MiniMax H3 measured: BGM vanishes at 480p; clip length buys it back
nakamep · reddit · 2026-08-20
Across 10+ controlled runs, MiniMax H3 renders no background music at 864x480 when dialogue is in the prompt. Prompt rewrites and raising steps from 20 to 30 do nothing; clip length does—going from 5s to 15s brings back drums and chords (accompaniment rises from -38.7 to -29.4 integrated LUFS). But one extra dialogue line (2.5s of speech) wipes out the 9 dB gain; drop dialogue entirely and 5s is already enough (-25.0, louder than the 720p reference).
Methodology is reproducible: stems split via demucs --two-stems=vocals, accompaniment measured with integrated LUFS, spectral flatness (0.094 noise vs 0.006 tonal music) and onset rate (0.89/s pad vs 4.97/s drums), all matching perception. Ambient sounds fare worse—audience murmur and costume rustle never rendered once; only a single transient tied to an on-screen flash appeared. Timecode instructions work unpredictably: one test nailed a crash at 3.5s, drum break at 7.5s, band return at 11s; another requesting a fade-in got the opposite—a steady decay from loudest at frame one.
More from Models
- Grok Bot Ran a 12-Hour Data Analysis Nonstop, Handling API Rate Limits Itself — Daniel_Farinax · 2026-10-16
- One year of Recursive Language Models: GPT-5-mini beats GPT-5 by 110% on 132k-token tasks, cheaper too — a1zhang · 2026-10-16
- Users Report Claude Models Suddenly More Compliant After Forced Compaction — RealSharpNinja · 2026-10-16
- Prediction: Opus 5.5-level models on laptops by 2027, backed by Qwen 27B data — cephaloform · 2026-10-16
- Users Say Krea 2 Output Degraded Sharply Since Krea 3 Beta Launched — szywis · 2026-10-16
- LLM-as-a-Verifier lifts Nemotron 3.5 Super by 9-10 points on agentic benchmarks — simonguozirui · 2026-10-16