WAN2.2 Workflow+LORA (I2V,T2V) 4GB-VRAM GGUF
Details
Download Files
Model description
Video Generation Workflow (WAN 2.2)
Optimized for RTX 3050 Laptop (4 GB VRAM) and ComfyUI Please use the 14b Models
With the Light2v Lora in the suggested resource section! if not use 6 for the cfg and 30-60 steps in total.
also make sure for the 14b versions the secound sampler has a low denoise setting 0.3-0.5
Video Explainer
🧠 Supported Workflows
Wan2.2 T2V (Text‑to‑Video) 14B
Wan2.2 I2V (Image‑to‑Video) 14B
Wan2.2 TI2V‑5B (Hybrid Text/Image‑to‑Video)
⚙️ Nodes Used (Install via ComfyUI Manager or links below)
Note: rgthree Only needed for Stack Lora Loader
⚡ WAN 2.2 Highlights
Mixture‑of‑Experts (MoE) architecture with high‑noise and low‑noise experts for better generation quality
Cinematic aesthetic control: flexible management of lighting, color, composition, and visual style
Enhanced motion modeling: stronger support for complex, smooth camera and object movement
Efficient compression: the TI2V‑5B model uses a high‑compression VAE to run on as little as 8 GB VRAM with ComfyUI offloading
📦 Model Downloads + VAE
*these are conversions from the original models to run on less VRAM.
All these GGUF conversions are done at the moment by:
https://huggingface.co/bullerwins
https://huggingface.co/QuantStack
*If you cant find the model you are looking for check out there profiles!
🧩 Additional Required Files (Do not downlaod from Model Downloads)
📥 What to Download & How to Use It
✅ Quantization Tips:
Q_5 – 🔥 Best balance of speed and quality
Q_3_K_M – Fast and fairly accurate
Q_2_K – Usable, but with some quality loss
5B models – ⚡ Super fast, lower detail (good for testing)
14B models – 🎯 High quality, slower and VRAM-heavy
Reminder: Lower "Q" = faster and less VRAM, but lower quality
Higher "Q" = better quality, but more VRAM and slower speed