MiniMax H3 takes up to 9 image, 3 video and 3 audio refs — directing, not prompting
LearnWithBishal · x · 2026-08-14
X creator @LearnWithBishal pitches the MiniMax H3 video model: it's currently 50% off on Magnific until September 1.
His point is that H3's biggest shift isn't quality but control — instead of hoping the model understands your vision, you guide it with up to 9 reference images, 3 videos and 3 audio clips, making generation feel "much closer to directing than prompting."
More from Multimodal
- Flux 3 Klien first impressions: occasional body horror like SD3, otherwise solid — rionix88 · 2026-10-16
- XGRIDS builds world's largest 3DGS town: 23.5 km² and 6.2 billion Gaussian Splats — bilawalsidhu · 2026-10-16
- IMMERSION: AI film compilation links 12 shorts with one continuous camera move — AI_Andrew · 2026-10-16
- A song with vocals, lyrics and 3D head animation created entirely by Opus 5.5 — repligate · 2026-10-16
- Alibaba open-sources CommerceGen-Studio for commercial-grade e-commerce images — alibabagroup · 2026-10-16
- EVA-Cap: Fudan and OPPO align video captioning with atomic audiovisual events via GRPO — Fudan-University · 2026-10-16