modelbeast/BENCHMARKS.md

# MODELBEAST Benchmarks (M3 Ultra, 256GB)

First measurements on this machine, recorded 2026-07-12. Fixtures are synthetic
(Blender-rendered Suzanne), so quality numbers are not representative of real
photography — these validate that the pipeline *runs* and how fast.

## Scan track (fully validated, no gated weights)
| Stage | Input | Settings | Result |
|---|---|---|---|
| `colmap_poses` | 48 frames @ 800×600 | sequential matcher, global (GLOMAP) mapper, OPENCV, CPU SIFT | **48/48 images registered, 1479 points, 0.60px mean reprojection error**; global mapper step ~1.0s; full job a few seconds |
| `brush_train` | above colmap_dataset | 1500 steps, max_res 800, sh 2 | ~30–60s wall, 465KB splat.ply, renders in the in-app SplatViewer |

A full-quality `brush_train` run is 30000 steps (the default) — expect minutes,
and a much crisper splat than the 1500-step preview above.

## Image generation — local FLUX lineup (M3 Ultra, mflux/MLX, 1024×1024, validated 2026-07-13)

All five installed and generating. **Warm** = weights cached (the real per-image cost):

| Model | Steps | Warm gen | Peak MLX mem | Gated? | Best for |
|---|---|---|---|---|---|
| **FLUX.2 Klein 4B** | 4 | **9.1s** | 18.0 GB | no (Apache) | volume / sprites — the default |
| FLUX.1 schnell-4bit | 4 | 18.5s | 19.1 GB | no (community quant) | fast draft, FLUX.1 look |
| **FLUX.2 Klein 9B** | 4 | **18.7s** | 28.4 GB | yes | **hero assets — best object accuracy of the whole lineup, cloud included** |
| FLUX.1 schnell | 4 | 20.4s | 25.0 GB | yes | fast draft (Klein 9B beats it at same speed) |
| FLUX.1 dev | 25 | 108.1s | 25.1 GB | yes | cinematic mood / DoF when you can wait ~2 min |

First-run downloads (one-time): schnell/dev ~31GB & ~16–19 min each, Klein 9B ~32GB, Klein 4B ~15GB. Believed among the first published M3 Ultra mflux FLUX.1/FLUX.2 numbers.

**Sweet spot: Klein 9B.** Same ~19s as schnell but far better object coherence → it dominates schnell. Klein 4B when speed matters (2× faster), dev only when you want the cinematic atmosphere. Note the `hf_xet` chunked downloader fails on these repos ("Unable to parse string as hex hash value") — the operator sets `HF_HUB_DISABLE_XET=1` to force the reliable HTTP path.

**First A/B (same prompt, 2026-07-12):** FLUX.2 Klein 4B local (8s, $0) vs nano-banana via OpenRouter (`google/gemini-2.5-flash-image`, 7.7s, $0.0387 exact-billed). Klein: cleaner product-photo subject. nano-banana: richer scene dressing (books/inkwell/quill, dust motes) + finer engraving detail. Verdict: Klein is the volume workhorse; nano-banana wins on scene storytelling per prompt-adherence expectations (Elo 1154 vs ~1083).

## Mesh-gen (local, validated 2026-07-13)
| Operator | Config | Result |
|---|---|---|
| `sf3d` | image → GLB, MPS, tex 1024 | **~5s**, ~9GB peak, 1.5MB GLB. Fast draft tier — good on solid objects, struggles on thin/open geometry. Needs `OMP_NUM_THREADS=1`+`KMP_DUPLICATE_LIB_OK` (segfaults otherwise). |
| `trellis_mac` | TRELLIS.2-4B, pipeline 1024, tex 2048, MPS | **289s (~4.8 min) generation** + 16s bake, 18.4MB GLB with PBR. SOTA-tier local quality — clean coherent geometry even on a thin-ringed astrolabe (dramatically better than SF3D). First run adds a one-time ~15GB download (~30 min); cached after. Needs `HF_HUB_DISABLE_XET=1` + the OMP guards. |
| `hunyuan3d_mlx` | Hunyuan3D 2.1, native MLX (fp16), shape+PBR, tex 2048, remesh 40k | **260s (~4.3 min) total** (shape 149s + texture 112s), peak 20.2GB, 7.6MB GLB (40k faces, 2048² baseColor+MR PBR). Weights **public — no HF login**. Needs `diffusers`+`fast_simplification` in the venv. **MLX-native → the one local 3D op that runs on M1 Ultra** (trellis_mac's torch-MPS bf16 is unverified there). |
| `bg_remove_local` | RMBG-2.0, MPS, 1024 | seconds; clean transparent cutout. Run before SF3D for a big geometry improvement. Note: keeps original RGB under alpha (upscale before cutout). |

### Head-to-head, same mermaid cutout (M3 Ultra, 2026-07-16)
| | speed | GLB | faces | face/detail quality |
|---|---|---|---|---|
| `trellis_mac` | 318s | 23MB | 175,842 | **sharper** — defined eyes/nose/mouth, individually raised tail scales, vivid colors |
| `hunyuan3d_mlx` | 260s | 7.6MB | 40,000 | softer — melted face, smoothed scales, muted texture |

**Verdict:** at defaults `trellis_mac` wins on quality (crisper face + geometry, richer color); `hunyuan3d_mlx` is faster, ~3× lighter, and the only local 3D op that runs on M1. `sf3d` stays the ~5s draft tier. All free/offline; fal cloud for on-demand SOTA without the local wait.

## Video gen — mlx-video / Wan 2.1 on M3 Ultra (2026-07-17, first numbers)
[Blaizzy/mlx-video](https://github.com/Blaizzy/mlx-video) (native MLX, MIT) at `vendor/mlx-video`. Wan2.1-T2V-1.3B, PyTorch weights → MLX via their converter (56s), raw `.pth` deleted after (MLX copy is what runs; 14GB).

| run | config | time | note |
|---|---|---|---|
| **T2V 3s** | 832×480, 49 frames, 50 steps, `--tiling aggressive` | **809.6s (13.5 min)** | T5 6.5s · **denoise 780.7s (15.6s/step)** · VAE decode 21.7s |
| smoke | 832×480, 5 frames, 2 steps | 11.8s | pipeline sanity |

Quality: genuinely good — photoreal fox walking through snow, coherent across all 49 frames, correct gait, no flicker/melting. Output registered in the library (`wan_first.mp4`).

**⚠️ THE GOTCHA — default `--tiling auto` gets Wan SIGKILLed past ~25 frames.** Verified bisect (2 steps, 832×480): 13f (seq 6240) OK · 25f (seq 10920) OK · **33f (seq 14040) SIGKILL**. It is **not** memory: it dies at **4GB RSS with 195GB free**, and MLX reports `max_buffer_length: 179GB` / working set 239GB on this machine. The `auto` heuristic never enables **temporal** tiling, so the VAE decode explodes. Worse, it dies *after* denoising completes — with defaults you wait 14 min for the expensive part, then lose it at the decode with no error. **Fix: `--tiling aggressive`** (→ spatial=256px, temporal=32f) — 33f then saves fine. `--tiling temporal` does *not* help.

Other mlx-video packaging bugs found (all upstream-PR-worthy): (1) `torch` is required to read the `.pth` T5/VAE but is **undeclared** — conversion dies *halfway*, leaving a plausible-looking broken model dir; (2) `librosa` is a hard dep though only LTX-2's *audio* VAE uses it — drags in ancient numba and **breaks install** on modern Python (should be an optional `[audio]` extra); (3) README documents `python -m mlx_video.wan2.convert` — that module doesn't exist (it's `mlx_video.models.wan_2.convert`).

**Wan2.2-TI2V-5B** (23GB MLX, text+image→video): 704², 49 frames, 40 steps = **337s (5.6 min)** — *faster than the 1.3B* at 832×480 (fewer steps + better-conditioned I2V). Note `Wan2.2-I2V-A14B` is **126GB** (~250GB with conversion) — not attempted.

### ❌ NEGATIVE RESULT: I2V-synthesized orbits do NOT reconstruct in 3D (2026-07-17)
Tested the appealing shortcut — *single image → I2V "camera orbit" → frames → colmap_poses → brush_train* — i.e. synthesize the turntable instead of filming it. Prompted TI2V-5B explicitly for camera orbit around a static subject (parallax being what COLMAP needs), fed the RMBG mermaid cutout on white.

| source | registered images | **3D points triangulated** |
|---|---|---|
| real iPad room video | 34 | **1006** ✅ |
| **synthesized I2V orbit** | 26 | **0** ❌ |

**Zero triangulated points.** COLMAP matched features and guessed poses but could not place a single consistent point in space (hence `Failed to fix Gauge … insufficient number of fixed points: 0`, and a "reconstruction" in 2.4s vs ~60s for real footage). Brush still emitted a 190MB .ply — trained faithfully on geometry-free input, i.e. garbage. **Bigger splat ≠ better.**

**Why:** the video visibly rotates (back of the head by frame 47) but the subject is **not rigid** — the tail dissolves, an orb materialises, the figure morphs into a different object. Video models optimise **temporal plausibility**, not **multi-view consistency**; 3D needs the latter. Each frame is individually gorgeous and mutually incompatible.

**Conclusion: you cannot synthesize your way out of capture.** For video→3D, film the real turntable (→ `bg_remove_local`/CorridorKey → colmap → brush/hunyuan3d_mlx). Use I2V for motion/B-roll, not as a multi-view source. Don't re-run this experiment — it's a property of the objective, not a tuning problem. (A purpose-built multi-view/NVS model — e.g. TRELLIS-class or a camera-controlled LoRA — is the thing that could work; a general I2V model can't.)

## CorridorKey (neural green-screen keyer, MLX) — Ultra tuning (2026-07-16)
Corridor Digital's keyer (`vendor/corridorkey`, 14.4k★) with the native MLX backend (`corridorkey-mlx`, resolved from git — not on PyPI). Benchmarked on a synthetic 12-frame 2048² green-screen set with **exact ground-truth alpha** (RMBG mermaid cutout + soft-alpha stripes/disk over chroma green); hint = 8× downscaled truth. Scores = alpha MAE / soft-IoU vs truth, steady-state after 2-frame warmup.

| config | M3 s/f | M1 s/f | MAE ↓ | IoU ↑ | peak GB |
|---|---|---|---|---|---|
| full 2048 (stock default) | 3.81 | 4.97 | 0.0091 | 0.925 | **28.2** |
| full 1024 | 0.61 | 0.75 | 0.0111 | 0.908 | 3.7 |
| full 512 | 0.25 | 0.31 | 0.0139 | 0.885 | 2.6 |
| tiled 512 (stock: compile forced off) | 3.64 | 4.65 | **0.0082** | **0.932** | 2.5 |
| **tiled 512 + our compile patch** | **2.48** | 4.19 | **0.0082** | **0.932** | **2.3** |

**Findings:** (1) `compile` gains nothing full-frame at 2048 on Ultras (within noise) — the documented "1.5–2×" is a small-res/laptop figure. (2) **Tiled-512 is the QUALITY winner**, not just the memory fallback — best alpha accuracy, model at native tile scale over full-res input. (3) Upstream hard-codes `compile=False` in tiled mode; tiles are fixed-shape so compilation applies — our 1-line patch (`vendor/corridorkey-mlx`, branch `modelbeast`, editable-installed into the app venv on M3+M1) makes tiled **1.47× faster on M3** (3.64→2.48 s/f), 1.11× on M1, output bit-identical. (4) tile 1024 tested worse (quality + speed) — 512 is the sweet spot. (5) full-frame 2048 peaks **28 GB** → 32GB Macs (M1 Max) should run **tiled** (2.3 GB) — which is also the best-quality config anyway.

**Fleet verdict:** best-quality config = `tile_size=512, overlap=64` + our patch: M3 ~0.40 fps, M1 ~0.24 fps at 2048², IoU 0.932, 2.3 GB — runs on every node including the M4 24GB. Throughput mode: full-1024 (1.65/1.34 fps, IoU 0.908). MLX gaps: blue-screen checkpoint + despill/despeckle not yet on MLX (torch backend covers those). License: CC BY-NC-SA (non-commercial).

**8-machine matrix (2026-07-17) — tiled keys 2048 on EVERY Mac, same accuracy.** `tile_size=768, overlap=64` + our compile patch, full 2048² input, **peak ~2.2GB and alpha MAE 0.00849 identical on all**: M3 Ultra 1949ms · M1 Ultra 3275ms · M1 Max(24c/32GB) 4062ms · M4 mini(10c/16GB) 7284ms · **OG M1 mini (2020, 8c/8GB) 14269ms**. Full-frame 2048 needs 26GB → unreachable on most Macs ever sold; **tiled isn't the fallback, it's the answer**. Also: hdim-56 fast-path fallback confirmed on all 8 (padding win 1.3×–4.9×, tracks fused-kernel quality **not** core count — the M1 Max refuted our width theory, the M4 mini then fit the corrected one as a prediction); M5's MLX gains live almost entirely in the fused kernel; fanless M3 Air does **not** throttle (1.01× at 5 min).

**Fleet ablation follow-up (2026-07-16, full report in the fork: `monster/corridorkey-mrp-mlx` → `docs/2026-07-16-m-series-fleet-ablation-results.md`):** ran upstream's own 6-toggle benchmark matrix on M3+M1+M4. New best config on BOTH Ultras = **tiled 768/64 + compile** (M3 1949ms, M1 3275ms per 2048² raster — vs 2788/5373 full-frame) at 2.4GB. Big discovery: **`sdpa` is a 4.5× regression on M1-class GPUs at 2048 (18.6s vs 4.1s) and free on M3; bf16 is neutral on both** — gate sdpa by GPU generation, the "M1 bf16 danger" assumption is wrong for this workload. `stage_gc` harmful on Ultras (0.53–0.66×). M4 24GB swaps at full-frame 2048 (~20–24s all configs) → tiled mandatory there; healthy ≤1024.

### Phase D — hunyuan Studio tuning (2026-07-16, M3 Ultra) → new operator defaults
Raised the config from the laptop-tuned defaults to `octree_resolution 384` + `remesh_faces 120000` + `texture_size 4096`. **4096² bake works on the Studio GPU — no Metal command-buffer watchdog** (the existing `extract_textiles` tiling handles it; `uv_feature_map` never needed patching). Result: **380s** (shape 160 + tex 221), peak 20.2GB, 21.5MB GLB, 78k verts / **120k faces**, 4096² baseColor+MR. Quality jump is real — the melted face gains defined eyes + structure, tail geometry sharpens, textures crisper; closes most of the gap to trellis (trellis still edges the face). Cost: ~46% slower + ~3× file size vs the 40k/2048 default. **These are now the `hunyuan3d_mlx` operator defaults** (all still param-overridable; drop to `remesh_faces 40000`/`texture_size 2048` for fast drafts). **4096 confirmed watchdog-free on the M1 Ultra too** (751s total — M1 runs it at ~2× M3 time, so speed-critical jobs prefer M3, which is first in the pool). Defaults are fleet-safe on both Ultras.

## Mesh-gen — earlier install notes (superseded by the table above)
| Operator | Install | Runtime status |
|---|---|---|
| `sf3d` | venv + Metal texture_baker/uv_unwrapper kernels compiled OK; torch 2.13 MPS available | Runs end-to-end; **weights gated** — `stabilityai/stable-fast-3d` returns `GatedRepoError` until the owner accepts the license + sets an HF token. Then expect seconds-to-a-minute on MPS. |
| `trellis_mac` | setup.sh built .venv (py3.11) + mtl* Metal kernels; torch 2.13 MPS available | Runs end-to-end; **weights gated** — needs HF access to `facebook/dinov3-vitl16-pretrain-lvd1689m` + `briaai/RMBG-2.0`. Expect ~3–5 min/gen once authed (M4 Pro reference; M3 Ultra should match or beat). |
| `fal_*` (trellis / trellis2 / hunyuan3d / rodin) | none (API) | Gated on `FAL_KEY`. Verified param surfaces; ~1s–1min server-side per fal docs. |

## To unblock the gated local operators
1. Accept the model licenses on HuggingFace (one-time, usually instant):
   - https://huggingface.co/stabilityai/stable-fast-3d
   - https://huggingface.co/facebook/dinov3-vitl16-pretrain-lvd1689m
   - https://huggingface.co/briaai/RMBG-2.0
2. Either `huggingface-cli login` on the machine, or paste an HF token into
   Settings → "HuggingFace token" (injected as `HF_TOKEN` for the operators).

## Fleet hardware baseline (perfcheck, torch 2.13.0/MPS, 2026-07-17)

What each node's GPU is *capable* of, independent of any model — the denominator for every
number above, and the reference for "is this machine still healthy". Measured nightly by
`scripts/perfcheck/`; see **[PERFCHECK.md](PERFCHECK.md)** for the full story.

| node | chip | RAM | matmul fp16 | bandwidth | sdpa 56/64 |
|---|---|---|---|---|---|
| m3ultra | M3 Ultra | 256GB | 25202 | 627 GB/s | 4.35x |
| m1 | M1 Ultra | 128GB | ~17100 | 639 GB/s | 3.89x |
| studio | M1 Max | 32GB | 6996 | 348 GB/s | 3.09x |
| m4 | M4 Pro | 24GB | 5678 | 241 GB/s | 1.96x |
| m4mini | M4 | 16GB | 3774 | 105 GB/s | 5.10x |
| mini | M1 (2020) | 8GB | 2310 | 61 GB/s | 5.26x |

GFLOP/s. The M3 Ultra is **~2.2x** the M1 Ultra on fp16 matmul but they share the same ~625
GB/s memory bandwidth — so bandwidth-bound stages (VAE decode, big texture bakes) run at
similar speed on both, while compute-bound ones (denoise) scale with the newer silicon. Both
Ultras reach only ~78% of their ~800 GB/s spec on a single kernel; the smaller Macs hit ~88%.

`sdpa 56/64` is the fused-attention cliff at CorridorKey's awkward `head_dim=56` — it's why
padding to 64 wins (`CORRIDORKEY.md`), reproduced independently on all 6 nodes, and it is
watched nightly so we learn if a future torch ever closes it.

## Method
Timings are wall-clock from the job runner (`started_at`→`finished_at`), single
job at a time (gpu lane = 1). Re-run `tests/smoke.sh` for the framework
regression suite (12 checks, ~30s).

**Benchmarking on this fleet is contended.** These are working machines; a neighbouring job
can make a node look 5–11x slower and it is invisible to load average (the M3 Ultra reads load
2.45 with its GPU pinned). Anything measured here should use min-of-runs and check
`ioreg … IOAccelerator` "Device Utilization %" *before* touching the GPU. See PERFCHECK.md.

## 2026-07-19 — fleet MLX capability matrix (mlx 0.32.0 everywhere)

Probe: 4096² matmul TFLOPS (fp16/bf16), 256MB elementwise GB/s, sdpa head_dim 56 vs 64 (B1·H8·L2048).

| host | chip | RAM | fp16 TF | bf16 TF | ~GB/s | sdpa d56/d64 |
|---|---|---|---|---|---|---|
| m3ultra | M3 Ultra | 256 | 23.48 | 23.54 | 400 | 1.51× |
| ultra | M1 Ultra | 128 | 16.31 | 13.16 | 297 | 0.94× |
| JING5 | M5 | 16 | 8.61 | 8.99 | 43* | **4.72×** |
| m4pro | M4 Pro | 24 | 5.24 | 3.86 | 225 | 1.29× |
| m4mini | M4 | 16 | 3.65 | 3.65 | 98 | 1.55× |
| AIRPRO | M3 | 16 | 1.55 | 1.64 | 87 | 0.95× |

*JING5 bandwidth measured under interactive load — retest idle before trusting.

Verdicts:
- **bf16-on-M1 question answered**: works, ~20% slower than fp16 (16.3→13.2). Ship fp16 on M1-class, bf16 free on M3 Ultra/M4/M5.
- **head_dim pad-to-64 matters MOST on newest chips** (M5 4.7×!, M4-class ~1.3-1.6×, M3 Ultra 1.5×) and ~zero on M1-class (no fused fast path there anyway — matches the corridorkey ablation).
- Fleet total ≈ 59 TFLOPS fp16; the two Ultras are 2/3 of it. TRELLIS.2-MLX lane assignment: m3ultra primary, ultra second (fp16), M4 Pro light ops.
- Probe script: reusable one-liner via `uv run --with mlx` — mlxprobe.py (this bench).

Addendum (same day) — m1max studio joins the matrix (login `studio@100.92.78.24`;
the fleet doc's `m1max@` user and "disk-critical" note are both stale — 170GB free):

| M1MAX | M1 Max | 32 | 7.09 | 6.04 | 336 | 1.22× |

Notable: 2nd-best bandwidth in the fleet (336 GB/s > M4 Pro's 225) despite mid-pack
compute — good fit for memory-bound MLX work (VAE decode, big-tensor elementwise).
Fleet total ≈ 66 TFLOPS fp16 across 7 machines.

## 2026-07-19 — trellis_mac rung 1: sparse-conv backend verified (m3ultra, anatomy image, seed 42, 1024-cascade)

| SPARSE_CONV_BACKEND | total gen+bake | wall | peak RSS |
|---|---|---|---|
| flex_gemm (Metal) | **124.8 s** | 248.8 s | 20.8 GB |
| none (pure-torch) | 214.4 s | 337.6 s | 20.8 GB |

Verdicts:
- The 124.8s baseline WAS the Metal fast path (log line "[SPARSE] Conv backend: flex_gemm") — no hidden 10x jackpot; metallib loads fine on macOS 26.5.
- Metal sparse-conv is worth **1.72x end-to-end** (saves 89.6s/gen). Keep flex_gemm the default; the conv_none fallback is a real 90s regression if the metallib ever breaks silently — worth a GODBOARD/log assertion.
- Remaining fat in the fast path: ~46s diffusion sampling (padded SDPA attention — next kernel frontier), ~50s decode/extract (incl. aten::segment_reduce falling back to CPU — patchable), 19s Metal bake.
- Decode actually produces 2.79M faces before the forced ~200k simplify — the quality-recovery headroom (rung 2) is bigger than the recon estimated.

## 2026-07-19 — Hunyuan paint-UNet mx.compile (Target 1, branch mlx-tune)

Micro gate (real arch, random fp16 weights, 6-view 64² latents, m3ultra):
eager 270.5 ms/fwd → compiled 260.5 ms/fwd = **1.04×**; parity max|diff|
3.0e-05 (PASS); first-trace 1.2 s; peak 4.9 GB. Env-gated
`HY3D_MLX_COMPILE=1`, default off (numerics untouched).

Verdict: keep (free, safe) but the paint UNet is GEMM-bound — fusion alone
is small. The bigger Hunyuan levers are Target 3 (MoE `.item()` sync-tax,
−4800 syncs/gen, unlocks whole-DiT compile) and Target 4 (manual→fused
SDPA in paint DINO). E2E flag-on-vs-off measurement pending after those.