Testing shows AI video audio can't mix two reference tracks simultaneously
Portable_Solar_ZA · reddit · 2026-08-29
A developer testing for the Comfy Sync competition found that current audio generation models cannot take two reference audio tracks that play at the same time: with both a reference voice and reference music, the music either disappears when the narrator speaks or comes out barely audible and garbled.
The working pattern is "1 reference + 1 model-generated": reference voice with model-generated music, or model-generated voice over reference music, both of which layer correctly. He had planned a trailer with custom narrator plus reference music and had to abandon that combination.
More from Multimodal
- Priming Models: Teacher-Free One-Step Generation Hits FID 1.19 With 12x Less Compute — du_yilun · 2026-10-15
- Same prompt, three takes: Claude vs ChatGPT vs Figma Make UI showdown — Tegadesigns · 2026-10-15
- Designer finds Opus 5.5 picks up nuanced aesthetic taste with startling speed — floguo · 2026-10-15
- The math: 5 rerolls at 480p plus a 4K upscale costs $5 vs $9 at 720p — HeyAmit_ · 2026-10-15
- Dreamina workflow: generate at 480p, upscale to 4K, cut video costs 44% — HeyAmit_ · 2026-10-15
- One-prompt GPT-6 Astra writes two Palestrina-style motets in LilyPond — deanwball · 2026-10-15