Generating Multi-Speaker Dialogue Audio in MiniMax H3 via 32x32 Pixel Trick
sktksm · reddit · 2026-08-09
A developer shared a hardcore workflow for generating multi-minute, multi-speaker dialogue audio using the MiniMax H3 omni-modal model in ComfyUI. The core trick is setting the video latent size to a tiny 32×32 pixels, which makes the video stream negligible and forces almost all of the model's capacity into generating high-quality audio.
Because H3's native output limit is about 15 seconds, generating longer clips causes quality to collapse into gibberish. To create seamless long-form dialogue, the author developed specific pacing controls:
- Segmented Generation: Scripts are split into chunks of 15 seconds or less.
- Word Budgeting: Natural speech runs at 2-2.5 words per second. A 15-second clip holds roughly 30-38 words. Overloading causes rushed, overlapping speech, while under-filling causes the model to invent unwanted mumbling.
- Event-Driven Endings: Timestamps in prompts are ignored by the model. Instead, pacing and endings must be controlled by explicit event descriptions, such as "final two seconds: only fire and wind, no voices."
More from coding & agent
- A vector DB is not LLM memory: long thread maps write/read paths, failure modes and evals — techNmak · 2026-10-16
- OpenUI 1.0 launches open source, turning agent text answers into live interactive UIs — heyshrutimishra · 2026-10-16
- 10 GitHub repos to see what your AI agent really did, from Langfuse to claude-code-log — victor_explore · 2026-10-16
- Mosaic: multi-session agentic RL with hierarchical advantage estimation — LARK-Lab · 2026-10-16
- Recursive Environment Scaling tanks six models' success rates by 57–77 points — Lingxiao Jing · 2026-10-16
- Harvard's MPT: decentralized agent swarms learn self-orchestration via bidding — Harvard · 2026-10-16