Near-silence into a hard transient — dynamic range

What this stresses: The quietest register plus a sharp peak in one clip. Yesterday every LoRA ran ~9 dB hot on whisper; this pairs that floor with a transient that must not clip.
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 wide shot of a library reading room at dusk, long oak tables, green banker's lamps, tall arched windows with cold blue light, a single young man with curly brown hair and a maroon scarf seated alone at the near table reading, absolutely still. From 0.0 to 2.8 seconds nothing happens: he reads, the room is silent, dust turns in a shaft of light. At the 2.9-second mark a heavy hardback slides off the far end of his table and hits the stone floor with a single flat crack, and he flinches hard, head snapping up. From 3.4 to 5.0 seconds the room returns to silence and he stares off toward the sound, breathing once. The camera is locked off on a tripod and never moves.

overall_soundscape:
Near-total silence for the first 2.8 seconds: only a distant HVAC breath and one faint page turn at 1.6 seconds, everything very quiet. At 2.9 seconds a single loud book impact on stone, sharp and dry with a short room tail, by far the loudest event in the clip. After 3.4 seconds the silence returns completely, with one soft intake of breath at 4.2 seconds.

non_diegetic_music:
N/A
baseline · 20 steps
No LoRA. The reference render.
165s · floor -57.3 · med -55.0 · peak -14.6 dB · jitter 3.9 · detail 0.328
ema-ckpt500 · 8 steps
larryvrh/drbaph Turbo LoRA, EMA weights. Round-3 winner.
76s · floor -61.1 · med -58.4 · peak -10.5 dB · jitter 8.8 · detail 0.518
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 -56.5 · med -53.8 · peak -9.2 dB · jitter 7.5 · detail 0.618
lightx2v · 4 steps · str 1.0
ModelTC reference settings.
45s · floor -56.2 · med -53.7 · peak 3.0 dB · jitter 76.4 · detail 0.495
lightx2v · 4 steps · str 1.0 · shift 12/6
Reference settings with our house audio shift.
44s · floor -67.5 · med -63.2 · peak -6.2 dB · jitter 76.7 · detail 0.477
lightx2v · 4 steps · str 1.0 · er_sde
Isolates the sampler at full strength.
44s · floor -68.7 · med -64.9 · peak -6.7 dB · jitter 48.3 · detail 0.478
lightx2v · 4 steps · str 0.75
Isolates strength on the reference sampler.
45s · floor -60.3 · med -58.3 · peak -2.1 dB · jitter 24.6 · detail 0.682
lightx2v · 4 steps · str 0.75 · shift 12/6
Lower strength with the house shift.
45s · floor -67.1 · med -63.6 · peak -5.4 dB · jitter 19.8 · detail 0.668
lightx2v · 4 steps · str 0.75 · er_sde
Kijai's recommended config. Pilot winner on jitter and luma.
44s · floor -67.3 · med -63.9 · peak -4.2 dB · jitter 20.9 · detail 0.489
lightx2v · 8 steps · str 1.0
Double the trained step count, reference settings.
76s · floor -55.5 · med -52.6 · peak -6.1 dB · jitter 67.9 · detail 0.417
lightx2v · 8 steps · str 1.0 · shift 12/6
Double steps with the house shift.
76s · floor -63.7 · med -59.6 · peak -4.7 dB · jitter 75.0 · detail 0.415
lightx2v · 8 steps · str 0.75 · er_sde
Kijai's config given twice the steps.
76s · floor -66.8 · med -63.8 · peak -7.3 dB · jitter 25.8 · detail 0.446
baseline · 20 steps + Spectrum
No LoRA, plus SpectrumApplyMiniMaxH3 (v0.1.9 defaults: degree 1, warmup 1, one-point bootstrap) skipping DiT evaluations.
106s · floor -57.2 · med -54.1 · peak -11.3 dB · jitter 7.1 · detail 0.418
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 -57.1 · med -54.3 · peak -8.3 dB · jitter 16.4 · detail 0.617
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 -61.7 · med -58.5 · peak -9.4 dB · jitter 6.6 · detail 0.430
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).
40s · floor -59.3 · med -56.9 · peak -7.7 dB · jitter 13.9 · detail 0.517
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 -58.2 · med -55.6 · peak -13.0 dB · jitter 4.2 · detail 0.323
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.
100s · floor -52.9 · med -49.9 · peak -12.0 dB · jitter 16.4 · detail 0.332
ema-ckpt850 · 8 steps
larryvrh's further-trained EMA checkpoint — the file he recommends — against the ckpt500 used everywhere else here.
76s · floor -58.2 · med -55.5 · peak -9.0 dB · jitter 7.4 · detail 0.554
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 -55.3 · med -51.5 · peak -8.0 dB · jitter 12.5 · detail 0.490
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 -54.2 · med -51.0 · peak -9.6 dB · jitter 7.4 · detail 0.498
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 -57.0 · med -54.0 · peak -10.1 dB · jitter 7.4 · detail 0.347
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 -43.0 · med -40.7 · peak 2.5 dB · jitter 105.9 · detail 0.496
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 -55.8 · med -52.9 · peak -9.6 dB · jitter 8.5 · detail 0.430
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 -65.0 · med -61.5 · peak -15.1 dB · jitter 7.7 · detail 0.309

Measured

configtimefloormedianpeakjitterdetailraw detailluma
baseline · 20 steps165s-57.3-55.0-14.63.90.32877046.7
ema-ckpt500 · 8 steps76s-61.1-58.4-10.58.80.51897848.6
ckpt500 (non-EMA) · 8 steps76s-56.5-53.8-9.27.50.618153652.8
lightx2v · 4 steps · str 1.045s-56.2-53.73.076.40.495174064.9
lightx2v · 4 steps · str 1.0 · shift 12/644s-67.5-63.2-6.276.70.477171064.7
lightx2v · 4 steps · str 1.0 · er_sde44s-68.7-64.9-6.748.30.478134557.0
lightx2v · 4 steps · str 0.7545s-60.3-58.3-2.124.60.682183158.1
lightx2v · 4 steps · str 0.75 · shift 12/645s-67.1-63.6-5.419.80.668184658.2
lightx2v · 4 steps · str 0.75 · er_sde44s-67.3-63.9-4.220.90.489103249.9
lightx2v · 8 steps · str 1.076s-55.5-52.6-6.167.90.417155560.9
lightx2v · 8 steps · str 1.0 · shift 12/676s-63.7-59.6-4.775.00.415152560.6
lightx2v · 8 steps · str 0.75 · er_sde76s-66.8-63.8-7.325.80.446117356.9
baseline · 20 steps + Spectrum106s-57.2-54.1-11.37.10.41869638.1
ema-ckpt500 · 8 steps + Spectrum61s-57.1-54.3-8.316.40.61775435.8
ema-ckpt500 · 8 steps + SageAttention50s-61.7-58.5-9.46.60.43088647.8
ema-ckpt500 · 8 steps + Sage + Spectrum40s-59.3-56.9-7.713.90.51780338.5
baseline · 20 steps + SageAttention98s-58.2-55.6-13.04.20.32375847.5
baseline · 20 steps + Sol-Attn100s-52.9-49.9-12.016.40.33283245.6
ema-ckpt850 · 8 steps76s-58.2-55.5-9.07.40.554151852.0
v4-step600 EMA · 8 steps77s-55.3-51.5-8.012.50.490117951.5
v4-step600 non-EMA · 8 steps76s-54.2-51.0-9.67.40.498128151.8
v4-step600 EMA · 8 steps · euler + beta76s-57.0-54.0-10.17.40.34771540.6
v4-step600 EMA · 4 steps45s-43.0-40.72.5105.90.496143269.7
v4-step600 EMA · 8 steps + SageAttention50s-55.8-52.9-9.68.50.430116551.8
ema-ckpt500 · 8 steps · euler + beta76s-65.0-61.5-15.17.70.30965744.5
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.