import torch
from diffusers import DiffusionPipeline
from diffusers.utils import load_image, export_to_video
# switch to "mps" for apple devices
pipe = DiffusionPipeline.from_pretrained("Wan-AI/Wan2.1-I2V-14B-720P", dtype=torch.bfloat16, device_map="cuda")
pipe.load_lora_weights("fwwrsd/ohwx-wan-lora")
prompt = "ohwx"
input_image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/guitar-man.png")
image = pipe(image=input_image, prompt=prompt).frames[0]
export_to_video(output, "output.mp4")
WAN 2.1 LoRA — ohwx
A personalized LoRA (Low-Rank Adaptation) trained on WAN 2.1 14B for generating video content with a specific identity. Works with both Image-to-Video and Text-to-Video WAN 2.1 pipelines.
Trained using dual-mode Musubi Tuner (high + low noise models → single LoRA file).
Quick Start
Direct Download URL
https://huggingface.co/fwwrsd/ohwx-wan-lora/resolve/main/lora.safetensors
ComfyUI Setup
- Download
lora.safetensors→ place inComfyUI/models/loras/ - Use WAN LoRA Loader node
- Set trigger word:
ohwx
Load Directly from URL (ComfyUI)
Many LoRA loader nodes support loading directly from a HuggingFace URL:
https://huggingface.co/fwwrsd/ohwx-wan-lora/resolve/main/lora.safetensors
No download needed — ComfyUI caches it automatically.
Download via Command Line
# wget
wget https://huggingface.co/fwwrsd/ohwx-wan-lora/resolve/main/lora.safetensors -O lora_ohwx.safetensors
# curl
curl -L https://huggingface.co/fwwrsd/ohwx-wan-lora/resolve/main/lora.safetensors -o lora_ohwx.safetensors
# huggingface-cli
huggingface-cli download fwwrsd/ohwx-wan-lora lora.safetensors
Recommended Settings
| Parameter | Image-to-Video | Text-to-Video |
|---|---|---|
| LoRA Strength (motion) | 0.3 — 0.4 | 0.3 — 0.4 |
| LoRA Strength (identity) | 0.85 — 0.95 | 0.85 — 0.95 |
| CFG Scale | 0.52 | 1.0 |
| Steps | 30 — 50 | 30 — 50 |
| Sampler | euler / dpmpp_2m | euler / dpmpp_2m |
Trigger word: ohwx — include in your prompt to activate the LoRA.
Training Details
| Parameter | Value |
|---|---|
| Base Model | Wan-AI/Wan2.1-I2V-14B-720P |
| Training Method | Musubi Tuner (dual-mode: high + low noise) |
| LoRA Rank | 16 |
| Learning Rate | 1e-4 |
| LR Scheduler | cosine with 5% warmup |
| Optimizer | adamw + LoRA+ (ratio=4) |
| Training Steps | ~unknown |
| Epochs | unknown |
| Resolution | 1024px |
| Dataset Size | unknown images |
| Captions | No (photos only) |
| Precision | fp16 (LoRA) + fp8 (base model) |
| Preset | standard |
| Created | 2026-06-20 |
| GPU | NVIDIA H200 SXM 141GB |
Architecture
This is a dual-mode LoRA trained with --timestep_boundary 875:
- High-noise model (timesteps > 875): Handles initial structure and motion
- Low-noise model (timesteps ≤ 875): Handles fine details and identity
Both models are trained simultaneously and packed into a single .safetensors file.
Compatible with any WAN 2.1 workflow that supports LoRA.
License
Apache 2.0 — free for personal and commercial use.
Trained with NanoBanana LoRA Bot on RunPod
- Downloads last month
- 20
Model tree for fwwrsd/ohwx-wan-lora
Base model
Wan-AI/Wan2.1-I2V-14B-720P