UltraGen CAP LoRA

LoRA checkpoints fine-tuned for accident-oriented video generation and editing. The checkpoints contain adapter weights only; the Wan2.1-T2V-1.3B base model, text encoder, and VAE are not included.

Training configuration

• Base model: Wan2.1-T2V-1.3B
• Training task: supervised fine-tuning (SFT)
• LoRA base module: DiT
• LoRA rank: 32
• Target modules: q, k, v, o, ffn.0, ffn.2
• Learning rate: 1e-4
• Weight decay: 0.01
• Epochs: 5
• Gradient accumulation steps: 1
• Training resolution: 1280 x 704
• Frames per training clip: 33
• Checkpoint interval: 500 steps

Checkpoints

The checkpoints/ directory contains checkpoints from step-500 through step-13000. Each checkpoint is a valid BF16 safetensors file containing 600 tensors and 43,745,280 LoRA parameters. step-13000.safetensors is the final checkpoint.

The matching Wan2.1-T2V-1.3B base model is required for inference.