Type | Workflows |
Stats | 423 0 |
Reviews | (16) |
Published | May 26, 2025 |
Base Model | |
Hash | AutoV2 3E281BF519 |
π Wan2.1-VACE-14B (LoRA Accelerated): 10x Speed with CausVid LoRA for 3-Step Video Creation
π¬ Skyrocket Your Video Creation: Achieve ~10x Speed with Wan2.1 & the CausVid LoRA! π¬
π Overview
The Wan2.1-VACE-14B video diffusion model, when supercharged by the CausVid LoRA, is designed for high-quality, highly efficient video generation. It particularly excels at 480p and 720p resolutions through a streamlined 3-step ComfyUI workflow. This guide will walk you through the setup process to unlock this accelerated video generation capability, including options for full precision and quantized models like the fast Q3KL GGUF.
π Key Components
Diffusion Model (14B):
Full Precision: wan2.1_vace_14B_fp16.safetensors (Recommended for compatibility with LoRA examples)
Quantized (Civitai): wan2.1_vace_14B_Q4KM.safetensors
Quantized (GGUF - Civitai): wan2.1_vace_14B_Q3kl.gguf (Used in the 5-min example, requires GGUF loader)
This isn't the same as the GGUF format from Hugging Face (they missing !). I tested that one, and it didn't work for vid2vid tasks. So, I developed my own types specifically designed to work well with vid2vid. These are optimized and structured differently to ensure compatibility and better results , if u need another type do "a comment" after test hugging face one!
Performance LoRA (Essential for Speed):
VAE:
Text Encoder: Choose one:
umt5_xxl_fp16.safetensors (Recommended to match Kijai's wrapper compatibility for the LoRA workflow)
umt5_xxl_fp8_e4m3fn_scaled.safetensors (Smaller, fp8 version)
π File Organization
Place the downloaded files in the following structure within your ComfyUI directory:
ComfyUI/
βββ models/
β βββ diffusion_models/
β β βββ wan2.1_vace_14B_fp16.safetensors # Or Q4KM.safetensors, or Q3kl.gguf
β βββ text_encoders/
β β βββ umt5_xxl_fp16.safetensors # Or the fp8 version
β βββ loras/
β β βββ Wan21_CausVid_14B_T2V_lora_rank32.safetensors
β βββ vae/
β βββ wan_2.1_vae.safetensors
π¨ Model Showcase: Rapid 720p Cinematic Shots
This setup, featuring Wan2.1-VACE-14B and the CausVid LoRA, excels at producing 720p (and 480p) video clips with remarkable speed, even faster with quantized GGUF models. It's ideal for quick iterations, creative experimentation, and efficient content creation, all streamlined by a 3-step workflow.
π‘ Usage Tips
Model & LoRA Configuration: For maximum speed and quality, ensure you are using the appropriate 14B model (e.g.,
wan2.1_vace_14B_fp16.safetensors
orwan2.1_vace_14B_Q3kl.gguf
) paired with theWan21_CausVid_14B_T2V_lora_rank32.safetensors
LoRA. The LoRA should be applied with a strength typically around 1.0.Text Encoder: The
umt5_xxl_fp16.safetensors
text encoder is recommended for best compatibility with existing examples and Kijai's original demonstrations. The fp8 version can save VRAM.Resolution: This setup is optimized for 480p and 720p video generation.
Performance Gains:
Without LoRA (fp16): An 81-frame 720p video might take ~40 minutes on an RTX 4090.
With CausVid LoRA (fp16): The same video can be generated in ~4 minutes on an RTX 4090.
With CausVid LoRA & Q3KL GGUF: Potentially even faster, around 5 minutes or less for similar output on capable hardware with a GGUF loader.
Workflow Simplicity: The primary advantage, beyond speed, is the reduction to a 3-step generation process once models are loaded. This typically involves: 1. Prompting (Text Input), 2. KSampler (or equivalent node with LoRA and chosen model), 3. Video Combine (Output).
π Credits & Acknowledgements
Original Wan 2.1 models repackaged for ComfyUI by Comfy-Org: Wan 2.1 ComfyUI Repackaged on Hugging Face. The performance-boosting CausVid LoRA (Wan21_CausVid_14B_T2V_lora_rank32.safetensors
) was extracted and shared by Kijai. Original announcement and details: Kijai's Reddit Post. Quantized GGUF and Safetensors versions available on Civitai, enabling broader accessibility and speed. Gratitude to the developers of the underlying CausVid technique (presumably available under an MIT License or similar open terms).
π¨βπ» Developer Information
This guide was created by Abdallah Al-Swaiti:
For additional tools and updates, check out my other repositories.
β¨ Create Dreamy Videos with WAN 2.1 VACE and Pastel Dream! β¨