YHLUO DEMO / MUNDO INFERENCE

The inference pipeline, made ready to ship.

Context photographs, a 3D proxy and a few camera keyframes go in; a collision-free camera path, the right context photos per chunk, a written prompt and the generated video come out. This page shows what changed in speed and what it costs in fidelity, measured against real photographs and against the published videos.

Sampling: the same chunk under every setting

One 124-frame chunk of the office case, eight A800 GPUs in tensor parallel, compared with the 50-step Euler sampling the page videos were made with. The latent error is the relative distance of the final latent from the reference; sharpness is the gradient energy of the frames relative to the reference (below 1 = softer).

Setting	s / chunk	latent error	PSNR	sharpness
Euler · 50 steps (reference)	195	0.000	138.1	1.000
Euler · 32	125	0.026	38.3	0.957
DPM++ 2M · 32	125	0.024	38.6	0.981
DPM++ 2M · 24	94	0.037	36.4	0.962
DPM++ 2M · 32 · reference tokens reused on every second step	86	0.035	37.0	0.974
DPM++ 2M · 24 · reused	64	0.049	33.5	0.951
Euler · 16	62	0.058	34.5	0.924

zoomed crops of the fast settings — The same region of frame 110 under the reference and the faster settings (pixel-doubled).

Held-out photographs: which context photos, which prompt

A camera path is laid through a few consecutive real photo poses; those photos and their neighbours are removed from the pool, so the frames at those poses can be compared with the real photographs (50-step sampling, same seed everywhere). Metrics are averaged over the key frames inside the photographed area.

Variant	PSNR	SSIM	LPIPS	cases
new selection · captioned prompt	14.43	0.437	0.409	10
new selection · captioned · averaged checkpoints 1400/1500/1600	14.51	0.437	0.407	10

auditorium a

5 held-out photo poses on the path, 121 of 128 photos left in the pool. Frame 62:

Variant	PSNR	SSIM	LPIPS
new selection · captioned prompt	12.68	0.336	0.636
new selection · captioned · averaged checkpoints 1400/1500/1600	12.75	0.336	0.637

bedroom a

8 held-out photo poses on the path, 277 of 300 photos left in the pool. Frame 70:

Variant	PSNR	SSIM	LPIPS
new selection · captioned prompt	11.96	0.400	0.634
new selection · captioned · averaged checkpoints 1400/1500/1600	11.96	0.399	0.633

church a

5 held-out photo poses on the path, 121 of 128 photos left in the pool. Frame 62:

Variant	PSNR	SSIM	LPIPS
new selection · captioned prompt	16.21	0.404	0.294
new selection · captioned · averaged checkpoints 1400/1500/1600	16.23	0.400	0.291

classroom a

6 held-out photo poses on the path, 288 of 300 photos left in the pool. Frame 74:

Variant	PSNR	SSIM	LPIPS
new selection · captioned prompt	8.45	0.398	0.755
new selection · captioned · averaged checkpoints 1400/1500/1600	8.49	0.399	0.753

courtroom a

5 held-out photo poses on the path, 57 of 64 photos left in the pool. Frame 62:

Variant	PSNR	SSIM	LPIPS
new selection · captioned prompt	12.77	0.302	0.480
new selection · captioned · averaged checkpoints 1400/1500/1600	12.72	0.298	0.484

lighthouse a

5 held-out photo poses on the path, 121 of 128 photos left in the pool. Frame 62:

Variant	PSNR	SSIM	LPIPS
new selection · captioned prompt	19.00	0.570	0.205
new selection · captioned · averaged checkpoints 1400/1500/1600	19.06	0.571	0.205

msm a

5 held-out photo poses on the path, 55 of 96 photos left in the pool. Frame 62:

Variant	PSNR	SSIM	LPIPS
new selection · captioned prompt	17.48	0.470	0.248
new selection · captioned · averaged checkpoints 1400/1500/1600	17.45	0.468	0.250

museum a

5 held-out photo poses on the path, 121 of 128 photos left in the pool. Frame 62:

Variant	PSNR	SSIM	LPIPS
new selection · captioned prompt	12.86	0.191	0.317
new selection · captioned · averaged checkpoints 1400/1500/1600	13.43	0.191	0.306

office a

8 held-out photo poses on the path, 286 of 300 photos left in the pool. Frame 70:

Variant	PSNR	SSIM	LPIPS
new selection · captioned prompt	16.50	0.675	0.241
new selection · captioned · averaged checkpoints 1400/1500/1600	16.46	0.673	0.240

office b

8 held-out photo poses on the path, 280 of 300 photos left in the pool. Frame 70:

Variant	PSNR	SSIM	LPIPS
new selection · captioned prompt	16.38	0.627	0.285
new selection · captioned · averaged checkpoints 1400/1500/1600	16.55	0.634	0.274

Project-page cases, re-rendered

Case	sharpness new / published	artifact score new / published	largest seam jump new / published	new vs published PSNR / SSIM / LPIPS
Alameda house	0.99	1 / 0	2.2 / 1.8	20.7 / 0.72 / 0.14
Bedroom	1.08	0 / 0	2.5 / 2.2	18.8 / 0.70 / 0.18
Mission church	1.28	3 / 3	1.4 / 1.4	16.3 / 0.40 / 0.33
Mont-Saint-Michel	1.28	0 / 0	1.9 / 1.9	18.4 / 0.64 / 0.27
Rotunda museum	1.14	1 / 1	3.2 / 2.8	16.0 / 0.39 / 0.26
Office	0.95	0 / 0	1.7 / 1.7	16.5 / 0.68 / 0.22

Sharpness is the gradient energy of the new video over the published one. The artifact score is the sum over chunks of a vision-language judge's 0–3 rating of four frames per chunk, run identically on both videos. The seam jump is the frame change at a chunk boundary relative to the typical change, both normalised by the proxy video's own motion. The last column compares the two renderings with each other (different seeds, contexts and prompts), so it measures how close they are, not which is right.

Office

Left: the published video (50-step Euler, hand-picked contexts, Codex-written prompt). Right: the new pipeline on the same camera path (32 steps dpmpp2m, automatic contexts and captions): 360 s of sampling for 3 chunks.

Bedroom

Alameda house

Rotunda museum

Mission church

Mont-Saint-Michel

End to end

The office case above (three chunks, 304 frames, a path planned from six keyframes) timed on one eight-GPU node, from the request to the finished video.

Mode	seconds	What the time goes to
Published pipeline (reference)	877	50-step Euler; loading the model the original way (175 s), encoding conditions in a separate process, 200 s per chunk; another 568 s the first time a node copies the weights
Cold start	650	nothing resident: 82 s to load the prompt and encoding models, 73 s of preparation (path 2, contexts 8, proxy 13, prompts 16, files 5, conditions 31), 26 s to load the video model, 357 s of sampling at 32 steps DPM++ 2M, 45 s of chunk hand-over and decoding
Resident services	481	prompt/encoding service and the eight-GPU sampler stay loaded: 45 s until the first chunk is handed to the sampler (path 2, contexts 10, proxy 13, prompts 11-16, files 6), then sampling of chunk 0 overlaps the encoding of chunks 1 and 2; 432 s of sampling and hand-over
Cold start, --fast	559	as the cold start, with the reference tokens reused on every second sampling step and the text tokens frozen: 248 s of sampling (80/84/84 s per chunk) instead of 357 s

Notes

Model: the same v3-mix step-1500 LoRA on MiniMax-H3 as the project page; 640×480, 24 fps, chunks of 124 frames with 34 overlapping.
Loading: the LoRA is merged once and the weights stored per GPU, so the eight-GPU model loads in about 12 s instead of 175 s and needs 13 GB per GPU instead of 39 GB; the outputs match the original loader within numerical noise.
Context photos are chosen per chunk by a per-pixel score (visibility in the proxy, resolution and viewing angle relative to the target camera, sharpness), keeping a second view of every surface where one exists.
Prompts are written by a local Qwen3-VL-8B from the photos and the planned camera moves; the camera moves themselves are computed from the path.