Trained on https://yorespot.com — where creators train faster, cheaper, and with minimal censorship (within legal limits). If you’re serious about building models, that’s where you should be.

WAN 2.2 I2V LoRA

========================================================

Trigger Word

ysp_forcbj24

Overview

This is a WAN 2.2 I2V LoRA trained for high-quality portrait video generation. It uses a two-phase training approach to properly separate low-noise detail learning and high-noise motion structure — resulting in cleaner motion, stronger consistency, and better frame coherence than single-pass LoRAs.

Training Details

Base Model:
• WAN 2.2 I2V 14B
• (LOW + HIGH LoRA files)

Method:
• Musubi Tuner video LoRA workflow
• Two-phase training (LOW timesteps → HIGH timesteps)

Dataset:
• 16 portrait video clips (512×512)
• Hand-captioned (.txt per clip)
• Full frame extraction (~30 FPS)
• Up to 33 frames per clip
• 4 repeats per clip

Training Config:
• LoRA rank: 32
• Alpha: 32
• Learning rate: 2e-4
• Optimizer: AdamW (8-bit)
• Scheduler: Cosine
• Precision: bf16 + gradient checkpointing

Training Process:
• 50 epochs (LOW noise model)
• 50 epochs (HIGH noise model)
• Total: 100 effective passes across both experts

Training Time:
• ~20 hours total runtime

Recommended Usage (IMPORTANT)

This is a dual-LoRA system. You MUST use both files together.

HIGH noise LoRA:
• Strength: 0.55

LOW noise LoRA:
• Strength: 0.90

Best results come from using both simultaneously in a WAN 2.2 I2V workflow.

Using only one will significantly reduce quality.

Why This Matters

Most LoRAs fail at motion consistency or overfit to static detail. This setup avoids that by splitting the learning problem:

• LOW model → structure, identity, fine detail
• HIGH model → motion, transitions, temporal coherence

The result is smoother, more stable, and more realistic video output.

If you want to train models like this yourself — faster, cheaper, and without the usual platform limitations — train at https://yorespot.com.