE 0.002

FramePackPSNR 11.85 · LPIPS 0.585 · ATE 0.162

MiniMax H3PSNR 11.75 · LPIPS 0.601 · ATE 0.098

References

Zhou et al. Stable Virtual Camera: generative view synthesis with diffusion models. arXiv 2503.14489.
Ren et al. GEN3C: 3D-informed world-consistent video generation with precise camera control. CVPR 2025.
Lan et al. Holo4D: holistic 4D reconstruction as geometric control for video diffusion. arXiv 2610.25276.
Szymanowicz et al. LagerNVS: latent geometry for fully neural real-time novel view synthesis. CVPR 2026.
Zhang and Agrawala. Packing input frame context in next-frame prediction models for video generation (FramePack). arXiv 2504.12626.
MiniMax. MiniMax-H3 (Ref2VA), open-weight video generation model.
Yeshwanth et al. ScanNet++: a high-fidelity dataset of 3D indoor scenes. ICCV 2023.
Ling et al. DL3DV-10K: a large-scale scene dataset for deep learning-based 3D vision. CVPR 2024.
Lazarow et al. Cubify Anything: scaling indoor 3D object detection (CA-1M). CVPR 2025.