Circle-Rotate LoRA

Implicit Motion Alignment for Rigid-Body Video Generation

Overview

Circle-Rotate LoRA enables smooth circular camera motion around static subjects in image-to-video generation. Fine-tuned on the Wan2.2 I2V model with only 1.4% parameter overhead, it achieves 15.1% subject drift reduction compared to baselines.

Key Features

Stable Subject: Objects remain stationary while camera orbits smoothly
Data-Centric: No explicit 3D pose supervision required
Lightweight: Only 1.4% LoRA parameters (high/low noise each 0.7%)
Plug & Play: Works with ComfyUI and standard inference pipelines

Demo Results

Baseline	Ours

Model Files

File	Description
`circle_rotate_h.safetensors`	High-noise LoRA for geometry alignment
`circle_rotate_l.safetensors`	Low-noise LoRA for texture refinement

Both files are required for optimal results.

Usage

With ComfyUI

Download both LoRA files to ComfyUI/models/loras/
Load workflow.json from the GitHub repo
Run inference

Command Line

# Clone the repository
git clone https://github.com/Jklaity/Circle-Rotate.git
cd Circle-Rotate

# Download LoRA weights
huggingface-cli download jk1741391802/circle-rotate-lora --local-dir ./checkpoints

# Run inference
python inference.py \
    --first_frame examples/first.png \
    --last_frame examples/last.png \
    --prompt "A car, camera smoothly orbits left" \
    --output output.mp4

Recommended Parameters

Parameter	Value
LoRA Strength	0.8 - 1.0
CFG Scale	1.6
Steps	4
Resolution	1280 x 720

Citation

@article{circle-rotate2025,
  title={Implicit Motion Alignment: A Data-Centric Empirical Study for Rigid-Body Video Generation},
  year={2025}
}

License

Apache License 2.0

Downloads last month: -