MUNDO · 20 FIXED CASES · CAMERA AND PROXY ABLATIONS

What changes without camera embeddings?

Compare the same target views against ground truth. No camera keeps the proxy video and context images but removes camera encoders and embeddings. No proxy keeps camera conditioning, injects the target camera embedding into target video tokens, and removes proxy visual tokens from both the renderer and Qwen.

Protocol. Ten fixed cases per dataset. 124 frames at 640 × 480; dpmpp2m, 32 steps, seed 42, TP4. Only the explicitly labeled no-proxy 1000 first-view variant fixes the anchor latent; all other rows have no fixed first-frame condition. Score keyframes 1–12, excluding the anchor, on the common 512 × 384 canvas. PSNR / SSIM / LPIPS-VGG use full images; clean-fid Inception-V3 FID uses 120 scored images per dataset and method.
Checkpoint scope. Resume 1200 is global mix-training step 1200. No camera 600 and no proxy 1000 start at mix step 1000 and receive 600 and 1000 additional ARKitScenes steps respectively. These training budgets differ. No proxy 1900 continues the same no-proxy run for another 900 steps (with a changed distributed batch configuration). Release camera-off uses unchanged release weights and removes camera encoder/embedding contributions only during inference; it is distinct from no-camera training. Release camera-on results reuse the original evaluation. Mix resume 1200 is evaluated on both datasets. FID is descriptive on this small set of correlated frames.

ScanNet++

Release camera-on results reuse the original evaluation. Mix resume 1200 is shown separately. The first-view variant is evaluated on ScanNet++ only.

Method	PSNR ↑	SSIM ↑	LPIPS ↓	FID ↓
Release · camera on	15.91	0.622	0.404	63.27
Mix resume · 1200	15.54	0.617	0.410	65.91
No camera trained · 600	15.49	0.610	0.421	65.07
No proxy · 1000	9.46	0.444	0.666	128.29
No proxy 1000 · first view	9.33	0.439	0.665	127.07
No proxy · 1900	9.64	0.451	0.660	114.13
Release · camera off (inference)	15.93	0.623	0.404	63.62

0529d56cce_near

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	17.88	0.607	0.394
Mix resume · 1200	16.72	0.585	0.413
No camera trained · 600	17.62	0.602	0.407
No proxy · 1000	8.87	0.403	0.716
No proxy 1000 · first view	7.82	0.310	0.738
No proxy · 1900	9.00	0.419	0.727
Release · camera off (inference)	17.91	0.607	0.393

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Matched ground truth and all evaluated settings

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy 1000 · first view

No proxy · 1900

Release · camera off (inference)

1ada7a0617_near

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	15.47	0.711	0.381
Mix resume · 1200	14.89	0.698	0.391
No camera trained · 600	14.63	0.685	0.413
No proxy · 1000	9.86	0.566	0.632
No proxy 1000 · first view	9.71	0.570	0.629
No proxy · 1900	9.99	0.573	0.631
Release · camera off (inference)	15.56	0.713	0.382

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy 1000 · first view

No proxy · 1900

Release · camera off (inference)

4a1a3a7dc5_near

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	15.17	0.564	0.428
Mix resume · 1200	14.98	0.558	0.435
No camera trained · 600	14.74	0.552	0.444
No proxy · 1000	9.44	0.373	0.670
No proxy 1000 · first view	9.31	0.365	0.677
No proxy · 1900	9.30	0.365	0.680
Release · camera off (inference)	15.16	0.564	0.429

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy 1000 · first view

No proxy · 1900

Release · camera off (inference)

75d29d69b8_near

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	16.95	0.638	0.363
Mix resume · 1200	16.67	0.632	0.367
No camera trained · 600	16.30	0.623	0.383
No proxy · 1000	8.21	0.329	0.710
No proxy 1000 · first view	7.90	0.340	0.691
No proxy · 1900	8.45	0.346	0.706
Release · camera off (inference)	16.95	0.638	0.362

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy 1000 · first view

No proxy · 1900

Release · camera off (inference)

7e7cd69a59_near

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	13.83	0.496	0.505
Mix resume · 1200	13.82	0.494	0.508
No camera trained · 600	13.60	0.475	0.521
No proxy · 1000	9.44	0.358	0.709
No proxy 1000 · first view	9.37	0.360	0.695
No proxy · 1900	9.55	0.362	0.706
Release · camera off (inference)	13.83	0.496	0.506

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy 1000 · first view

No proxy · 1900

Release · camera off (inference)

98b4ec142f_near

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	12.27	0.490	0.456
Mix resume · 1200	12.20	0.488	0.455
No camera trained · 600	12.09	0.486	0.463
No proxy · 1000	7.93	0.319	0.662
No proxy 1000 · first view	8.16	0.329	0.655
No proxy · 1900	8.31	0.336	0.650
Release · camera off (inference)	12.28	0.491	0.455

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy 1000 · first view

No proxy · 1900

Release · camera off (inference)

accad58571_near

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	15.16	0.600	0.465
Mix resume · 1200	15.10	0.604	0.460
No camera trained · 600	15.01	0.600	0.466
No proxy · 1000	10.73	0.540	0.597
No proxy 1000 · first view	10.70	0.541	0.600
No proxy · 1900	11.31	0.538	0.587
Release · camera off (inference)	15.16	0.600	0.464

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy 1000 · first view

No proxy · 1900

Release · camera off (inference)

b20a261fdf_near

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	18.20	0.746	0.356
Mix resume · 1200	18.00	0.742	0.361
No camera trained · 600	18.10	0.741	0.365
No proxy · 1000	11.21	0.568	0.646
No proxy 1000 · first view	11.16	0.587	0.642
No proxy · 1900	11.60	0.594	0.622
Release · camera off (inference)	18.22	0.747	0.356

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy 1000 · first view

No proxy · 1900

Release · camera off (inference)

d6d9ddb03f_near

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	17.24	0.704	0.300
Mix resume · 1200	16.13	0.701	0.317
No camera trained · 600	16.30	0.677	0.338
No proxy · 1000	8.74	0.464	0.655
No proxy 1000 · first view	8.76	0.468	0.656
No proxy · 1900	8.73	0.454	0.649
Release · camera off (inference)	17.23	0.705	0.299

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy 1000 · first view

No proxy · 1900

Release · camera off (inference)

da8043d54e_near

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	16.99	0.669	0.394
Mix resume · 1200	16.88	0.671	0.392
No camera trained · 600	16.54	0.658	0.412
No proxy · 1000	10.22	0.515	0.664
No proxy 1000 · first view	10.34	0.519	0.666
No proxy · 1900	10.16	0.525	0.642
Release · camera off (inference)	17.02	0.669	0.394

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy 1000 · first view

No proxy · 1900

Release · camera off (inference)

Tanks and Temples

Release camera-on results reuse the original evaluation. Mix resume 1200 is shown separately. T&T uses the existing Pi3X camera estimates; the fixed list includes Playground and Temple.

Method	PSNR ↑	SSIM ↑	LPIPS ↓	FID ↓
Release · camera on	15.41	0.384	0.451	54.00
Mix resume · 1200	15.14	0.374	0.458	54.40
No camera trained · 600	15.16	0.378	0.462	51.49
No proxy · 1000	11.59	0.273	0.634	94.37
No proxy · 1900	11.64	0.269	0.630	101.08
Release · camera off (inference)	15.42	0.383	0.451	53.49

Auditorium_near

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	17.85	0.619	0.444
Mix resume · 1200	17.26	0.603	0.460
No camera trained · 600	17.79	0.617	0.448
No proxy · 1000	12.32	0.425	0.617
No proxy · 1900	12.48	0.428	0.616
Release · camera off (inference)	17.84	0.619	0.444

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy · 1900

Release · camera off (inference)

Ballroom_near

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	13.70	0.232	0.467
Mix resume · 1200	13.53	0.227	0.466
No camera trained · 600	12.77	0.206	0.499
No proxy · 1000	10.41	0.168	0.668
No proxy · 1900	9.96	0.159	0.652
Release · camera off (inference)	13.71	0.231	0.467

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy · 1900

Release · camera off (inference)

Church_near

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	17.70	0.336	0.404
Mix resume · 1200	17.11	0.305	0.434
No camera trained · 600	16.50	0.302	0.440
No proxy · 1000	14.42	0.255	0.584
No proxy · 1900	14.96	0.263	0.581
Release · camera off (inference)	17.70	0.335	0.403

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy · 1900

Release · camera off (inference)

Courtroom_near

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	18.89	0.533	0.365
Mix resume · 1200	18.47	0.518	0.370
No camera trained · 600	17.39	0.471	0.419
No proxy · 1000	12.34	0.307	0.672
No proxy · 1900	12.41	0.311	0.677
Release · camera off (inference)	18.88	0.533	0.364

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy · 1900

Release · camera off (inference)

Meetingroo_near

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	15.06	0.423	0.462
Mix resume · 1200	14.68	0.415	0.467
No camera trained · 600	14.98	0.425	0.481
No proxy · 1000	11.07	0.252	0.654
No proxy · 1900	10.22	0.244	0.653
Release · camera off (inference)	15.05	0.423	0.462

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy · 1900

Release · camera off (inference)

Ballroom_near_n2

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	10.76	0.171	0.496
Mix resume · 1200	10.18	0.165	0.500
No camera trained · 600	11.27	0.170	0.494
No proxy · 1000	7.87	0.122	0.652
No proxy · 1900	7.94	0.124	0.651
Release · camera off (inference)	10.75	0.166	0.504

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy · 1900

Release · camera off (inference)

Church_near_n2

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	17.08	0.333	0.386
Mix resume · 1200	16.27	0.299	0.425
No camera trained · 600	18.02	0.398	0.336
No proxy · 1000	13.65	0.234	0.583
No proxy · 1900	14.24	0.251	0.549
Release · camera off (inference)	17.10	0.334	0.385

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy · 1900

Release · camera off (inference)

Meetingroo_near_n2

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	15.32	0.540	0.468
Mix resume · 1200	16.01	0.553	0.445
No camera trained · 600	15.22	0.542	0.475
No proxy · 1000	11.29	0.431	0.632
No proxy · 1900	11.39	0.402	0.635
Release · camera off (inference)	15.37	0.542	0.467

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy · 1900

Release · camera off (inference)

Playground_near

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	15.15	0.327	0.493
Mix resume · 1200	15.54	0.337	0.480
No camera trained · 600	15.50	0.329	0.497
No proxy · 1000	12.83	0.278	0.622
No proxy · 1900	12.20	0.264	0.634
Release · camera off (inference)	15.17	0.327	0.491

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy · 1900

Release · camera off (inference)

Temple_near

↑ Dataset

Method	PSNR ↑	SSIM ↑	LPIPS ↓
Release · camera on	12.61	0.324	0.528
Mix resume · 1200	12.35	0.322	0.530
No camera trained · 600	12.17	0.317	0.533
No proxy · 1000	9.74	0.254	0.655
No proxy · 1900	10.56	0.238	0.652
Release · camera off (inference)	12.62	0.324	0.525

Matched scored views at 2 frames/s for inspection. Labels identify every setting; read left to right, then top to bottom.

Inspect matched keyframes at full resolution

Keyframe 1 / 12

Full generated videos · synchronized controls

Release · camera on

Mix resume · 1200

No camera trained · 600

No proxy · 1000

No proxy · 1900

Release · camera off (inference)