MiniMax H3 round-up: new LoRAs, speech-tag voice control, AMD Strix Halo local runs
optimisticalish · reddit · 2026-10-05
- New LoRAs: a trigger-word-free 1980s fantasy-movie style LoRA (not good for combat scenes) and a five-view-of-one-character LoRA.
- Age slider: an experimental age-slider for REF2VA H3 models, previously only available for FL2VA.
- Consistent scenes: a user encoded 20 slightly-overlapping room photos (1216×2040px, 4x5 grid) into RefMod for consistent locations with a dynamic camera—strict panorama stitching isn't required.
- Voice control: dialogue can be steered without <d> tags using only <inhale>, <pause>, <long pause>, <softer>, <breathe> tags plus <i> italics for word emphasis.
- Local runs: the h3-hrx GitHub project gets MiniMax H3 running on AMD Strix Halo AI PCs (about £3.5k in the UK).
More from Multimodal
- Creator releases radio edit remastered with Suno and Perplexity Computer — inductionheads · 2026-10-05
- Crowd-sourced Westminster photos dropped at exact spots using Google 3D Tiles, no scanning — MonaJalal_ · 2026-10-05
- HeyGen launches Video built on MiniMax H3, generating 720p 15-second clips — HeyGen · 2026-10-05
- Opus vs Astra 'jelly world' islands: better water costs $15.54 vs $10.51 — i_dg23 · 2026-10-05
- Image model A/B tests are back, hinting at a new version in hidden testing — flowersslop · 2026-10-05
- ComfyUI help wanted: MiniMax H3 ref2video with checkpoint and LoRA loaders — aanabobby31 · 2026-10-05