3DGS
§ 02Models
| Model | Judged | Tier | Self-best | Rounds | Hours | API cost | |
|---|---|---|---|---|---|---|---|
| Claude Opus 5.5 | 5.82× | N1 | 5.85× | 20 | 10.0 h | $65.4 | trace → |
| Claude Sonnet 5.5 | 4.44× | N1 | 4.06× | 20 | 10.0 h | $56.5 | trace → |
| GPT-6.2 Astra | 3.58× | N1 | 3.54× | 57 | 10.0 h | $97.4 | trace → |
| MiniMax M3.1 Flash | 2.54× | N1 | 1.60× | 26 | 10.0 h | $5.47 | trace → |
| GPT-6.2 Sol | 2.35× | N1 | 2.35× | 48 | 10.0 h | $22.3 | trace → |
| Claude Haiku 5.5 | 2.35× | N1 | 2.34× | 39 | 10.0 h | $58.8 | trace → |
| GPT-6 Astrasuspected nerf | 2.04× | N1 | 2.07× | 22 | 10.0 h | $97.5 | trace → |
| GLM-5.3 | 1.76× | N1 | 1.74× | 21 | 10.0 h | $55.2 | trace → |
| DeepSeek V4.1 Flash-1017 | 1.71× | N1 | 1.72× | 52 | 9.7 h | $4.06 | trace → |
| Kimi k3.1-pro | 1.68× | N1 | 1.78× | 32 | 10.0 h | $39.6 | trace → |
§ 03Key levers
| Model | Judged | Smaller tiles + rewritten binning | Bypass autograd, activations folded into raster kernel | Lazy Adam (update only visible Gaussians) | Fused loss / SSIM | Other main levers |
|---|---|---|---|---|---|---|
| Claude Opus 5.5 | 5.82× | ✓ 8×8, later 8×4; keys generated per slot, tested by bounding-box range; with the new forward / backward kernels 1.23 → 2.01 | ✓ GS_FAST_PIPE=1: whole iteration in one CUDA file, activations and chain rule done in per-Gaussian kernels, no autograd, one sync per step; kernels per step 144 → ~30, 2.01 → 3.02 | ✓ GS_LAZY_ADAM=1, all param groups (positions also made lazy in round 18); 3.42 → 3.68; after folding pending grads into one update and looping 8 rows per thread 4.12 → 4.49 | SSIM forward's three convolutions merged into one pass, ground-truth mean / variance cached (57 → 38 µs); loss grad fusion tried, no gain | branch-free, barrier-free render and key kernels (render 376 → 208 µs, 4.76 → 5.21), shuffle-free backward, Gaussians reordered in Morton order GS_FAST_SORT=500, depth sort only for visible Gaussians, fused densification, fixed-point atomic adds (training bitwise reproducible) |
| Claude Sonnet 5.5 | 4.44× | ✓ 8×8 + exact tile enumeration via two-pass warp cooperation (no padding entries in the sort); 8×8 alone nets 0, with enumeration 1.25 → 1.52 | partial: GS_OPT_RAWRENDER=1 has rasterization take raw params, activations and their backward folded into the kernel (1.52 → 1.86); training loop and raster calls still in PyTorch | ✓ SH only (48 of 59 floats) GS_OPT_LAZYSH=1, 3.31 → 3.51; position / opacity / scale / rotation stay dense | ✓ GS_OPT_FASTLOSS=1 loss grad without building an autograd graph (2.56 → 2.62); rewritten SSIM kernel (forward 100 → 44 µs, 3.14 → 3.20) | custom single-launch Adam (2.20 → 2.56), SH update on a low-priority stream concurrent with the binning chain (3.81 → 4.06), sort by depth then stable sort by tile (4.06 → 4.24), backward passes only scalar color accumulators (→ 4.35), single-instruction exp / rcp (→ 4.51), fixed-point atomic adds |
| MiniMax M3.1 Flash | 2.54× | ✓ 8×8 alone flat (1.59 vs 1.60); 1.76 with tight bounding boxes; 1.98 with one warp per Gaussian in binning | ✗ only quaternion normalization made into one bitwise-identical kernel GS_FUSED_NORMALIZE=1 (2.21 → 2.40) | ✗ | last commit fused L1 and its grad (2.403 → 2.410, per commit message); SSIM untouched | bitwise-identical fused Adam, desync, backward per-pixel state packed into float4 (inner loop 176 → 129 instructions, 1.43 → 1.62), unused depth path off (→ 2.06), contiguous SH grad writes (→ 2.21) |
| Claude Haiku 5.5 | 2.35× | ✗ stays 16×16; instead warp-level culling in forward (1.92 → 2.08), binning with one thread per (Gaussian, tile) slot (1.73 → 1.86) | ✗ only normalization made into one kernel GS_FUSED_NORMALIZE=1 (1.55 → 1.63) | ✗ dense; rest-SH Adam folded into backward preprocess GS_FUSED_REST_ADAM=1 (1.86 → 1.91) | two SSIM kernel attempts (16×16 blocks, redundant flushes removed), both within noise, not committed | --use_fast_math (1.35 → 1.43), custom dense Adam, coalesced SH grad writes (1.63 → 1.73), forward color and inverse depth in shared memory (+5.2%), backward visits only possibly contributing pixels, tiles launched in descending length order |
| GPT-6 Astrasuspected nerf | 2.04× | ✓ only in round 20: GS_OPT_TILE_SHAPE=8x8 (keeps the 16×16 parent-tile coverage test, 166.2 → 161.2 s); round 21 compacted sort input GS_OPT_COMPACT_SORT=1 (161.2 → 150.5 s) | ✗ activations via torch.compile GS_OPT_COMPILED_ACTIVATIONS=1 (234.7 → 225.2 s; of the first two versions one errored, one only 1.08×) | ✗ dense Adam in Triton (211.9 → 196.4 s) | ✗ not in the log | warp-cooperative binning for large rectangles (225.2 → 216.0 s), cooperative staging of backward pixel state (187.7 → 170.6 s), background term pre-folded, depth skipped, memory cache kept |
| GLM-5.3 | 1.76× | ✗ tile size unchanged; last round tightened candidate rectangles GS_OPT_TIGHT_RECT=1 (radix sort −39%); binning one warp per Gaussian (1.55 → 1.58) | ✗ activations as a pair of standalone kernels outside rasterization GS_OPT_ACT_FUSE=1 (1.60 → 1.70) | ✗ dense; SH grad and Adam fused into one kernel GS_OPT_SH_ADAM=1 (first version regressed to 1.28, after three reorderings 1.34 → 1.46) | only clamp backward fused (−11 µs / step); SSIM only changed to not zero its output | depth skipped, backward pixel state in shared memory (1.46 → 1.53), four small param groups in one Adam kernel, __expf, densification stats kernel (+0.003) |
| DeepSeek V4.1 Flash-1017 | 1.71× | tried, reverted: round 3 switched tiles straight to 8×8, 1.230 vs 1.248 (binning and sort get costlier with tile count); never returned to this direction | activations ✓ folded into preprocess (FUSED_ACT on by default, 1.354 → 1.433); autograd not bypassed | ✗ log states "state decay needed even for zero grads, the 68% cull rate can't be exploited" | ✗ | custom float4 dense Adam (1.26 → 1.35), __expf, forward color staging (kernel −15%), backward per-pixel state as float4 (kernel −16%), two warps per block in backward, bucket boundary state merged into one float4 |
| Kimi k3.1-pro | 1.68× | ✗ tile size not in the log; binning only hoisted loop invariants (184.3 → 179.4 s) | ✗ activation fusion not in the log | ✗ dense; custom single-kernel Adam only in round 29 (177.8 → 172.5 s) | tried, reverted: SSIM scalar backward 0.8 s slower, three revisions crashed within 3000 steps; L1 backward indistinguishable across two timings | depth skipped (247.5 → 237.8 → 231.3 s), black-background specialization, sample state and pixel state packed into float4, shared forward color, cooperative prefetch of backward state, __expf |