| # Minimal RENI++ Inference |
|
|
| This directory evaluates the thesis RENI++ illumination prior with a small |
| CPU environment containing PyTorch, NumPy and Pillow. It does not install |
| Nerfstudio, tiny-cuda-nn, COLMAP, or the full `ns_reni` package. |
|
|
| Download the `minimal` release group from the repository root: |
|
|
| ```bash |
| python scripts/download_models.py model-storage/reni --group minimal |
| ``` |
|
|
| Then render a deterministic random sample on CPU: |
|
|
| ```bash |
| cd model-storage/reni/minimal |
| uv run render.py --weights decoder.pt --output-dir render |
| ``` |
|
|
| The release directory is self-contained: it includes the decoder artifact, |
| reference implementation, rendering example, Apache 2.0 licence and locked |
| CPU environment. From the source-tree copy of this example, pass the |
| downloaded `model-storage/reni/minimal/decoder.pt` path explicitly. |
|
|
| The command writes: |
|
|
| - `environment_linear_hdr.pt`, a `[H, 2H, 3]` linear-HDR tensor; |
| - `environment_preview.png`, a display-tonemapped preview; |
| - `latent.pt`, the `[100, 3]` latent used for the render. |
|
|
| The output is in the fixed luminance gauge used to train the two-bracket |
| model. Apply a scalar exposure in linear HDR when absolute scene exposure is |
| required. |
|
|
| `reni_decoder.py` is a release-specific reference implementation. It includes |
| the learned joint Vector Neuron frame, its Gram-Schmidt orthonormalisation, the |
| directional encoding, attention decoder, and two-bracket HDR reconstruction. |
| The full Nerfstudio checkpoint remains the source for continued training and |
| for reproducing analyses of the learned training latents. |
|
|
