Inference Performance

This report provides practical runtime and GPU-memory references for running 4DAnyone on common GPUs.

Turbo vs. Base

Generation Quality

The 4-step Turbo model achieves quality comparable to a 50-step Base reference:

Model	Denoising steps	PSNR ↑	SSIM ↑	LPIPS ↓
4DAnyone-Turbo	4	25.48	0.895	0.072
4DAnyone-Base	50	25.69	0.899	0.073

Denoising Speedup

For the 6-view example, Turbo uses 4 denoising steps and Base uses 24. The values show Turbo's speedup over Base on the same GPU and attention backend.

GPU	SDPA	FlashAttention-3	SageAttention
H200	5.49×	5.35×	5.12×
H20-3E	5.83×	5.78×	5.73×
RTX 5090 D v2	5.77×	—	5.70×
RTX 4090	5.85×	—	5.78×
RTX 3090	5.85×	—	5.80×
RTX 5880 Ada	5.79×	—	5.80×
RTX A6000	5.77×	—	5.75×

4DAnyone-Turbo Performance

The performance tables below use the following terms:

• SDPA, FlashAttention-3, and SageAttention: denoising time for each attention backend.
• Peak CUDA allocated: maximum live tensor memory during generation.
• GPU memory: total memory capacity per GPU.
• —: backend or workload unavailable on that GPU.

6-View Full Orbit

The 6-view full-orbit benchmark uses the following command:

python inference.py \
    --video_path "data/source/pexels/2785536-uhd_2160_3840_25fps.mp4" \
    --views_per_layer 6

GPU	SDPA (min)	FlashAttention-3 (min)	SageAttention (min)	Peak CUDA allocated (GB)	GPU memory (GB)
H200	0.97	0.76	0.80	24.06	139.81
H20-3E	3.21	2.54	2.09	24.06	139.81
RTX 5090 D v2	2.06	—	1.61	22.11	23.40
RTX 4090	2.71	—	2.13	22.18	23.52
RTX 3090	5.48	—	4.58	21.55	23.56
RTX 5880 Ada	3.47	—	2.99	24.06	47.37
RTX A6000	3.78	—	3.69	24.06	47.53

24-View Full Orbit

The 24-view full-orbit benchmark runs 4DAnyone-Turbo on a single GPU:

python inference.py \
    --video_path "data/source/pexels/2785536-uhd_2160_3840_25fps.mp4" \
    --views_per_layer 24

For 24-view generation, the pipeline first denoises six RCP views to use as references and then denoises the 24 target views. The table reports the total time for both stages.

GPU	SDPA (min)	FlashAttention-3 (min)	SageAttention (min)	Peak CUDA allocated (GB)	GPU memory (GB)
H200	5.11	3.90	3.90	24.06	139.81
H20-3E	17.66	13.86	11.24	24.06	139.81
RTX 5090 D v2	11.22	—	—	21.94	23.40
RTX 4090	16.20	—	—	22.18	23.52
RTX 3090	31.98	—	—	21.94	23.56
RTX 5880 Ada	18.84	—	16.14	24.06	47.37
RTX A6000	20.55	—	19.98	24.06	47.53

4DAnyone-Base Performance

The 6-view 4DAnyone-Base benchmark uses 24 denoising steps:

python inference.py \
    --video_path "data/source/pexels/2785536-uhd_2160_3840_25fps.mp4" \
    --views_per_layer 6 \
    --enable_turbo=False

GPU	SDPA (min)	FlashAttention-3 (min)	SageAttention (min)	Peak CUDA allocated (GB)	GPU memory (GB)
H200	5.33	4.05	4.09	24.06	139.81
H20-3E	18.68	14.70	11.98	24.06	139.81
RTX 5090 D v2	11.87	—	9.17	22.11	23.40
RTX 4090	15.86	—	12.29	22.18	23.52
RTX 3090	32.03	—	26.52	21.55	23.56
RTX 5880 Ada	20.10	—	17.36	24.06	47.37
RTX A6000	21.82	—	21.20	24.06	47.53

Preprocessing Performance

GVHMR and skeleton extraction run before denoising and are shared by both models. These 6-view timings can vary with CPU and storage performance.

• GVHMR: tracking, ViTPose, image-feature extraction, and motion prediction.
• Skeleton extraction: target-view skeleton-condition rendering.

GPU	GVHMR (min)	Skeleton extraction (min)
H200	1.20	1.25
H20-3E	1.08	1.27
RTX 5090 D v2	0.72	0.82
RTX 4090	0.72	0.73
RTX 3090	1.66	1.72
RTX 5880 Ada	0.98	1.00
RTX A6000	1.57	2.12