Open-source LoRA turns MiniMax H3 into seamless 360-degree orbit videos from a single photo
camenduru · x · 2026-10-15
camenduru shared MiniMax-H3-360-Orbit-LoRA, an open-source LoRA (by pablodawson, trained with ai-toolkit) for MiniMax H3's first-last-frame video model.
- How it works: feed the same image as both first and last keyframes, and the LoRA orbits the camera a full 360 degrees around the subject while the scene stays frozen in time.
- Why it matters: the base model outputs a near-still clip when both keyframes match, and drifts away without returning when given only the first frame; this LoRA produces a real geometry-constrained orbit that lands exactly back on the starting frame.
- Chaining: because the clip closes on its own first frame, orbits can be stitched into longer shots with no visible seam.
Comparisons with the same seed, resolution and step count show a clear difference between the base model and the LoRA-enhanced renders.
More from Multimodal
- Scale AI's Argon claims first 60%+ image score, dominates long-video benchmarks — zacharynado · 2026-10-15
- AI filmmaking enters a new phase: full camera control via Blender animatics to Dreamina — Uncanny_Harry · 2026-10-15
- AI video confidently claims "they put Halo into Modern Warfare 2", and it looks real — ahuja_priyank · 2026-10-15
- Screenshot of a video call where multiple "participants" appear fully AI-generated — Left-Hotel904 · 2026-10-15
- Neon Noir Elegance prompt template goes viral with striking pink-blue two-tone outputs — LudovicCreator · 2026-10-15
- Krea 2 Turbo + px16 LoRA generates game-ready pixel art in 5s, where GPT Image 2.5 fails — Angaisb_ · 2026-10-15