Accent and code-switch — Spanish into accented English mid-line

What this stresses: Phoneme range beyond neutral English. Distillation tends to regress toward a generic accent; switching language mid-sentence makes that collapse audible.
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 close-up of an older Mexican woman in her sixties with silver-streaked hair pinned up, deep laugh lines and a floral apron, standing at a sunlit kitchen counter rolling masa between her palms, terracotta tiles and hanging copper pans behind her, warm window light from camera right. From 0.0 to 1.1 seconds she works the dough, glancing down at her hands. At the 1.2-second mark she looks up into the lens and (S1), in a warm weathered alto with a strong Mexican accent, says: <d>[Spanish] Mira, asi lo hacia mi abuela,</d> then continues without pause in accented English: <d>[English] and she never once used a recipe.</d> finishing at 4.3 seconds. From 4.4 seconds she laughs softly and slaps the disc of masa flat between her hands. Static camera on a tripod, shallow depth of field.

overall_soundscape:
A lived-in kitchen: a low gas flame, the soft slap of dough, a distant radio too quiet to make out words. (S1)'s voice is close and warm with clear Spanish vowel colour carrying into the English half of the line. A sharp dough slap at 4.6 seconds.

non_diegetic_music:
N/A
baseline · 20 steps
No LoRA. The reference render.
165s · floor -46.1 · med -12.1 · peak -0.7 dB · jitter 134.8 · detail 0.097
ema-ckpt500 · 8 steps
larryvrh/drbaph Turbo LoRA, EMA weights. Round-3 winner.
76s · floor -58.0 · med -12.2 · peak 0.3 dB · jitter 123.7 · detail 0.098
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 -46.7 · med -12.0 · peak -0.2 dB · jitter 163.3 · detail 0.209
lightx2v · 4 steps · str 1.0
ModelTC reference settings.
45s · floor -47.5 · med -12.5 · peak -0.1 dB · jitter 207.5 · detail 0.150
lightx2v · 4 steps · str 1.0 · shift 12/6
Reference settings with our house audio shift.
45s · floor -54.7 · med -13.8 · peak -0.4 dB · jitter 216.3 · detail 0.156
lightx2v · 4 steps · str 1.0 · er_sde
Isolates the sampler at full strength.
45s · floor -51.7 · med -13.0 · peak 0.2 dB · jitter 191.7 · detail 0.105
lightx2v · 4 steps · str 0.75
Isolates strength on the reference sampler.
45s · floor -48.9 · med -12.8 · peak -0.1 dB · jitter 158.1 · detail 0.152
lightx2v · 4 steps · str 0.75 · shift 12/6
Lower strength with the house shift.
45s · floor -57.7 · med -14.4 · peak 0.1 dB · jitter 162.1 · detail 0.152
lightx2v · 4 steps · str 0.75 · er_sde
Kijai's recommended config. Pilot winner on jitter and luma.
45s · floor -55.2 · med -12.5 · peak 0.6 dB · jitter 154.1 · detail 0.131
lightx2v · 8 steps · str 1.0
Double the trained step count, reference settings.
76s · floor -47.8 · med -13.1 · peak -0.6 dB · jitter 240.0 · detail 0.134
lightx2v · 8 steps · str 1.0 · shift 12/6
Double steps with the house shift.
76s · floor -49.9 · med -12.7 · peak -0.5 dB · jitter 238.7 · detail 0.125
lightx2v · 8 steps · str 0.75 · er_sde
Kijai's config given twice the steps.
76s · floor -51.6 · med -13.6 · peak 0.6 dB · jitter 154.1 · detail 0.129
baseline · 20 steps + Spectrum
No LoRA, plus SpectrumApplyMiniMaxH3 (v0.1.9 defaults: degree 1, warmup 1, one-point bootstrap) skipping DiT evaluations.
107s · floor -49.9 · med -12.5 · peak -0.3 dB · jitter 134.0 · detail 0.092
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 -60.8 · med -29.2 · peak -9.2 dB · jitter 107.9 · detail 0.094
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 -52.7 · med -13.4 · peak 0.4 dB · jitter 128.6 · detail 0.096
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 -60.8 · med -27.3 · peak -8.1 dB · jitter 103.8 · detail 0.083
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 -47.9 · med -12.6 · peak -0.6 dB · jitter 138.9 · detail 0.098
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 -47.5 · med -12.4 · peak 0.0 dB · jitter 127.1 · detail 0.147
ema-ckpt850 · 8 steps
larryvrh's further-trained EMA checkpoint — the file he recommends — against the ckpt500 used everywhere else here.
77s · floor -49.4 · med -12.2 · peak -0.7 dB · jitter 149.3 · detail 0.184
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 -50.1 · med -12.1 · peak -0.0 dB · jitter 139.4 · detail 0.116
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 -49.9 · med -12.2 · peak -0.4 dB · jitter 145.1 · detail 0.134
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 -47.2 · med -13.2 · peak 0.5 dB · jitter 133.8 · detail 0.114
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 -56.3 · med -14.5 · peak 0.5 dB · jitter 146.9 · detail 0.147
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 -49.5 · med -12.1 · peak -0.4 dB · jitter 143.3 · detail 0.123
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 -55.6 · med -14.2 · peak 0.4 dB · jitter 131.3 · detail 0.094

Measured

configtimefloormedianpeakjitterdetailraw detailluma
baseline · 20 steps165s-46.1-12.1-0.7134.80.09741174.9
ema-ckpt500 · 8 steps76s-58.0-12.20.3123.70.09844280.6
ckpt500 (non-EMA) · 8 steps76s-46.7-12.0-0.2163.30.20992881.9
lightx2v · 4 steps · str 1.045s-47.5-12.5-0.1207.50.15084490.6
lightx2v · 4 steps · str 1.0 · shift 12/645s-54.7-13.8-0.4216.30.15688989.8
lightx2v · 4 steps · str 1.0 · er_sde45s-51.7-13.00.2191.70.10553080.1
lightx2v · 4 steps · str 0.7545s-48.9-12.8-0.1158.10.15276485.8
lightx2v · 4 steps · str 0.75 · shift 12/645s-57.7-14.40.1162.10.15274786.5
lightx2v · 4 steps · str 0.75 · er_sde45s-55.2-12.50.6154.10.13158575.9
lightx2v · 8 steps · str 1.076s-47.8-13.1-0.6240.00.13471586.6
lightx2v · 8 steps · str 1.0 · shift 12/676s-49.9-12.7-0.5238.70.12567985.1
lightx2v · 8 steps · str 0.75 · er_sde76s-51.6-13.60.6154.10.12958182.4
baseline · 20 steps + Spectrum107s-49.9-12.5-0.3134.00.09235473.8
ema-ckpt500 · 8 steps + Spectrum61s-60.8-29.2-9.2107.90.09433075.0
ema-ckpt500 · 8 steps + SageAttention50s-52.7-13.40.4128.60.09643682.5
ema-ckpt500 · 8 steps + Sage + Spectrum40s-60.8-27.3-8.1103.80.08329672.8
baseline · 20 steps + SageAttention98s-47.9-12.6-0.6138.90.09841975.3
baseline · 20 steps + Sol-Attn99s-47.5-12.40.0127.10.14752470.6
ema-ckpt850 · 8 steps77s-49.4-12.2-0.7149.30.18486180.5
v4-step600 EMA · 8 steps77s-50.1-12.1-0.0139.40.11653480.1
v4-step600 non-EMA · 8 steps77s-49.9-12.2-0.4145.10.13461681.2
v4-step600 EMA · 8 steps · euler + beta76s-47.2-13.20.5133.80.11449172.9
v4-step600 EMA · 4 steps45s-56.3-14.50.5146.90.14771190.1
v4-step600 EMA · 8 steps + SageAttention50s-49.5-12.1-0.4143.30.12353979.7
ema-ckpt500 · 8 steps · euler + beta76s-55.6-14.20.4131.30.09432071.4
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.