RLSplat

This model repository contains the inference-only RLSplat checkpoint in Safetensors format. Try it in the interactive Hugging Face demo with 2–4 images of the same scene. Training code will be released in a future update.

Checkpoint	Training data	Training context views	Folder
RLSplat	RE10K	dynamic 2–24 (maximum 24)	re10k-24view

The checkpoint was exported from the submitted 5,000-step RE10K Lightning checkpoint. Optimizer state, training metadata, and training-only tensors are not included.

Use the model with the RLSplat evaluation code:

python -m src.main '+evaluation=re10k' \
  checkpointing.load=hf://Knowing/RLSplat/re10k-24view/model.safetensors \
  'dataset.re10k.roots=[/path/to/re10k]' \
  dataset.re10k.view_sampler.index_path=assets/evaluation_index_re10k.json

See the repository README for installation, data preparation, other context-view selections, and full budget-sweep examples.