H3 inference / schedule comparison
Six MiniMax H3 execution paths receive the same first frame, prompt, seed, checkpoint, resolution, and duration. The Turbo profile and attention schedule change; measured wall time is reported for every sequence. A complete 55-render integer handoff sweep follows at the bottom.
01 / 4-step family
The dense and 2→2 paths use four evaluations. The 2→4 path adds dense subdivisions across the trajectory's final half for six evaluations total.
4-step family
The dense four-step sequence completed in 99.8 seconds. It is the timing baseline for the other four-step schedules.
4-step family
Making the first half sparse while keeping four total evaluations reduced wall time to 75.2 seconds—a saving of 24.5 seconds, or 24.6%.
4-step family
Spending four dense evaluations over the final half completed in 94.7 seconds. That saved 5.0 seconds, or 5.0%, against the dense four-step baseline while using six evaluations.
02 / 8-step family
All three paths use the same eight-step Turbo adapter, shifts, sampler, scheduler, and complete sigma schedule. The attention path is the changing variable.
8-step family
The dense eight-step sequence completed in 184.3 seconds. It is the timing baseline for the other eight-step schedules.
8-step family
Keeping all eight evaluations sparse completed in 126.8 seconds. That saved 57.6 seconds, or 31.2%, against the dense eight-step baseline.
8-step family
Changing from sparse to dense attention at the midpoint completed in 149.1 seconds. That saved 35.2 seconds, or 19.1%, against the dense eight-step baseline.
integrated_multimodal_description: <Picture 1> is the exact first frame. Preserve the exact on-screen subject visible in this image, including its identity, species-defining anatomy, head and facial structure, surface covering, proportions, clothing, background, framing, and lighting. Do not humanize the subject or change its species. [Shot 1] One continuous, natural medium close-up with a locked camera. The subject looks directly into the camera, blinks naturally, makes small restrained head and hand movements, and speaks clearly with calm confidence. Keep the subject's face stable and mouth movement precisely synchronized. (S1), the on-screen subject, says: <d>[English] Atlas began with a simple idea: creative tools should work together instead of getting in the way. We connected the first pieces, tested them on real projects, and kept refining them until Atlas became a practical creative partner.</d> The subject finishes the sentence, closes its mouth naturally, and holds eye contact for the final moment. No cuts, no camera movement, no captions, no subtitles, no logos, and no visible text. overall_soundscape: Clean close-miked speech from (S1), subtle natural room tone, quiet breathing, and faint clothing movement. No other voices or prominent environmental sounds. non_diegetic_music: N/A
55-run timing overview
Each line moves from an all-dense schedule at 0% sparse to an all-sparse schedule at 100%. Lower is faster. The panels use labeled but different vertical ranges so the shape of each Turbo family remains readable; the exact ledger below carries every measured value.
21 measured sequences
34 measured sequences
| Profile | Total NFE | Sparse → dense path | Wall time | Vs. dense | Reduction |
|---|---|---|---|---|---|
| Turbo 4 | 4 | All dense · 4 steps | 85.9 s | baseline | — |
| 1 sparse → 3 dense | 86.1 s | 0.3 s longer | -0.3% | ||
| 2 sparse → 2 dense | 79.7 s | 6.2 s saved | 7.2% | ||
| 3 sparse → 1 dense | 73.1 s | 12.8 s saved | 14.9% | ||
| All sparse · 4 steps | 67.9 s | 18.0 s saved | 20.9% | ||
| Turbo 4 | 6 | All dense · 6 steps | 139.3 s | baseline | — |
| 1 sparse → 5 dense | 131.3 s | 8.0 s saved | 5.7% | ||
| 2 sparse → 4 dense | 124.2 s | 15.1 s saved | 10.8% | ||
| 3 sparse → 3 dense | 110.9 s | 28.4 s saved | 20.4% | ||
| 4 sparse → 2 dense | 107.4 s | 31.8 s saved | 22.9% | ||
| 5 sparse → 1 dense | 101.8 s | 37.4 s saved | 26.9% | ||
| All sparse · 6 steps | 94.8 s | 44.5 s saved | 32.0% | ||
| Turbo 4 | 8 | All dense · 8 steps | 181.8 s | baseline | — |
| 1 sparse → 7 dense | 173.7 s | 8.1 s saved | 4.5% | ||
| 2 sparse → 6 dense | 166.9 s | 14.9 s saved | 8.2% | ||
| 3 sparse → 5 dense | 158.0 s | 23.8 s saved | 13.1% | ||
| 4 sparse → 4 dense | 146.1 s | 35.7 s saved | 19.6% | ||
| 5 sparse → 3 dense | 142.1 s | 39.7 s saved | 21.9% | ||
| 6 sparse → 2 dense | 131.5 s | 50.3 s saved | 27.7% | ||
| 7 sparse → 1 dense | 126.4 s | 55.4 s saved | 30.4% | ||
| All sparse · 8 steps | 123.5 s | 58.3 s saved | 32.1% | ||
| Turbo 8 | 4 | All dense · 4 steps | 95.5 s | baseline | — |
| 1 sparse → 3 dense | 88.8 s | 6.7 s saved | 7.0% | ||
| 2 sparse → 2 dense | 81.8 s | 13.8 s saved | 14.4% | ||
| 3 sparse → 1 dense | 72.6 s | 23.0 s saved | 24.0% | ||
| All sparse · 4 steps | 68.0 s | 27.6 s saved | 28.9% | ||
| Turbo 8 | 6 | All dense · 6 steps | 139.1 s | baseline | — |
| 1 sparse → 5 dense | 130.7 s | 8.4 s saved | 6.0% | ||
| 2 sparse → 4 dense | 126.0 s | 13.1 s saved | 9.4% | ||
| 3 sparse → 3 dense | 111.7 s | 27.5 s saved | 19.7% | ||
| 4 sparse → 2 dense | 104.5 s | 34.7 s saved | 24.9% | ||
| 5 sparse → 1 dense | 99.2 s | 39.9 s saved | 28.7% | ||
| All sparse · 6 steps | 94.7 s | 44.4 s saved | 31.9% | ||
| Turbo 8 | 8 | All dense · 8 steps | 182.7 s | baseline | — |
| 1 sparse → 7 dense | 175.0 s | 7.8 s saved | 4.2% | ||
| 2 sparse → 6 dense | 165.8 s | 16.9 s saved | 9.3% | ||
| 3 sparse → 5 dense | 158.4 s | 24.3 s saved | 13.3% | ||
| 4 sparse → 4 dense | 146.2 s | 36.6 s saved | 20.0% | ||
| 5 sparse → 3 dense | 137.7 s | 45.0 s saved | 24.6% | ||
| 6 sparse → 2 dense | 130.8 s | 51.9 s saved | 28.4% | ||
| 7 sparse → 1 dense | 126.4 s | 56.3 s saved | 30.8% | ||
| All sparse · 8 steps | 125.7 s | 57.1 s saved | 31.2% | ||
| Turbo 8 | 12 | All dense · 12 steps | 263.6 s | baseline | — |
| 1 sparse → 11 dense | 262.1 s | 1.5 s saved | 0.6% | ||
| 2 sparse → 10 dense | 251.4 s | 12.2 s saved | 4.6% | ||
| 3 sparse → 9 dense | 241.6 s | 22.0 s saved | 8.4% | ||
| 4 sparse → 8 dense | 229.1 s | 34.5 s saved | 13.1% | ||
| 5 sparse → 7 dense | 220.2 s | 43.4 s saved | 16.5% | ||
| 6 sparse → 6 dense | 216.1 s | 47.5 s saved | 18.0% | ||
| 7 sparse → 5 dense | 209.1 s | 54.5 s saved | 20.7% | ||
| 8 sparse → 4 dense | 201.6 s | 62.0 s saved | 23.5% | ||
| 9 sparse → 3 dense | 193.8 s | 69.8 s saved | 26.5% | ||
| 10 sparse → 2 dense | 187.0 s | 76.5 s saved | 29.0% | ||
| 11 sparse → 1 dense | 186.1 s | 77.4 s saved | 29.4% | ||
| All sparse · 12 steps | 183.3 s | 80.3 s saved | 30.5% |
Time-equivalent depth
These two views isolate the practical exchange: a longer schedule can fit inside—or very near—the wall-time envelope of a shorter fully dense run when most early evaluations are sparse and the final evaluation is dense.
ComfyUI replication kit
H3-Optimizations 0.3.0 adds the exact integer-boundary sampler used here and no new Python dependencies. Import the workflow matching the Turbo adapter, choose the total schedule length in BasicScheduler, then set the sampler's sparse step count. The remaining evaluations are dense automatically.
Focused A/B comparison
Match every all-dense control against the schedule that keeps all but its final evaluation sparse. Start either player independently; starting one automatically pauses the other so their audio never overlaps.
A · Dense control
B · One dense finish
55-render addendum
Choose the four- or eight-step Turbo family, select a total NFE budget, then move through each sparse-to-dense boundary in sets of three. Every sequence keeps the same first frame, prompt, seed, base checkpoint, Euler sampler, Simple schedule, and continuous latent. The sparse stage uses a 30% video-KV budget; the dense stage uses Comfy Kitchen INT8.
21 sequences
4 total NFE
dense timing baseline
4 total NFE
0.3 s longer · 0.3%
4 total NFE
6.2 s saved · 7.2%
Showing 1–3 of 5 schedules at 4 total NFE
34 sequences
4 total NFE
dense timing baseline
4 total NFE
6.7 s saved · 7.0%
4 total NFE
13.8 s saved · 14.4%
Showing 1–3 of 5 schedules at 4 total NFE