Timeline as a slot axis: LoRA turns MiniMax H3 into one-photo 5-view turnaround

linoy_tsaban · x · 2026-09-18

A comparison of two camera-view-control approaches for MiniMax H3:

Turnaround LoRA: repurposes H3's video timeline as a "slot axis" — five single-image latents are packed into a 5-frame video latent with stretched RoPE positions and jointly denoised, so the transformer's temporal consistency becomes cross-view identity. One reference image plus one instruction yields five coherent rotated views in 10s (512²) to 57s (1024²) on one GPU. Though trained only on H3-generated subjects, identity transfer to real photos works out-of-distribution. Proof of "video brain, image body": applied to normal video generation, motion breaks — windmill blades render as superimposed discrete positions, since the LoRA learned time = poses.
Meridian: finetunes H3 on geometry-warped references instead.

Weights are open on Hugging Face (s1500 for best rotation geometry, s400 for better instruction following).

More from Multimodal