We provide multiple checkpoints with different Multi-HMR encoders on HuggingFace, including ViT-672S/B/L and ViT-896L. Once downloaded you need to place them into the src directory.

# Download multiple Human3R checkpoints
huggingface-cli download faneggg/human3r human3r_672S.pth --local-dir ./src
huggingface-cli download faneggg/human3r human3r_672B.pth --local-dir ./src
huggingface-cli download faneggg/human3r human3r_672L.pth --local-dir ./src
huggingface-cli download faneggg/human3r human3r_896L.pth --local-dir ./src

Here is an evaluation of their Accuracy vs. Speed Trade-off:

	Local Human			Global Human			Runtime
Models	PA-MPJPE↓	MPJPE↓	PVE↓	WA-MPJPE↓	W-MPJPE↓	RTE↓	FPS↑
Human3R w/ ViT-S/672	56.1	87.8	103.1	129.9	314.2	2.2	15
Human3R w/ ViT-B/672	49.3	79.6	94.3	122.1	292.9	2.2	11
Human3R w/ ViT-L/672	48.5	83.1	96.7	113.6	291.7	2.2	7
Human3R w/ ViT-L/896	44.1	71.2	84.9	112.2	267.9	2.2	5

Benchmark Setup: All reported speeds are measured on an NVIDIA RTX 4090 GPU with dual Intel Xeon Gold 6530 CPUs.

Real-Time Tier: ViT-S (15 FPS) offers a strong balance for global motion estimation (WA-MPJPE 129.9, RTE 2.2).

High-Fidelity Tier: ViT-L (5-7 FPS) provides more detailed human-mesh reconstruction (WA-MPJPE 112.2, RTE 2.2), suitable for application requiring fine-grained pose and shape.

Inference Speed

Additionally, we benchmark runtime across diverse datasets and backbones. As shown below, the ViT-S/672 variant indeed supports real-time applications (~15 FPS), while larger models trade speed for detail:

	FPS
Models	3DPW (288×512)	BEDLAM (512×288)	RICH (512×368)	EMDB (384×512)	Bonn (512×384)	TUM-D (512×384)
Human3R w/ ViT-S/672	15.87	15.64	14.28	13.75	13.65	13.59
Human3R w/ ViT-B/672	13.33	12.69	11.89	11.68	11.67	12.41
Human3R w/ ViT-L/672	9.17	8.73	8.38	8.27	8.27	8.61
Human3R w/ ViT-L/896	5.38	5.3	5.15	5.09	5.06	5.15

We provide a script to evaluate the model inference speed (forward pass only), excluding data loader/saver operations.

Note: It is recommended to use inference_only.py instead of eval/global_human/launch.py for timing purposes, as the evaluation script includes significant overhead from ground-truth data processing, metric computation, visualization, and I/O operations.

    # Example:
    CUDA_VISIBLE_DEVICES=0 python inference_only.py \
        --model_path src/human3r_672S.pth --size 512 \
        --seq_path /path/to/3DPW/imageFiles/downtown_runForBus_00 \
        --use_ttt3r --reset_interval 100