Which LoRA Acceleration Is Right for You? The Blind-Test Leaderboard
Where this leaderboard comes from
H3 Acceleration Arena pits ~26 acceleration variants (LoRAs, full checkpoints, an API option and the 28-step baseline) against each other on pre-generated clips from a ~200-prompt corpus, via blind A/B voting: voters never see the variant names, just two clips, and pick the better one. Scores are Elo with confidence intervals; i2v and t2v are separate slices. (Sources: the Arena Space and arena/README.md.)
The tables below were fetched from the Arena's public API (/api/leaderboard; 26,593 i2v votes and
27,618 t2v votes at fetch time) on 2026-09-20 and re-ordered by Elo — no re-computation.
Timing reference: RTX PRO 6000 Blackwell 96GB, bf16, no CPU offload (source: Arena timing_reference).
i2v slice · top 10
| # | Variant | NFE | Elo ±CI | Votes | Win rate | Timed speedup | Source repo |
|---|---|---|---|---|---|---|---|
| 1 | LightX2V Turbo 8-step v1.0 · native | 8 | 1086 ±63 | 4105 | 42% | 3.07× | lightx2v/Minimax-h3-Turbo |
| 2 | larryvrh Turbo v4 · step600 EMA | 6 | 1084 ±62 | 3871 | 45% | 4.12× | larryvrh/MiniMax-H3-Turbo-Lora |
| 3 | Plaguekind Parasyte Turbo | 6 | 1041 ±62 | 4142 | 46% | 4.09× | Plaguekind/H3-Lora |
| 4 | MiniMax H3-Max (hosted via fal) | — | 1035 ±70 | 4032 | 44% | — | fal.ai |
| 5 | silveroxides 4-to-8-step DARE merge | 6 | 1026 ±66 | 4167 | 43% | 4.09× | silveroxides/MiniMax-H3_tests |
| 6 | VDN-H3 8-step (OpenVDN full ckpt) | 8 | 1010 ±59 | 938 | 39% | 3.17× | OpenVDN/vdn-minimax-h3 |
| 7 | LightX2V Turbo 8-step v1.0 768p | 8 | 1007 ±56 | 3760 | 38% | 3.07× | lightx2v/Minimax-h3-Turbo |
| 8 | LightX2V Turbo 4-step v0.1 | 4 | 1005 ±61 | 2994 | 32% | 6.18× | lightx2v/Minimax-h3-Turbo |
| 9 | FastVideo FastH3 Preview v0.2 | 4 | 999 ±56 | 1444 | 36% | — | FastVideo/…-Preview-v0.2 |
| 10 | silveroxides DARE-TIES merge | 6 | 999 ±56 | 4123 | 44% | 4.12× | silveroxides/MiniMax-H3_tests |
t2v slice · top 10
| # | Variant | NFE | Elo ±CI | Votes | Win rate | Timed speedup | Source repo |
|---|---|---|---|---|---|---|---|
| 1 | larryvrh Turbo v4 · step600 EMA | 6 | 1110 ±62 | 3308 | 49% | 4.12× | larryvrh/MiniMax-H3-Turbo-Lora |
| 2 | MiniMax-H3 · 28-step baseline | 27 | 1073 ±66 | 3543 | 47% | 1.0× | MiniMax official weights |
| 3 | silveroxides DARE-TIES merge | 6 | 1065 ±54 | 3173 | 47% | 4.12× | silveroxides/MiniMax-H3_tests |
| 4 | Tutu AudioVideo 20→8 · ckpt100 | 8 | 1042 ±67 | 2018 | 39% | 3.09× | tutututututu/Tutu-…-LoRA |
| 5 | LightX2V Turbo 8-step v1.0 · native | 8 | 1029 ±63 | 3211 | 44% | 3.07× | lightx2v/Minimax-h3-Turbo |
| 6 | VDN-H3 8-step (OpenVDN full ckpt) | 8 | 1029 ±71 | 1114 | 45% | 3.17× | OpenVDN/vdn-minimax-h3 |
| 7 | FastVideo FastH3 v1 VSA-DataFree · 6 steps (community config) | 6 | 1028 ±73 | 4141 | 40% | 10.34× | FastVideo/…-VSA-DataFree |
| 8 | Plaguekind Parasyte Turbo | 6 | 1025 ±79 | 3149 | 49% | 4.09× | Plaguekind/H3-Lora |
| 9 | alibaba-pai MiniMax-H3-Acc (PDD) 8-step | 8 | 1018 ±65 | 3097 | 45% | 3.09× | alibaba-pai/MiniMax-H3-Acc-LoRAs |
| 10 | MiniMax H3-Max (hosted via fal) | — | 1006 ±67 | 3218 | 41% | — | fal.ai |
Three findings worth noting
- larry v4 leads both: i2v #2 (1084) and t2v #1 (1110) — the only entry in the top three of both slices; the i2v crown goes to LightX2V 8-step (native, 1086), so the same family ranks very differently across slices (source: tables above).
- Fast ≠ good-looking: the 28-step baseline still ranks #2 on t2v, while the 4-step family (6.18× timed speedup) sits at i2v #8 (source: tables above).
- 6-NFE adapters remain the sweet spot: 3 of the i2v top 5 and 2 of the t2v top 3 are 6 NFE (~4.1× timed speedup) — nearly 4-step speed with near-baseline preference (source: tables above; Arena
conventions: NFE = actual forwards, 6 NFE = 7 steps).
Mapping to this site (selectable in the calculator)
| Arena entry | i2v / t2v rank | WhichH3 calculator entry |
|---|---|---|
| Plaguekind Parasyte Turbo | #3 / #8 | Parasyte (6 NFE) |
| larryvrh Turbo v4 step600 EMA | #2 / #1 | larry v4 8-step |
| alibaba-pai PDD 8-step | #14 / #9 | PDD 8-step (official distill) |
| silveroxides DARE-TIES | #10 / #3 | silveroxides (6 NFE) |
| LightX2V 8-step v1.0 (768p profile) | #7 / #16 | LightX2V Turbo 8-step |
| LightX2V 4-step v1.1 768p | #17 / #21 | LightX2V Turbo 4-step v1.1 |
| FastVideo FastH3 (VSA) | #12 / #7 | FastH3 4-step (VSA) |
| TaoMate-H3 3-step | #19 / #24 | TaoMate 3-step |
Pick any acceleration in step 2 of the calculator and its arena rank is shown right below (with the source). For stacking rules see the acceleration comparison; downloads live in the model library.
How to choose, by goal
- Fast but not ugly (mostly i2v): the 6-NFE family — Parasyte / silveroxides DARE-TIES / larry v4 (~4.1× timed speedup, i2v top 5).
- Safest, least fiddling: LightX2V 8-step (same as official turbo_mode) or PDD 8-step (official distill, needs its dedicated node — source: alibaba-pai).
- t2v with quality first: silveroxides DARE-TIES (t2v #1), or expect less from aggressive step cuts.
- Maximum speed: 4-step options (e.g. LightX2V 4-step) ≈6.2× timed speedup, but preference ranks clearly lower — spot-check before a batch.
- Small VRAM / long clips: fix resolution and VRAM first (low-VRAM setups), then stack acceleration; attention backends (Sage/Kitchen/SLA/VSA) stack orthogonally.
Limitations (read before quoting)
- It's community voting: every vote counts equally, one account can vote many times, and the sample is not everyone (source: Arena README, “Going public”).
- The list mixes full checkpoints (OpenVDN, FastVideo) and an API option (fal H3-Max) with LoRAs — not the same usage model (source: Arena
kindfield). - Versions of the same family are separate entries (v1.0 / v1.1 / v0.2…); don't merge them (source: Arena
variants.json). - Timed speedups only mean something on the same reference hardware (RTX PRO 6000 96GB, bf16, no offload); your absolute numbers will differ (source: Arena
timing_reference). - WhichH3's calculator reports a resolution × duration matrix estimate (calibrated on 30+ community anchors) — a different measurement than the arena's timed runs (source: this site's data notes).
- The leaderboard evolves; our snapshot is 2026-09-28 (votes still growing) — after that, the Arena is the source of truth.
Sources
- Leaderboard & votes: H3 Acceleration Arena (HF Space, multimodalart), API fetched 2026-09-28
- Method & caveats: arena/README.md · caveats.json · variants.json
- Per-variant repos: see the “Source repo” column in the tables above
- Our archived dataset copy: whichh3-data/data/acceleration_arena.json