--- license: mit tags: - photon-memory - rendering - density-estimation - test-time-training library_name: pytorch --- # Photon Memory checkpoints (Rough Flower / PureLight comparison) Companion weights for the Photon Memory reader that matches / beats PureLight on directional NLL for the released Rough Flower luminaire. Code: [ANABUR920/renderTTT](https://github.com/ANABUR920/renderTTT) (`lightmem/`). ## Files | Path | Role | Size | |------|------|------| | `checkpoints/v3_ksched_model.pt` | Meta-trained PhotonKernel reader (`k_feat=True`) | ~0.7 MB | | `checkpoints/big_1e8_k512_ttt.pt` | Reader-only TTT on 1e8 memory (`k=512`) | ~0.7 MB | | `checkpoints/big_match_res2b_best.pt` | Intermediate matched TTT + residual (~NLL −4.54) | ~42 MB | | `checkpoints/big_match_res2c_best.pt` | Final matched TTT + residual (**NLL −6.00**, render default) | ~42 MB | | `logs/*.jsonl` | Training / eval logs for the above TTT runs | small | ## Download ```bash hf download ANABUR/photon-memory --local-dir ./photon-memory-ckpts # Photons (separate dataset): hf download ANABUR/photon-memory-photons --repo-type dataset --local-dir ./photon-memory-photons ``` ## Reproduce the −6.00 NLL claim (same protocol as paper) Needs: CUDA GPU (~40 GB for 1e8 + residual), traced Rough Flower photons (`flower_surface_*` 11e6 held-out + `flowerbig_surface_*` 1e8 memory), and this repo's `scripts/ttt_big.py`. Winning warm-start chain (do not claim a cold start from the meta reader equals −6.00 in one shot): 1. Start from `v3_ksched_model.pt` → reader-only TTT → `big_1e8_k512_ttt.pt` (~−0.53). 2. Joint reader + residual matched TTT → `big_match_res2b_best.pt` (~−4.54). 3. Continue: ```bash python scripts/ttt_big.py \ --data flowerbig_surface --N 100000000 --k 512 --wmult 2 \ --ttt_matched --n_ttt_q 1000000 --steps 8000 --residual \ --res_K 32 --res_log2 19 --res_finest 2048 \ --init_ttt checkpoints/big_match_res2b_best.pt \ --lr 5e-6 --res_lr_grid 5e-3 --res_lr_head 5e-4 \ --name big_match_res2c ``` Zero-shot (no TTT) with `v3_ksched_model.pt` at `N=1e8, k=512` reaches RGB-mean NLL ≈ **−0.460**, matching the PureLight teacher plateau (~−0.46). The −6.00 number is **matched test-time adaptation** on the non-eval remainder of the same 11e6 photon file as the held-out set (memory may be a separately traced 1e8 bank). Leave-one-out TTT on the 1e8 memory alone stalls near −0.53. It is not zero-shot and not a fresh-retrace claim. ## Honest notes for reviewers / friends - Held-out queries are always the last 20k of `default_rng(0).permutation(11_000_000)` on `flower_surface_*`. - `--ttt_matched` trains on the non-eval remainder of that same 11e6 file while retrieving from the separately traced 1e8 memory. - A frozen-reader hash residual alone peaks ~−0.38 and does **not** beat PureLight. - **Photons** for Rough Flower are published separately: [`ANABUR/photon-memory-photons`](https://huggingface.co/datasets/ANABUR/photon-memory-photons) (`flower_surface_*` 11e6 + `flowerbig_surface_*` 1e8, channels mean/r/g/b). PureLight teacher weights and photon maps remain third-party; otherwise re-trace with `scripts/gen_flower.py`. ## Photons (required for reproduction) Rough Flower OptiX traces live in [`ANABUR/photon-memory-photons`](https://huggingface.co/datasets/ANABUR/photon-memory-photons): ```bash hf download ANABUR/photon-memory-photons --repo-type dataset --local-dir ./photon-memory-photons ``` Layout: `flower_surface_{mean,r,g,b}/val/rough_flower.npz` (11e6) and `flowerbig_surface_{mean,r,g,b}/val/rough_flower.npz` (1e8). ## Load in code ```python import torch from lightmem.models import PhotonKernel, ResidualLobes sd = torch.load("checkpoints/big_match_res2c_best.pt", map_location="cpu") reader = PhotonKernel(k=512, k_feat=True) reader.load_state_dict(sd["reader"]) residual = ResidualLobes(K=32, finest=2048, log2_table=19) residual.load_state_dict(sd["residual"]) # sd["nll"] ≈ -6.00, sd["step"] == 8000 ```