pablodawson/MiniMax-H3-360-Orbit-LoRA · Hugging Face

Source: https://huggingface.co/pablodawson/MiniMax-H3-360-Orbit-LoRA

https://huggingface.co/pablodawson/MiniMax-H3-360-Orbit-LoRA#minimax-h3-360%C2%B0-orbit-lora-firstlast-frameMiniMax-H3 360° Orbit LoRA (first–last frame)

A LoRA forMiniMax-H3(FL2VA variant) that turns one photo into a frozen-time 360° camera orbit, and lands back on the exact frame it started from.

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.

https://huggingface.co/pablodawson/MiniMax-H3-360-Orbit-LoRA#resultsResults

Each comparison shows three renders of the same input with the same seed, prompt, resolution and step count:

leftmiddlerightbase model, first frame onlybase model, first + last framebase + this LoRA, first + last frame

The previews are downscaled animated WebP. Full-resolution MP4s: comparisonsskate·man2·girl2·man; LoRA-only clipsskate·man2·girl2·man.

https://huggingface.co/pablodawson/MiniMax-H3-360-Orbit-LoRA#about-the-comparisonsAbout 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.

https://huggingface.co/pablodawson/MiniMax-H3-360-Orbit-LoRA#why-this-loraWhy 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.

https://huggingface.co/pablodawson/MiniMax-H3-360-Orbit-LoRA#promptPrompt

Use this prompt verbatim.

One frozen instant. Only the camera moves. In a continuous 360 orbit. Preserve every person and object in exactly the same world position, orientation, shape and pose throughout the shot. Airborne objects remain suspended at the captured height and angle: no wobbling, shaking, spinning, drifting, falling or continued action. Keep faces, hands, clothing, liquids and the background motionless while retaining their natural appearance. Camera parallax is the only source of apparent movement. No cuts, zoom, morphing or added objects.

https://huggingface.co/pablodawson/MiniMax-H3-360-Orbit-LoRA#recommended-settingsRecommended settings

settingvaluebase weightsMiniMax-H3 FL2VA, pruned (Comfy-Org/MiniMax-H3int8 ConvRot repack)keyframesthe same image as firstandlast frame, for a full 360° loopresolution768 × 768frames73 (about 3 s at 24 fps)steps28guidancenone. MiniMax-H3 is guidance-distilled, so there is no CFG or negative promptLoRA strength1.0audiooff

https://huggingface.co/pablodawson/MiniMax-H3-360-Orbit-LoRA#usageUsage

These weights were trained and tested withostris/ai-toolkitand itsminimax\_h3model extension. The keys use ComfyUI naming (diffusion\_model\.\*).

https://huggingface.co/pablodawson/MiniMax-H3-360-Orbit-LoRA#training-detailsTraining details

https://huggingface.co/pablodawson/MiniMax-H3-360-Orbit-LoRA#dataData

Hand-selected high-qualityhuman Gaussian splatscollected from the internet, with an orbit video rendered around each one. Because the splats are static 3D scenes, every frame is geometrically consistent by construction. The dataset shows exactly the motion the LoRA should learn: the camera moves and nothing else does.

clips28 orbit rendersclip format768 × 768, 73 frames, 24 fpscaptionsthe single prompt above, for every clip; caption dropout 0.05audionone

https://huggingface.co/pablodawson/MiniMax-H3-360-Orbit-LoRA#hyperparametersHyperparameters

trainerostris/ai-toolkit, archminimax\_h3, partitionfl2va\_prunedbase modelMiniMax-H3 FL2VA pruned, int8 ConvRot (convrot8); Qwen3-VL-32B text encoder in nvfp4training adapterostris/minimax\_h3\_training\_adapterv1, active during training only and not needed for inferenceconditioningfirst frame of each clip as the keyframe (i2v)LoRArank 16, alpha 16, all transformer linear layers exceptadaln\_proj(208 modules)steps3000 (≈107 epochs)batch size1, no gradient accumulationoptimizerAdamW 8-bit, lr 1e-4, weight decay 1e-4precisionbf16, gradient checkpointingnoise scheduleflow matching, shifted timestepslossMSE, guidance loss target 3.5resolution buckets512, 768hardware1 × NVIDIA A100 80 GB, 10 h 24 min (about 12.5 s/step)

https://huggingface.co/pablodawson/MiniMax-H3-360-Orbit-LoRA#limitationsLimitations

• **The domain is narrow.**The training data is 28 human-centric square clips of 73 frames. Other subjects, aspect ratios and clip lengths are untested.
• You may still get some subtle movements, blinks, etc.

https://huggingface.co/pablodawson/MiniMax-H3-360-Orbit-LoRA#authorAuthor

Pablo Dawson