Building realtime interactive video on 4-step distilled MiniMax H3, stuck on character LoRA transfer
Big-Set9728 · reddit · 2026-09-07
A developer built a realtime interactive scene on FastVideo's FastH3 (4-step DMD2 distill of H3): you play a cop at a traffic stop, type your lines, and the woman in the car answers with generated audio. Clips are 5s and generation runs slightly faster than realtime.
Identity currently relies on reference conditioning, which holds up better than expected, but the author wants a trained character LoRA to lock one person across hundreds of clips and stop paying for reference conditioning per generation.
Progress so far:
- FastH3 publishes 6s for a 5s clip on 4x B200, with a lorapath/lorastrength hook
- diffusion-pipe and ai-toolkit both list H3 support
- Pitfall: VSA weights don't survive a stock ComfyUI conversion — 50 togatecompress tensors get silently dropped, producing noise, so people fall back to dense and miss the published speed numbers
Open questions:
- Does a LoRA trained on base H3 transfer cleanly to the 4-step distilled checkpoint (works in the Wan/Flux turbo world, unverified for H3)?
- Can a character LoRA stack with the distillation LoRA, or must you use the full distilled checkpoint?
- Is a T2I stills dataset enough for identity, or does a talking character need T2VA clips?
- Any real numbers for H3 sequence parallelism across datacentre GPUs?
The author is happy to pay for a consult or hands-on tuning.
Related event: Dev Tests Real-Time Interactive Video with MiniMax H3(2 posts)→
More from coding & agent
- User burns $6,000 of Claude Code usage in 7 days, mostly cache read tokens — guess_wat · 2026-09-07
- Grok Bot opens template marketplace; Haggle Bot found $100K+ in savings in a week — FinanceYF5 · 2026-09-07
- Developer builds AI agent harness on iMessage that trades and pays via PayBox — kleffew94 · 2026-09-07
- Codex tasked with designing its own sheet-metal part via DFM/quote MCP — PaulYacoubian · 2026-09-07
- Copilot CLI's Astra fixes stuck PowerPoint on Mac via computer-use MCP tools — DanWahlin · 2026-09-07
- Reddit devs ask: are AI coding agents' costs worth the productivity gains at scale? — lowkeyskibidi · 2026-09-07