import requests

response = requests.post(
    "https://api.wavespeed.ai/api/v3/wavespeed-ai/wan-2.2/image-to-video-lora",
    headers={
        "Content-Type": "application/json",
        "Authorization": "Bearer YOUR_API_KEY"
    },
    json={
        "image": "https://example.com/input_image.jpg",
        "prompt": "lexy_vox, gentle head turn, warm smile, natural movement",
        "size": "832x480",
        "duration": 5,
        "low_noise_loras": [{"path": "https://huggingface.co/dustin2050/lexy-vox-i2v-lora-training-wan2.2/resolve/main/lexy_vox_i2v_v2_diffusers.safetensors", "scale": 1.0}],
        "high_noise_loras": [{"path": "https://huggingface.co/dustin2050/lexy-vox-i2v-lora-training-wan2.2/resolve/main/lexy_vox_i2v_v2_diffusers.safetensors", "scale": 1.0}]
    }
)

Important: Wan 2.2 uses a dual-DiT architecture. You must pass the LoRA to both low_noise_loras and high_noise_loras. Using only one will produce incorrect results.

Local Inference (musubi-tuner)

Use the kohya format files (kohya/ directory) with both --lora_weight and --lora_weight_high_noise:

bash

python wan_generate_video.py \
  --task i2v-14B \
  --prompt "lexy_vox, gentle head turn, warm smile" \
  --image input.jpg \
  --lora_weight kohya/lexy_vox_i2v_v2.safetensors \
  --lora_weight_high_noise kohya/lexy_vox_i2v_v2.safetensors \
  --video_size 832 480 \
  --video_length 33 \
  --seed 42

Trigger Word

Use lexy_vox in your prompts to activate the character identity.

Files

File	Format	Usage
`lexy_vox_i2v_v2_diffusers.safetensors`	Diffusers	Wavespeed API, HF Diffusers pipeline
`lexy_vox_i2v_v2_diffusers-000005..020.safetensors`	Diffusers	Epoch checkpoints (5, 10, 15, 20)
`kohya/lexy_vox_i2v_v2.safetensors`	Kohya	Local inference (musubi-tuner), ComfyUI
`kohya/lexy_vox_i2v_v2-000005..020.safetensors`	Kohya	Epoch checkpoints (5, 10, 15, 20)

The final checkpoints (without epoch number) are from epoch 25 (best).

Format Difference

Diffusers: Key pattern diffusion_model.blocks.0.cross_attn.k.lora_A.weight (800 keys)
Kohya: Key pattern lora_unet_blocks_0_cross_attn_k.lora_down.weight (1200 keys incl. alpha)

Both contain identical weights, just different key naming conventions. Convert between them using musubi-tuner's convert_lora.py.

Training Details

Parameter	Value
Base Model	Wan 2.2 I2V 14B (dual-DiT)
Framework	musubi-tuner (kohya-ss)
Task	`i2v-A14B`
LoRA Rank (dim)	32
LoRA Alpha	32
Optimizer	Prodigy (`prodigyopt.Prodigy`)
Learning Rate	1.0 (Prodigy-managed)
LR Scheduler	Cosine (50 warmup steps)
Epochs	25
Dropout	0.1
Precision	fp16 (fp8 base model)
Gradient Accumulation	2
Timestep Sampling	shift
Discrete Flow Shift	5.0
Timestep Boundary	900
Seed	42
Final Loss	~0.013
Training Time	~8.4 hours
Hardware	NVIDIA RTX PRO 6000 Blackwell (102 GB VRAM)

Training Dataset

Dataset	Count	Resolution	Repeats	Purpose
Close-up portraits	22 images	1536x1536	x3	Face fidelity
Regular images	24 images	1536x1536	x2	Full body / scenes
Videos	23 clips	640x640 @ 33 frames	x1	Motion learning

Total: 160 batches per epoch, 4000 total steps (25 epochs x 160 batches / 2 grad accum).

Training Data

This repo includes the full training dataset in training_data/:

training_data/images_closeups/ – 22 close-up portraits with captions
training_data/images_upscaled/ – 24 regular images with captions
training_data/videos/ – 23 video clips with captions

Training scripts and configs are in scripts/.

T2V LoRA: dustin2050/lexy-vox-t2v-lora-training-wan2.2

dustin2050/lexy-vox-i2v-lora-training-wan2.2

Lexy Vox – Wan 2.2 I2V LoRA (v2)

Quick Start (Wavespeed API)