Repeated canaries
Ten-run medians
Four-second compatibility workloads. They describe repeatability inside each lane, not full-length relative quality.
Sanitized evidence index
Benchmark data
A human-readable overview and a filterable record of measured H3 executions. No source benchmark media, raw prompts, personal references, provider identifiers, or absolute local paths are included.
Fixed canonical shot
Full-length highlights
These rows use the 768 × 1344, 15.084-second, 362-frame native-audio fixture and canonical seed unless the recipe explicitly contains a multi-seed preview stage.
| Workflow | Total | Sampling | Recipe | Observation |
|---|---|---|---|---|
| Standard | 611.380 s | 558.131 s | 20 steps · SageAttention2 + FirstBlockCache Safe | Reference baseline for the canonical 15-second shot. |
| PDD + Sage | 375.042 s median | 323.581 s median | 8 steps · PDD shifts 12/3 · SageAttention2 | Two uncached runs; 1.63× the Standard throughput. |
| Seed Hunter direct | 285.765 s | 232.650 s | 12-step low-resolution pass · 3D latent upscale · 4-step refine | 2.14× Standard throughput; identical to the same selected-seed interactive output. |
| 10Eros beta4 author recipe | 373.961 s | 319.065 s | 8 steps · Comfy-Kitchen · Euler/simple · shifts 12/7 | Exact dialogue; a camera-perspective inconsistency was observed. |
| 10Eros + Seed Hunter three-preview | 328.394 s | 234.172 s | Three 4-step previews · 3D latent upscale · 4-step refine | 1.86× Standard throughput; exact dialogue with a camera-perspective inconsistency. |
| 10Eros + Seed Hunter direct | 243.521 s | 184.942 s | One 4-step preview · 3D latent upscale · 4-step refine | 2.51× Standard throughput; exact dialogue with the same perspective caveat. |
| 10Eros + Seed Hunter camera-POV selected path | 265.346 s warm | 208.832 s | Clarified prompt · one 4-step preview · 3D latent upscale · 4-step refine | 2.30× Standard throughput; exact dialogue and no visible phone in the dense review. |
Additional measurements
Quality and memory
Historical matched sets
Documented results without public media
These values remain useful as data, but they belong to different workloads and should only be compared within matching rows.
SageAttention2
1344 × 768 · 15.084 s · 20 steps
851.534 s808.815 s samplingSage + FBC Safe
1344 × 768 · 15.084 s · 20 steps
647.308 s607.727 s samplingLightX2V 768p
1344 × 768 · 15.084 s · 8 steps
363.357 s322.094 s samplingSLA dense wrapper
576 × 1024 · 15.084 s · Ref2VA
351.201 s324.715 s samplingSLA 85% request
576 × 1024 · 15.084 s · Ref2VA
294.211 s270.284 s samplingComplete normalized index
Explore every timing record
Cached replays are hidden by default. Search by workload, backend, package, or source evidence locator.
| Record | Lane | Workload | Total | Sampling | Backend / package | Evidence locator |
|---|