Loopable 360° camera orbits are deceptively hard: standard reference-to-video models either freeze the scene or drift off the start frame, which prevents clean stitching. This LoRA teaches MiniMax‑H3 FL2VA the motion pattern of a true orbit while keeping every object motionless in world space, so the clip both looks geometrically consistent and closes on the original frame for seamless chaining.
What Sets It Apart
- Learned, geometry-consistent orbit behavior: trained from rendered orbit sequences so the model moves the camera while preserving world positions and poses; this means parallax is the only apparent motion, enabling clean visual continuity when clips are concatenated.
- FL2VA-first+last integration: by conditioning on the same image for both first and last keyframes, the LoRA enforces an exact start/end match — so orbits don’t drift and can be looped or stitched without visible seams.
- Narrow, high-fidelity training domain: trained on 28 human-centric Gaussian-splat renders (768×768, 73 frames). That focused dataset yields strong performance on similar subjects (people, portrait-style scenes) but limits generalization to arbitrary scenes or aspect ratios.
- Lightweight, Comfy/ai-toolkit friendly: LoRA rank 16 adapter weights intended for use with MiniMax‑H3 FL2VA extensions (tested with ostris/ai-toolkit and Comfy-style model keys), making it easy to drop into existing FL2VA pipelines.
Who It's For
Great fit if you need short, loopable 360° orbits from a single reference photo (for product shots, portrait loops, or stitched sequences) and you can work within square, human-centric inputs and MiniMax‑H3 FL2VA. Look elsewhere if your scenes are wide-aspect, non-human, require long durations, or must include complex scene dynamics — the LoRA was trained on a small, specialized dataset and may still exhibit subtle blinks or micro-movement. Recommended inference: identical first+last keyframe, 768×768, ~73 frames (≈3 s at 24fps), LoRA strength 1.0, guidance none, audio off.
