experiments/realhumans_h3d_diffusion_span360/README.md

RealHumans Human3Diffusion 360-Degree Reconstruction

Overview

This experiment fine-tunes the official Human3Diffusion 3D reconstruction model on the RealHumans turntable dataset. It preserves the original diffusion formulation while replacing four MVD-generated pseudo views during training with four real views from the same subject.

Each training item contains:

• one clean frontal context image;
• four real target views distributed around the full 360-degree orbit;
• Gaussian-noised VAE latents for those four views at a sampled DDPM timestep;
• four held-out real views used for RGB, LPIPS, and Gaussian regularization losses.

At strict single-view evaluation time, the official pretrained MVD model generates the four auxiliary views from one real frontal image. The fine-tuned 3D reconstructor then predicts a 3D Gaussian representation and renders the held-out cameras.

Data

• Dataset: RealHumans, 10,078 subjects, 81 frames per subject.
• Training split: the first 10,000 sorted subjects.
• Test split: the final 78 sorted subjects.
• Training sampling: frame 0 is the clean frontal context; four additional views are sampled around approximately 72, 144, 216, and 288 degrees with bounded jitter. Four non-input views supervise rendering.
• Resolution: 256 x 256 model input and 512 x 512 rendered supervision.
• Masks: foreground masks are derived from the image alpha/background convention and are used by the renderer losses.

The RGB archive and Video Depth Anything preprocessing output are stored at the root of the Hugging Face dataset repository. The depth data belongs to the earlier Prometheus experiment and is not consumed by this Human3Diffusion run.

Model And Optimization

• Initialization: official Human3Diffusion 3D reconstructor model_1.safetensors.
• Diffusion prior: ashawkey/imagedream-ipmv-diffusers VAE and DDPM scheduler.
• Trainable component: Human3Diffusion 3D generation/reconstruction network (g_phi).
• Frozen components: VAE and LPIPS network. The pretrained MVD network is not jointly fine-tuned.
• Diffusion input: four noisy VAE latents, camera ray embeddings, one clean context latent, and its ray embedding.
• Output: a 3D Gaussian representation rendered into four held-out target cameras.
• Losses: masked RGB MSE, LPIPS, and Gaussian distortion regularization.
• GPUs: 8.
• Batch size: 8 per GPU, global batch size 64.
• Precision: bfloat16.
• Optimizer: AdamW, learning rate 5e-5, weight decay 0.05.
• Schedule: OneCycleLR.
• Duration: 40 epochs; one checkpoint per epoch.

Strict Single-View Evaluation

Evaluation uses one fixed frontal image per test subject. The official MVD checkpoint generates four pseudo views, followed by 50-step DDIM reverse sampling through the 3D reconstructor. Inputs, predictions, ground truth images, metrics, and fixed subject/view selections are saved with the experiment.

Representative fixed-sample metrics:

Epoch	Full PSNR	Foreground PSNR	SSIM
19	17.1009	8.7132	0.8678
24	17.0899	8.6547	0.8683
29	17.0457	8.6584	0.8676
34	17.0101	8.6155	0.8679
39	17.1096	8.7056	0.8682

These values are nearly flat after epoch 19. Epoch 39 is retained as the final model, while epoch 19 is a reasonable metric-selected alternative.

Reproduction

On the original server:

BASE=/mnt/localssd/colligo/contrib/Mori/collections/foundation_exploration/configs/genrender6/pixel_diffusion/lx_human
cd "$BASE/Human3Diffusion"
source "$BASE/envs/human3diffusion_cuda13/bin/activate"
bash launch_realhumans_h3d_diffusion_span360.sh

Important files:

• train_realhumans_h3d_diffusion.py: distributed diffusion fine-tuning loop.
• core/dataset_realhumans_rgb4.py: deterministic split, 360-degree five-view sampling, and held-out supervision.
• core/models_timeImage_cond_xt_gof.py: diffusion-conditioned 3D Gaussian reconstructor.
• core/attention.py: multi-view attention implementation.
• monitor_realhumans_h3d_singleview_eval.sh: periodic strict single-view evaluation and HF upload.
• eval_singleview_metrics.py: PSNR/foreground-PSNR/SSIM calculation.

Backup Layout

The Hugging Face dataset repository linxin02/human contains:

• root dataset archives and depth preprocessing output;
• code/ source archives;
• docs/ experiment documentation;
• experiments/realhumans_h3d_diffusion_span360/ checkpoints, logs, fixed inputs, metrics, and visualizations;
• earlier 90-, 180-, and 360-degree RGB5 experiments under separate experiment directories;
• earlier Prometheus checkpoints and offline visualizations under experiments/prometheus_stage1_span90/.

The environment directory and downloadable third-party pretrained checkpoints are intentionally excluded. Their identifiers and launch settings are recorded above.