Dense polyphony with a foreground voice

What this stresses: Many simultaneous sources with one that must stay intelligible. Yesterday's crowd scene was where the non-EMA noise bed collapsed hardest; this stacks more layers on top.
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 shot inside a busy hotel kitchen during service: stainless steel passes, six cooks in white moving fast in the background, steam and gas flames, hard overhead light. In the foreground, right of frame, a heavyset head chef in his fifties with a grey moustache and a sweat-soaked collar leans toward the camera. From 0.0 to 1.2 seconds he reads a ticket, the kitchen roaring around him. At the 1.3-second mark he looks up and (S1), in a hoarse carrying shout that cuts straight through the noise, says: <d>[English] Table nine is walking in four minutes, I need that sauce now.</d> ending at 4.0 seconds. From 4.2 seconds he slams the ticket onto the pass rail and turns away. The background cooks never stop moving; the camera drifts slowly left throughout.

overall_soundscape:
A dense wall of kitchen noise for the entire clip: roaring gas burners, a loud extractor hood, pans clattering, plates stacking, a ticket printer chattering, two other cooks calling to each other indistinctly, running water. (S1)'s shout from 1.3 seconds sits clearly above all of it and stays fully intelligible without the background dropping in level. A hard metallic ticket-rail slam at 4.3 seconds.

non_diegetic_music:
N/A
baseline · 20 steps
No LoRA. The reference render.
166s · floor -33.4 · med -21.7 · peak -0.8 dB · jitter 119.1 · detail 0.157
ema-ckpt500 · 8 steps
larryvrh/drbaph Turbo LoRA, EMA weights. Round-3 winner.
77s · floor -34.8 · med -20.3 · peak -0.6 dB · jitter 137.4 · detail 0.114
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.
77s · floor -29.8 · med -16.8 · peak 3.9 dB · jitter 181.6 · detail 0.236
lightx2v · 4 steps · str 1.0
ModelTC reference settings.
45s · floor -35.8 · med -15.2 · peak 3.3 dB · jitter 329.3 · detail 0.215
lightx2v · 4 steps · str 1.0 · shift 12/6
Reference settings with our house audio shift.
45s · floor -36.7 · med -17.6 · peak 3.8 dB · jitter 327.0 · detail 0.202
lightx2v · 4 steps · str 1.0 · er_sde
Isolates the sampler at full strength.
45s · floor -34.9 · med -15.1 · peak 2.3 dB · jitter 266.3 · detail 0.138
lightx2v · 4 steps · str 0.75
Isolates strength on the reference sampler.
45s · floor -47.4 · med -24.3 · peak -0.0 dB · jitter 262.8 · detail 0.237
lightx2v · 4 steps · str 0.75 · shift 12/6
Lower strength with the house shift.
45s · floor -47.5 · med -24.3 · peak 0.7 dB · jitter 267.1 · detail 0.235
lightx2v · 4 steps · str 0.75 · er_sde
Kijai's recommended config. Pilot winner on jitter and luma.
45s · floor -36.5 · med -16.4 · peak 2.4 dB · jitter 223.8 · detail 0.131
lightx2v · 8 steps · str 1.0
Double the trained step count, reference settings.
77s · floor -31.1 · med -12.9 · peak 1.1 dB · jitter 324.2 · detail 0.129
lightx2v · 8 steps · str 1.0 · shift 12/6
Double steps with the house shift.
77s · floor -37.2 · med -12.6 · peak 2.7 dB · jitter 326.8 · detail 0.137
lightx2v · 8 steps · str 0.75 · er_sde
Kijai's config given twice the steps.
77s · floor -36.5 · med -14.4 · peak 1.8 dB · jitter 247.7 · detail 0.132
baseline · 20 steps + Spectrum
No LoRA, plus SpectrumApplyMiniMaxH3 (v0.1.9 defaults: degree 1, warmup 1, one-point bootstrap) skipping DiT evaluations.
107s · floor -34.0 · med -21.1 · peak -1.4 dB · jitter 143.8 · detail 0.130
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 -29.9 · med -15.4 · peak 2.5 dB · jitter 129.3 · detail 0.148
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 -31.9 · med -18.8 · peak -0.9 dB · jitter 144.7 · detail 0.123
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 -30.4 · med -18.1 · peak 2.3 dB · jitter 123.7 · detail 0.160
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 -33.3 · med -22.8 · peak -0.5 dB · jitter 130.8 · detail 0.144
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.
98s · floor -33.7 · med -20.5 · peak 1.1 dB · jitter 157.5 · detail 0.261
ema-ckpt850 · 8 steps
larryvrh's further-trained EMA checkpoint — the file he recommends — against the ckpt500 used everywhere else here.
77s · floor -31.0 · med -18.2 · peak 2.2 dB · jitter 157.6 · detail 0.169
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.
78s · floor -38.2 · med -22.2 · peak -3.3 dB · jitter 175.5 · detail 0.196
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).
77s · floor -30.9 · med -16.4 · peak 2.9 dB · jitter 164.4 · detail 0.191
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.
77s · floor -39.0 · med -21.7 · peak -2.5 dB · jitter 156.1 · detail 0.173
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 -37.5 · med -19.9 · peak -0.2 dB · jitter 173.0 · detail 0.181
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 -32.3 · med -17.1 · peak -0.6 dB · jitter 153.0 · detail 0.159
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.
77s · floor -34.8 · med -23.0 · peak -2.3 dB · jitter 136.0 · detail 0.120

Measured

configtimefloormedianpeakjitterdetailraw detailluma
baseline · 20 steps166s-33.4-21.7-0.8119.10.15742070.7
ema-ckpt500 · 8 steps77s-34.8-20.3-0.6137.40.11434874.6
ckpt500 (non-EMA) · 8 steps77s-29.8-16.83.9181.60.23668671.1
lightx2v · 4 steps · str 1.045s-35.8-15.23.3329.30.21578285.6
lightx2v · 4 steps · str 1.0 · shift 12/645s-36.7-17.63.8327.00.20276785.7
lightx2v · 4 steps · str 1.0 · er_sde45s-34.9-15.12.3266.30.13849580.9
lightx2v · 4 steps · str 0.7545s-47.4-24.3-0.0262.80.23792787.1
lightx2v · 4 steps · str 0.75 · shift 12/645s-47.5-24.30.7267.10.23591287.1
lightx2v · 4 steps · str 0.75 · er_sde45s-36.5-16.42.4223.80.13144178.8
lightx2v · 8 steps · str 1.077s-31.1-12.91.1324.20.12949684.5
lightx2v · 8 steps · str 1.0 · shift 12/677s-37.2-12.62.7326.80.13752084.1
lightx2v · 8 steps · str 0.75 · er_sde77s-36.5-14.41.8247.70.13250882.0
baseline · 20 steps + Spectrum107s-34.0-21.1-1.4143.80.13034668.6
ema-ckpt500 · 8 steps + Spectrum61s-29.9-15.42.5129.30.14834771.2
ema-ckpt500 · 8 steps + SageAttention50s-31.9-18.8-0.9144.70.12335772.7
ema-ckpt500 · 8 steps + Sage + Spectrum41s-30.4-18.12.3123.70.16036670.3
baseline · 20 steps + SageAttention98s-33.3-22.8-0.5130.80.14441176.2
baseline · 20 steps + Sol-Attn98s-33.7-20.51.1157.50.26163781.3
ema-ckpt850 · 8 steps77s-31.0-18.22.2157.60.16955173.6
v4-step600 EMA · 8 steps78s-38.2-22.2-3.3175.50.19656173.5
v4-step600 non-EMA · 8 steps77s-30.9-16.42.9164.40.19151468.5
v4-step600 EMA · 8 steps · euler + beta77s-39.0-21.7-2.5156.10.17348873.3
v4-step600 EMA · 4 steps45s-37.5-19.9-0.2173.00.18174094.5
v4-step600 EMA · 8 steps + SageAttention50s-32.3-17.1-0.6153.00.15949473.3
ema-ckpt500 · 8 steps · euler + beta77s-34.8-23.0-2.3136.00.12033470.0
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.