LoomVideo CAP LoRA

LoRA checkpoints fine-tuned for realistic traffic-accident video generation and editing on CAP data. Only inference-relevant adapter weights and resolved configuration files are published; datasets, ordinary logs, smoke-test checkpoints, generated videos, and evaluation outputs are excluded.

Base components

• LoomVideo pretrained checkpoint
• Wan2.2-TI2V-5B-Diffusers generator backbone
• Qwen3-VL-8B-Instruct understanding model

The matching base checkpoints are required for inference.

Training configuration

• Training steps: 3,200
• Checkpoint interval: 400 steps
• Gradient accumulation steps: 1
• Learning rate: 1e-4
• LoRA rank: 16
• LoRA alpha: 16
• LoRA dropout: 0.05
• Target modules: to_q, to_k, to_v, to_out.0
• LoRA parameters: 23,592,960
• Saved dtype: BF16
• Frames per clip: 32
• Training FPS: 24
• Resolution bucket: 640
• Seed: 0

Checkpoints

The repository includes checkpoints for steps 400, 800, 1200, 1600, 2000, 2400, 2800, and 3200. Each lora.pth contains 480 LoRA tensors. The final adapter is:

checkpoints/step-3200/lora.pth