Pablo Dawson's MiniMax-H3 LoRA Spins a Full 360 Orbit From One Photo
A community LoRA teaches MiniMax-H3's video model to spin the camera around a subject in a seamless, loopable 360 degree orbit from a single photo.
- New LoRA for MiniMax-H3 generates seamless 360 degree camera orbits from a single image.
- Uses first-last-frame mode with identical keyframes so orbits loop cleanly with no visible seam.
- Trained on 28 Gaussian splat orbit renders for guaranteed 3D geometric consistency.
- Rank 16 LoRA, 3000 steps on one A100 80GB in about 10.5 hours using ostris/ai-toolkit.
- Runs at 768x768, 73 frames, 28 steps, no CFG since H3 is guidance-distilled.
- Limited to human subjects and square aspect ratio; other domains untested.
A Rank-16 LoRA Teaches MiniMax-H3 to Orbit a Still
Pablo Dawson has released a community LoRA that generates a continuous 360-degree camera orbit from one photograph. The adapter model card targets the first-and-last-frame variant of MiniMax-H3, using the same image for both keyframes so the 73-frame clip returns to its starting view.
LoRA, short for low-rank adaptation, adds a compact set of learned weights without retraining the base model. This rank-16 adapter teaches a narrow behavior: move the camera around a subject while keeping the scene frozen in time.
MiniMax-H3’s reference-to-video mode treats input images as appearance guides, so its output has no fixed final view. The first-and-last-frame mode, labeled FL2VA, can constrain both endpoints. When given identical endpoint images, however, the base model tends to produce almost no motion. The adapter overrides that tendency and steers the model toward a complete orbit.
The result remains a generated view rather than a recovered 3D scene. A single photograph contains no information about hidden surfaces, so MiniMax-H3 must synthesize the back and sides of the subject.
Three Runs Reveal the Control
The model card compares three renders made with the same prompt, random seed, and inference-step count. Each run isolates the effect of the endpoint conditioning and the adapter.
- Base model, first frame only: The camera moves but drifts from the initial composition and does not return.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
