MiniMax H3 — Turbo LoRA comparisons

25 sampler/LoRA configurations of MiniMax H3 rendered across ten stress-test scenes, 5.17 s each with native audio, all at 960×544 and seed 424242, locally on an RTX 5080. The configs span the base model at 20 steps, larryvrh's Turbo LoRA (EMA and non-EMA) at 8 steps on both the v1 ckpt500/850 line and the newer v4 step-600 recipe, lightx2v's 4-step distillation across strength, sampler and audio-shift settings, and the SageAttention / Sol-Attn / Spectrum accelerators.

Pick a scene to see all 25 side by side, or use the A/B slider to overlay any two of them with a wipe and switch which one you hear.

• 
  Sustained singing — pitch stability across a held note

  Melodic voice. Distillation is known to hurt the quietest and the most sustained vocal registers; a held note exposes pitch drift and warble that speech hides.

  25/25 configs
• 
  Overlapping dialogue — two voices talking across each other

  Speaker separation under simultaneity. Yesterday's two-speakers scene alternated cleanly; this one deliberately collides the turns, which is where voice identity smears.

  25/25 configs
• 
  Accent and code-switch — Spanish into accented English mid-line

  Phoneme range beyond neutral English. Distillation tends to regress toward a generic accent; switching language mid-sentence makes that collapse audible.

  25/25 configs
• 
  Off-screen sound source — audio with no visual anchor

  Audio that cannot be read off the pixels. Tests whether the audio branch invents a source or honours a described one it can never see.

  25/25 configs
• 
  Near-silence into a hard transient — dynamic range

  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.

  25/25 configs
• 
  Four hard cuts in five seconds — scene-change handling

  Editing inside a single generation. Tests whether hard cuts survive distillation, or smear into dissolves, and whether the audio bed cuts with the picture.

  25/25 configs
• 
  Mirror reflection — subject and reflection must agree

  Geometric self-consistency. A reflection doubles every error: the model has to keep two views of one face in agreement for the whole clip.

  25/25 configs
• 
  Hands close-up — ten fingers doing precise work

  The classic generative failure. Full-frame hands at 0.5 MP with fine bimanual coordination, plus synchronised micro-foley.

  25/25 configs
• 
  Metronome count — audio events on an exact grid

  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.

  25/25 configs
• 
  Dense polyphony with a foreground voice

  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.

  25/25 configs

Configurations

Earlier rounds

The round-3 grid (10 scenes × 4 configs) was rendered before Kijai's MiniMax audio-sampler fix (bdcb886a, merged 2026-08-06), which reworked how the audio stream is carried through the sampler. Its audio measurements are not directly comparable with the pages above.