pablodawson/MiniMax-H3-360-Orbit-LoRA — new model trending #30 on Hugging Face
A MiniMax-H3 LoRA generates seamless 360-degree frozen-time camera orbits from one photo, closing on the start frame.
Pablo Dawson released MiniMax-H3-360-Orbit-LoRA, a rank-16 adapter for the MiniMax-H3 FL2VA video model that turns one photo into a frozen-time 360-degree camera orbit ending on the same frame. Using the identical image as first and last keyframes, the LoRA keeps people and objects still while the camera circles, so clips can be looped without a seam. It was trained for 3,000 steps on 28 human Gaussian-splat orbit clips at 768x768 and 73 frames, using ostris/ai-toolkit on one A100 80GB. Recommended inference is 768x768, 73 frames, 28 steps, LoRA strength 1.0, and no classifier-free guidance.
- LoRA makes MiniMax-H3 orbit a frozen scene and return to the start frame.
- Base model freezes or drifts when both keyframes are the same photo.
- Trained 3,000 steps, rank 16, on 28 splat-rendered human orbit clips.
- Best at 768x768, 73 frames, 28 steps, strength 1.0, guidance off.
- Author warns the domain is narrow and subtle motion can remain.
Full article850 words · extracted from huggingface.co · click to collapse
# MiniMax-H3 360° Orbit LoRA (first–last frame)
**A LoRA for [MiniMax-H3](https://huggingface.co/MiniMaxAI/MiniMax-H3) (FL2VA variant) that turns one photo into a frozen-time 360° camera orbit, and lands back on the exact frame it started from.**
<p align="center">
<img src="media/hero.webp" alt="Four 360° orbits generated by the LoRA from a single photo each" width="768">
</p>
Give the model the same image as the first and the last keyframe, and the LoRA orbits the camera all the way around the subject while the scene stays frozen. The clip closes on its own first frame, so orbits can be chained or stitched into longer shots without a visible seam.
## Results
Each comparison shows three renders of the same input with the same seed, prompt, resolution and step count:
| left | middle | right |
|---|---|---|
| base model, first frame only | base model, first + last frame | **base + this LoRA, first + last frame** |
<p align="center"><img src="media/skate_comparison.webp" alt="skate: base first-only vs base first+last vs LoRA first+last" width="100%"></p>
<p align="center"><img src="media/man2_comparison.webp" alt="man2: base first-only vs base first+last vs LoRA first+last" width="100%"></p>
<p align="center"><img src="media/girl2_comparison.webp" alt="girl2: base first-only vs base first+last vs LoRA first+last" width="100%"></p>
<p align="center"><img src="media/man_comparison.webp" alt="man: base first-only vs base first+last vs LoRA first+last" width="100%"></p>
The previews are downscaled animated WebP. Full-resolution MP4s: comparisons [skate](media/skate_comparison.mp4) · [man2](media/man2_comparison.mp4) · [girl2](media/girl2_comparison.mp4) · [man](media/man_comparison.mp4); LoRA-only clips [skate](media/skate_lora.mp4) · [man2](media/man2_lora.mp4) · [girl2](media/girl2_lora.mp4) · [man](media/man_lora.mp4).
### About the comparisons
- **Base, first + last frame:** when both keyframes are the same image, the base model reads the clip as a still and barely moves.
- **Base, first frame only:** the camera moves, but drifts away and never returns to the starting view.
- **With the LoRA:** the camera travels the full orbit and returns to the starting frame, with smooth motion and no cuts.
## Why this LoRA?
Reference-to-video (Ref2VA) treats its input images as loose appearance references, so clips don't end on a known frame and can't be joined cleanly. First–last-frame generation (FL2VA) pins both ends, but out of the box it freezes when both ends are the same image. This LoRA teaches the model a real, geometry-consistent orbit, while FL2VA keyframe pinning at inference keeps both ends exact.