Skip to main content
Back

MiniMax-H3 LoRAs Let Anyone Spin Still Photos Into 360° Videos

Two community LoRAs give MiniMax-H3 a full 360-degree camera orbit from a single photo, using opposite tricks with very different geometry results.

Model, Full Access
  • Two community LoRAs give MiniMax-H3 a 360-degree camera orbit from a single photo
  • MATLOWAI ORB360 uses Ref2VA with 3 views (front, 120, 240) for true back-of-subject geometry
  • pablodawson FL2VA loops seamlessly from one image but sidesteps rather than fully circling
  • ORB360 step 1500 trained on 345 Blender orbits, 23 assets, 1500 steps, ~26 hours on one RTX PRO 6000
  • Measured: ORB360 lands within 2.9 degrees of true orbit per frame; CardSpin softens detail to roughly a third
  • Height-offset prompts, mid-clip altitude changes, and single-image back views remain unsolved

Two community low-rank adaptation files give MiniMax-H3 the same headline capability: turning a still image into a video of a camera circling a frozen subject. A LoRA is a compact adapter that changes a model’s behavior without replacing the base model. These releases use different MiniMax-H3 workflows, which affects input preparation, rear-view accuracy, loop closure, and image sharpness.

MATLOWAI’s ORB360-CardSpin repository contains a rank-32 LoRA for reference-to-video, or Ref2VA. Its step-1500 checkpoint produces clockwise camera orbits, while the CardSpin files rotate the source photograph like a thin physical card. Pablodawson’s 360-Orbit-LoRA targets first/last-frame-to-video, or FL2VA, and uses matching endpoint frames to create a closed loop.

The conditioning method sets the trade-off

A grounded Ref2VA orbit works best with several views of the subject, typically at 0°, 120°, and 240°. Each reference receives a caption with a precise timestamp and angle. The rear views provide actual visual evidence for hair, clothing, surfaces, and background geometry that the model would otherwise have to invent. Ref2VA treats those images as appearance and viewpoint references, so it does not guarantee that the final frame will match the first.

FL2VA conditions generation on exact endpoint images. The base model often freezes when the same photograph appears at both ends, so the 360-Orbit-LoRA teaches it to move through a geometry-consistent arc between those keyframes. In the published examples, supplying the same photograph at frames 0 and 73 produced pixel-identical endpoints and a loop without a visible seam.