Metronome count — audio events on an exact grid

What this stresses: Timing precision. Five evenly spaced beats give a measurable grid: any drift between the spoken count, the clap and the picture shows up as milliseconds.
960×544 · 124f (5.17 s) · 24 fps · seed 424242 · identical prompt in every cell. Same seed does not mean same take — steps, sampler and LoRA all change the trajectory, so judge character rather than shot-for-shot identity.

prompt
integrated_multimodal_description:
[Shot 1] A photorealistic medium close-up of a bald Black man in his forties in a black t-shirt, standing in a bright empty dance studio with a mirrored wall and a pale sprung floor, facing the camera squarely, arms raised ready to clap. He claps his hands together sharply exactly five times, at 0.5, 1.5, 2.5, 3.5 and 4.5 seconds, evenly spaced one second apart, and on each clap he calls out the number in a strong clear baritone. (S1) says: <d>[English] One. Two. Three. Four. Five.</d> with each number landing exactly on its clap. Between claps his hands draw apart to shoulder width and come back together; the motion is metronomic and identical every time. The camera is locked off and never moves; his feet stay planted.

overall_soundscape:
A large reflective studio with a short bright slap-back. Five sharp hand claps at 0.5, 1.5, 2.5, 3.5 and 4.5 seconds, each identical in level and attack, each followed by a short room reflection. (S1)'s counted numbers are close, clear and land exactly with each clap, not before or after. Almost no other sound between the claps.

non_diegetic_music:
N/A
baseline · 20 steps
No LoRA. The reference render.
165s · floor -39.8 · med -28.3 · peak 2.8 dB · jitter 101.5 · detail 0.054
ema-ckpt500 · 8 steps
larryvrh/drbaph Turbo LoRA, EMA weights. Round-3 winner.
76s · floor -52.8 · med -31.5 · peak 2.9 dB · jitter 100.8 · detail 0.051
ckpt500 (non-EMA) · 8 steps
Same checkpoint without EMA averaging. Dragged a noise bed in round 3 — re-tested here under the fixed audio path.
76s · floor -45.1 · med -24.5 · peak 3.3 dB · jitter 101.2 · detail 0.068
lightx2v · 4 steps · str 1.0
ModelTC reference settings.
45s · floor -51.2 · med -20.1 · peak 3.6 dB · jitter 192.5 · detail 0.135
lightx2v · 4 steps · str 1.0 · shift 12/6
Reference settings with our house audio shift.
45s · floor -58.4 · med -27.4 · peak 2.9 dB · jitter 195.3 · detail 0.125
lightx2v · 4 steps · str 1.0 · er_sde
Isolates the sampler at full strength.
44s · floor -59.8 · med -32.0 · peak 2.2 dB · jitter 167.6 · detail 0.039
lightx2v · 4 steps · str 0.75
Isolates strength on the reference sampler.
45s · floor -53.4 · med -25.4 · peak 3.8 dB · jitter 140.5 · detail 0.117
lightx2v · 4 steps · str 0.75 · shift 12/6
Lower strength with the house shift.
45s · floor -57.8 · med -27.9 · peak 3.2 dB · jitter 134.4 · detail 0.112
lightx2v · 4 steps · str 0.75 · er_sde
Kijai's recommended config. Pilot winner on jitter and luma.
44s · floor -58.8 · med -33.4 · peak 2.7 dB · jitter 126.3 · detail 0.043
lightx2v · 8 steps · str 1.0
Double the trained step count, reference settings.
76s · floor -54.5 · med -27.6 · peak 4.0 dB · jitter 195.8 · detail 0.057
lightx2v · 8 steps · str 1.0 · shift 12/6
Double steps with the house shift.
76s · floor -55.0 · med -22.9 · peak 2.5 dB · jitter 186.6 · detail 0.056
lightx2v · 8 steps · str 0.75 · er_sde
Kijai's config given twice the steps.
76s · floor -59.9 · med -24.4 · peak 3.0 dB · jitter 138.2 · detail 0.053
baseline · 20 steps + Spectrum
No LoRA, plus SpectrumApplyMiniMaxH3 (v0.1.9 defaults: degree 1, warmup 1, one-point bootstrap) skipping DiT evaluations.
107s · floor -51.1 · med -31.5 · peak 3.4 dB · jitter 94.8 · detail 0.048
ema-ckpt500 · 8 steps + Spectrum
Turbo LoRA and Spectrum stacked. Only possible since Spectrum v0.1.9 — the old warmup_steps=5 default left no forecastable window at 8 steps.
61s · floor -49.3 · med -34.7 · peak 2.7 dB · jitter 104.9 · detail 0.056
ema-ckpt500 · 8 steps + SageAttention
Turbo LoRA with SageAttention 2.2.0 (KJNodes patch, backend auto). Unlike Spectrum, Sage preserves the take — 0.955 frame correlation against plain emck8.
50s · floor -51.8 · med -34.2 · peak 3.3 dB · jitter 91.6 · detail 0.048
ema-ckpt500 · 8 steps + Sage + Spectrum
Everything stacked: the fastest usable config here, at the cost of a different take (the drift is Spectrum's, not Sage's).
41s · floor -47.3 · med -33.7 · peak 2.8 dB · jitter 105.9 · detail 0.057
baseline · 20 steps + SageAttention
Isolates the attention backend with no LoRA in play. Near-identical output to the reference render (0.974 correlation).
98s · floor -41.7 · med -31.0 · peak 3.7 dB · jitter 93.8 · detail 0.061
baseline · 20 steps + Sol-Attn
NVIDIA Sol-Attn sparse attention via the H3 zero-copy patch (tau 1.0, sink_conditioning exact_kv). Changes the take substantially — 0.723 correlation.
99s · floor -41.8 · med -26.8 · peak 2.3 dB · jitter 120.7 · detail 0.054
ema-ckpt850 · 8 steps
larryvrh's further-trained EMA checkpoint — the file he recommends — against the ckpt500 used everywhere else here.
76s · floor -46.5 · med -30.7 · peak 3.2 dB · jitter 99.4 · detail 0.068
v4-step600 EMA · 8 steps
larryvrh's new v4 training recipe (step 600, EMA) — a different training line, not a further ckpt of the 500/850 series. Single-variable swap against emck8, the round-4 winner: same steps, sampler, scheduler, shift and strength.
77s · floor -42.9 · med -28.5 · peak 3.0 dB · jitter 97.6 · detail 0.059
v4-step600 non-EMA · 8 steps
The same v4 checkpoint without EMA averaging. Re-runs the EMA axis on the new recipe — on the v1 line the gap was 6.3 dB of noise floor (emck8 −5.6 vs ck8 +0.7).
76s · floor -42.2 · med -28.3 · peak 3.6 dB · jitter 92.4 · detail 0.064
v4-step600 EMA · 8 steps · euler + beta
drbaph's recommended config for the pruned conversion we load: 8 steps, euler, beta scheduler. Varies sampler *and* scheduler against the rest of the grid, so it is only readable against v4e8, not against the res_multistep columns.
76s · floor -50.3 · med -32.3 · peak 3.3 dB · jitter 85.4 · detail 0.065
v4-step600 EMA · 4 steps
Probes the one regression larryvrh documents for v4: motion smear / trailing ghosting at 4 steps under fast motion, which 6–8 steps is said to remove. rapid-cuts, hands-dexterity and polyphony-foreground are where it should show.
45s · floor -40.9 · med -33.4 · peak 2.8 dB · jitter 144.9 · detail 0.249
v4-step600 EMA · 8 steps + SageAttention
Re-bases the production accelerator onto v4. emck8-sage (45 s, 3.66×, 0.955 correlation) is the current production config; if v4e8 dethrones emck8 this is what replaces it. Sage is an attention-backend swap and so should be weight-independent — this checks that.
50s · floor -46.6 · med -30.0 · peak 3.3 dB · jitter 97.4 · detail 0.092
ema-ckpt500 · 8 steps · euler + beta
Control for v4e8-eb, which varies LoRA, sampler and scheduler at once. Putting drbaph's euler+beta on the weights we already have 10 scenes of isolates the sampler/scheduler axis, so an eb win can be attributed to the config rather than to v4.
76s · floor -48.9 · med -33.4 · peak 2.8 dB · jitter 96.6 · detail 0.055

Measured

configtimefloormedianpeakjitterdetailraw detailluma
baseline · 20 steps165s-39.8-28.32.8101.50.054250153.6
ema-ckpt500 · 8 steps76s-52.8-31.52.9100.80.051208154.6
ckpt500 (non-EMA) · 8 steps76s-45.1-24.53.3101.20.068283150.1
lightx2v · 4 steps · str 1.045s-51.2-20.13.6192.50.135719136.6
lightx2v · 4 steps · str 1.0 · shift 12/645s-58.4-27.42.9195.30.125669137.6
lightx2v · 4 steps · str 1.0 · er_sde44s-59.8-32.02.2167.60.039200145.0
lightx2v · 4 steps · str 0.7545s-53.4-25.43.8140.50.117622135.6
lightx2v · 4 steps · str 0.75 · shift 12/645s-57.8-27.93.2134.40.112605136.1
lightx2v · 4 steps · str 0.75 · er_sde44s-58.8-33.42.7126.30.043212145.9
lightx2v · 8 steps · str 1.076s-54.5-27.64.0195.80.057283141.1
lightx2v · 8 steps · str 1.0 · shift 12/676s-55.0-22.92.5186.60.056279140.9
lightx2v · 8 steps · str 0.75 · er_sde76s-59.9-24.43.0138.20.053261150.0
baseline · 20 steps + Spectrum107s-51.1-31.53.494.80.048196160.9
ema-ckpt500 · 8 steps + Spectrum61s-49.3-34.72.7104.90.056206160.9
ema-ckpt500 · 8 steps + SageAttention50s-51.8-34.23.391.60.048186155.9
ema-ckpt500 · 8 steps + Sage + Spectrum41s-47.3-33.72.8105.90.057208159.6
baseline · 20 steps + SageAttention98s-41.7-31.03.793.80.061279160.3
baseline · 20 steps + Sol-Attn99s-41.8-26.82.3120.70.054254164.0
ema-ckpt850 · 8 steps76s-46.5-30.73.299.40.068287154.2
v4-step600 EMA · 8 steps77s-42.9-28.53.097.60.059244150.8
v4-step600 non-EMA · 8 steps76s-42.2-28.33.692.40.064270150.8
v4-step600 EMA · 8 steps · euler + beta76s-50.3-32.33.385.40.065268150.6
v4-step600 EMA · 4 steps45s-40.9-33.42.8144.90.249792131.8
v4-step600 EMA · 8 steps + SageAttention50s-46.6-30.03.397.40.092379145.5
ema-ckpt500 · 8 steps · euler + beta76s-48.9-33.42.896.60.055242155.5
Rendered on ComfyUI 0.30.0 at commit 344b4398 (2026-08-07, "Support asym w4a8_int (#15308)") · torch 2.12.1+cu130 · NVIDIA GeForce RTX 5080, 15.4 GB
Custom nodes: ComfyUI-Spectrum-MiniMax-H3 1a06e20 (v0.1.9) — used by the two Spectrum columns only · comfyui-kjnodes 1.4.9 · ComfyUI-sol-attn 251c27f
All 14 columns rendered on this single ComfyUI commit. It postdates bdcb886a (MiniMax audio-sampler rework, merged 2026-08-06), which is why round-3 clips are archived separately rather than mixed in.