GenLit_ICLight_Relighting

GenLit IC-Light Relighting — raw WSG samples

Raw (pre-VAE) samples for the WSM v1 IC-Light relighting pilot (iclight_relighting_v3). Each sample is a self-describing named-key .npz: a source appearance, an IBL light-probe condition, and the relit target. The matching FLUX.2 latents live in the companion latent repo World-Snapshot/WSM_v1_Latent_PT_WSG_v1.8_for_Flux2 (custom_pilots/iclight_relighting_v3/cache_named_v2/).

Source

Derived from the public GenLit dataset (gbmpi/genlit_dataset): Objaverse objects rendered under Poly Haven HDRI environments with a single moving point light. Within a GenLit scene the point-light direction is fixed while its intensity ramps 0 → 30 → 60 → 90 → 100 → 120 and then plateaus at 120 (as the point light rises the environment intensity falls 1.0 → 0.2); after the ramp the direction drifts. Frame 0 is environment-only (point intensity 0).

How each sample is produced

For a chosen scene/frame we render one npz with three named modalities (all uint8, 480×480, H×W×3; the trainer maps pixels with x/127.5 − 1):

The task: given the source appearance + the probe describing the target lighting, produce the relit image. The probe carries all per-sample lighting and the source carries object identity/geometry, so no text caption is used (matching the official inverse_rendering heads).

Layout

named/          base split      — 1 mid-trajectory frame per scene (intensity 120)   10,274 samples / 411 objects
named_refill/   intensity aug   — per scene, 2 random distinct levels sampled from
                                  {0,30,60,90,100,120} at the scene's fixed direction  20,548 samples
hdris/          the Poly Haven HDRIs (2k .hdr) used to render the probes

Total 30,822 samples across 411 objects. The refill split exists because the base split's frame is always on the intensity-120 plateau; the refill adds same-object/same-direction, varying-intensity pairs so the model learns intensity control.

Packaging

Files are shipped as uncompressed multi-part tars GenLit_ICLight_Relighting_partNNN.tar (each < 15 GB; the .npz are already zlib-compressed, so no zstd). ..._parts_manifest.jsonl lists each part's file count, byte size and SHA-256; ..._parts_summary.json has the totals.

Extract

mkdir -p data/custom_pilots/iclight_relighting_v3
for f in GenLit_ICLight_Relighting_part*.tar; do
  tar -xf "$f" -C data/custom_pilots/iclight_relighting_v3/
done
# -> data/custom_pilots/iclight_relighting_v3/{named,named_refill,hdris}/

License

MIT for the packaging/processing. Underlying assets follow their original terms (Objaverse; Poly Haven HDRIs are CC0; GenLit).