WAN 2.2 Animate – AI Video Subject Replacement (Sage, Flash Attention, Low VRAM)

Replace subjects in any video with the new WAN 2.2 Animate. I created a Windows installer that includes everything needed to get up and running—ComfyUI, a fully pre-configured workflow and all dependencies. This update brings Sage and Flash Attention optimizations, making high-quality AI video generation accessible even on mid-range hardware, with step-by-step ease and no manual setup required.
 
What is WAN 2.2 Animate? (Brief Summary)
 
WAN 2.2 Animate is a next-generation, multimodal video generative model featuring a Mixture-of-Experts (MoE) architecture for superior video quality. The model excels in cinematic-level aesthetic control, advanced motion capture, semantic accuracy, and seamless subject replacement or animation. Whether generating entirely new scenes or animating existing characters, WAN 2.2 delivers high-fidelity, realistic results with advanced control over lighting, color, and composition—ideal for video creators, animators, and content professionals seeking top-notch AI-driven video generation.
 
Preloaded Models within the Installer (Low VRAM)
umt5_xxl_fp8_e4m3fn_scaled.safetensors (ComfyUI\models\clip)
Patched for fast, accurate video-text understanding.
wan_2.1_vae.safetensors (ComfyUI\models\vae)
High-compression ratio VAE optimized for resource efficiency.
Wan2.2-Animate-14B-Q2_K.gguf (ComfyUI\models\unet)
Efficient UNet architecture to enable smooth video generation.
Wan21_T2V_14B_lightx2v_cfg_step_distill_lora_rank32.safetensors (ComfyUI\models\loras)
LoRA supports flexible scene relighting and compositional accuracy.
 
Note: The standard high-fidelity Wan2.2-Animate diffusion models (FP16 and FP8) are not included in the installer for size reasons. For best results, download them separately from Kijai’s Hugging Face repository!
WAN 2.2 Animate Diffusion Model Downloads
Find the full FP16/FP8 diffusion backbone here for enhanced output:
Kijai Wan2.2 Animate Diffusion Models.
 
Speed
Generate crisp AI-animated 480p videos in roughly 10–15 minutes on an RTX 4050 6GB VRAM GPU. The workflow dynamically adapts to your system, so faster render times and higher resolutions (720p/24fps) are achievable with higher-end GPUs (e.g., RTX 4090).
 
System Requirements
Nvidia RTX 30XX, 40XX, or 50XX series GPU (FP16 support required; GTX 10XX/20XX not tested)
CUDA-compatible GPU (minimum 6GB VRAM; more recommended)
Windows OS
At least 40GB free storage.
 
What’s Included
Portable, pre-configured ComfyUI and Wan2.2 install—no manual ComfyUI setup.
Automated download of all custom nodes and required models.
Complete audio-to-video workflow supporting advanced features like Sage/Flash Attention.
Workflow designed for simplicity and speed, including easy toggling and modular Fast Groups Bypasser for custom tasks.
 
Usage Notes
Compatible with both GGUF quantized models and standard diffusion models.
Portrait (single-character) clips yield best results—just enter your description and let the AI handle the rest.
Instantly bypass workflow nodes when you want to test or iterate rapidly.
To replace a subject: Upload a reference image plus a video clip (5 seconds or less). Use the Points Editor node to define and refine the subject’s area.
Finish the workflow tip:
Use the points editor node to outline your subject precisely—hold shift and left click to add green indicators for your target subject, and shift + right click to indicate background boundaries. This helps the AI separate foreground and background for flawless character swapping, especially useful for complex costumes or movement. Adjust the grow mask as needed for larger subjects, and switch between mix and move modes for different video compositing styles.
 
Learn More
This is NOT my tutorial video, but here’s a fantastic walkthrough showing subject selection and workflow tips:
https://www.youtube.com/watch?v=tSaJuj0yQkI