Expand description
Screen-space neural GI from a G-buffer and a lighting probe.
The path tracer is the teacher. At inference the network sees first-hit albedo, normal and depth, a multi-bounce lighting probe, quarter-res probe context, DDGI-lite world bins, and encoded hit positions. Heads: Direct irradiance residual, 7×7 kernel gather, hybrid (KPN + capped residual), hops (à-trous light jumps + world-grid teleport), plus an optional NRC-style 1×1 MLP for per-pixel world residuals. The training family includes jittered furnace-like shells so hops learn uniform-bright identity.
Built from conv2d / relu / concat / a fixed 2× pool, so the same
graph trains on Metal or CUDA and runs on the CPU the browser already has.
Re-exports§
pub use dataset::Dataset;pub use ddgi::bounds_from_world;pub use ddgi::encode_frame;pub use ddgi::encode_world_planes;pub use ddgi::DdgiContext;pub use ddgi::DDGI_GRID;pub use ddgi::DDGI_GRID_Y;pub use ddgi::DDGI_VIS_PLANE;pub use infer::ProbeGi;pub use infer::ProbeNet;pub use model::init_params;pub use model::Head;pub use model::ProbeArch;pub use model::Widths;pub use model::ALBEDO_FLOOR;pub use model::GRAPH_ROOM_TINT;pub use model::HOPS_IRRADIANCE_CAP;pub use model::HYBRID_IRRADIANCE_CAP;pub use model::IN_CHANNELS;pub use model::OUT_CHANNELS;pub use nrc::fuse_beauty;pub use nrc::fuse_beauty_with;pub use nrc::init_params as init_nrc_params;pub use nrc::load_nrc;pub use nrc::load_nrc_bytes;pub use nrc::mean_albedo;pub use nrc::nrc_apply_mask;pub use nrc::nrc_eligible;pub use nrc::nrc_features;pub use nrc::nrc_input_channels;pub use nrc::pack_nrc_batch;pub use nrc::pack_nrc_batch_with;pub use nrc::probe_colour_relative_sq;pub use nrc::save_nrc;pub use nrc::train_nrc;pub use nrc::train_nrc_dataset;pub use nrc::train_nrc_from;pub use nrc::NrcArch;pub use nrc::NrcNet;pub use nrc::HOPS_NRC_CAP;pub use nrc::HOPS_NRC_PROBE_CONFIDENCE;pub use nrc::NRC_CAP;pub use nrc::NRC_IN;pub use nrc::NRC_IN_DDGI;pub use nrc::NRC_MAX_PROBE;pub use nrc::NRC_MIN_ALBEDO;pub use nrc::NRC_OUT;pub use nrc::NRC_PROBE_CONFIDENCE;pub use pack::colour_planes;pub use pack::geometry_hits;pub use pack::planes;pub use pack::PlaneExtras;pub use pack::DEPTH_PLANE;pub use pack::SKY_DEPTH_COMPRESSED;pub use train::Batch;pub use train::TrainConfig;pub use train::Trainer;pub use train::GI_LOSS_BOOST;pub use train::HOPS_CHROMA_LOSS_BOOST;pub use train::HOPS_GI_LOSS_BOOST;pub use train::HOPS_NEUTRAL_LOSS_BOOST;pub use train::LOSS_EPSILON;pub use train::TRAIN_EPSILON;
Modules§
- checkpoint
- Trained weights on disk.
- dataset
- Training tiles on disk.
- ddgi
- DDGI-lite: bin probe radiance into a coarse world-space grid, then broadcast each cell back to every pixel that falls in it — cheap distant context without extra rays.
- infer
- Running a trained network: forward graph only.
- live_
cornell - Browser-side Cornell box: first-hit G-buffer + a cheap multi-bounce probe.
- model
- A guided reconstructor: probe + G-buffer + demodulated irradiance in, reconstructed colour out.
- nrc
- NRC-style tiny MLP: per-pixel world features → irradiance residual.
- pack
- Pack probe + G-buffer + reference into the planar layout the network reads.
- train
- Training: relative L2, autodiff, Adam.
Functions§
- compress
- Compress unbounded radiance into
[0, 1)for the convolution stack. - expand
- Inverse of
compress. Values at 1 (compressed infinity) stay large rather than dividing by zero. - preferred_
device - The device this crate actually trains and infers on.