Which LoRA Acceleration Is Right for You? The Blind-Test Leaderboard

Updated 2026-09-20 · Leaderboard source: H3 Acceleration Arena (by multimodalart, Hugging Face Space) · fetched 2026-09-20

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; 25,905 i2v votes and 26,883 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

Source: H3 Acceleration Arena public API, fetched 2026-09-20. “Timed speedup” is the Arena's own timing vs its 28-step baseline (not a WhichH3 estimate).

t2v slice · top 10

Same source. Full 23 (i2v) / 27 (t2v) entries with confidence intervals: the Arena Space. Our archived copy: whichh3-data: acceleration_arena.json.

Three findings worth noting

• The two slices disagree: Parasyte tops i2v while silveroxides DARE-TIES tops t2v, and the same variants trade places across slices (source: the tables above / Arena slices).
• Fast ≠ good-looking: the 28-step baseline ranks #2 on t2v, while the 4-step family (6.18× timed speedup) sits at i2v #8/#16/#18 (source: tables above).
• 6-NFE “8-step” adapters are the current sweet spot: 4 of the i2v top 5 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)

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 kind field).
• 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 2026-09-20 — after that, the Arena is the source of truth.

Sources

• Leaderboard & votes: H3 Acceleration Arena (HF Space, multimodalart), API fetched 2026-09-20
• 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